diff --git a/be/src/agent/agent_server.h b/be/src/agent/agent_server.h
index 11521eb09a..3aebc3d1f7 100644
--- a/be/src/agent/agent_server.h
+++ b/be/src/agent/agent_server.h
@@ -18,14 +18,11 @@
 #pragma once
 
 #include <butil/macros.h>
-#include <gen_cpp/AgentService_types.h>
 
 #include <memory>
 #include <string>
 #include <vector>
 
-#include "runtime/exec_env.h"
-
 namespace doris {
 
 class TaskWorkerPool;
diff --git a/be/src/agent/heartbeat_server.h b/be/src/agent/heartbeat_server.h
index 65e228dd8c..e08603043f 100644
--- a/be/src/agent/heartbeat_server.h
+++ b/be/src/agent/heartbeat_server.h
@@ -24,8 +24,6 @@
 #include <mutex>
 
 #include "common/status.h"
-#include "olap/olap_define.h"
-#include "runtime/exec_env.h"
 
 namespace doris {
 class ExecEnv;
@@ -34,7 +32,6 @@ class TMasterInfo;
 
 const uint32_t HEARTBEAT_INTERVAL = 10;
 class StorageEngine;
-class Status;
 class ThriftServer;
 
 class HeartbeatServer : public HeartbeatServiceIf {
diff --git a/be/src/agent/task_worker_pool.cpp b/be/src/agent/task_worker_pool.cpp
index ab7877fbf2..f93cee6a6b 100644
--- a/be/src/agent/task_worker_pool.cpp
+++ b/be/src/agent/task_worker_pool.cpp
@@ -43,7 +43,7 @@
 #include "common/logging.h"
 #include "common/status.h"
 #include "gutil/ref_counted.h"
-#include "gutil/stringprintf.h"
+#include "gutil/strings/numbers.h"
 #include "gutil/strings/substitute.h"
 #include "io/fs/file_system.h"
 #include "io/fs/local_file_system.h"
diff --git a/be/src/agent/task_worker_pool.h b/be/src/agent/task_worker_pool.h
index cd9b274d98..2f90f7b685 100644
--- a/be/src/agent/task_worker_pool.h
+++ b/be/src/agent/task_worker_pool.h
@@ -19,7 +19,6 @@
 
 #include <butil/macros.h>
 #include <gen_cpp/AgentService_types.h>
-#include <gen_cpp/HeartbeatService_types.h>
 #include <gen_cpp/Types_types.h>
 #include <stdint.h>
 
@@ -31,11 +30,8 @@
 #include <mutex>
 #include <set>
 #include <string>
-#include <utility>
-#include <vector>
 
 #include "common/status.h"
-#include "olap/data_dir.h"
 #include "olap/tablet.h"
 #include "util/countdown_latch.h"
 #include "util/metrics.h"
diff --git a/be/src/agent/topic_listener.h b/be/src/agent/topic_listener.h
index 1ad66b31a6..0ef9c597f1 100644
--- a/be/src/agent/topic_listener.h
+++ b/be/src/agent/topic_listener.h
@@ -17,7 +17,7 @@
 
 #pragma once
 
-#include "gen_cpp/AgentService_types.h"
+#include <gen_cpp/AgentService_types.h>
 
 namespace doris {
 
diff --git a/be/src/agent/topic_subscriber.h b/be/src/agent/topic_subscriber.h
index 3151cbe54c..490bd35d2b 100644
--- a/be/src/agent/topic_subscriber.h
+++ b/be/src/agent/topic_subscriber.h
@@ -20,13 +20,9 @@
 #include <gen_cpp/AgentService_types.h>
 
 #include <map>
-#include <mutex>
 #include <shared_mutex>
-#include <thread>
 #include <vector>
 
-#include "agent/topic_listener.h"
-
 namespace doris {
 class TopicListener;
 
diff --git a/be/src/agent/user_resource_listener.h b/be/src/agent/user_resource_listener.h
index b26ae3f279..6fd25bef67 100644
--- a/be/src/agent/user_resource_listener.h
+++ b/be/src/agent/user_resource_listener.h
@@ -18,13 +18,9 @@
 #pragma once
 
 #include <gen_cpp/AgentService_types.h>
-#include <gen_cpp/HeartbeatService_types.h>
 #include <stdint.h>
 
-#include <string>
-
 #include "agent/topic_listener.h"
-#include "runtime/exec_env.h"
 
 namespace doris {
 
diff --git a/be/src/agent/utils.cpp b/be/src/agent/utils.cpp
index 0e91eeaa49..19792fd1f8 100644
--- a/be/src/agent/utils.cpp
+++ b/be/src/agent/utils.cpp
@@ -17,7 +17,8 @@
 
 #include "agent/utils.h"
 
-#include <errno.h>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
 #include <gen_cpp/FrontendService.h>
 #include <gen_cpp/HeartbeatService_types.h>
 #include <gen_cpp/Types_types.h>
diff --git a/be/src/agent/utils.h b/be/src/agent/utils.h
index 8cd1895d6a..eea8f9f8cf 100644
--- a/be/src/agent/utils.h
+++ b/be/src/agent/utils.h
@@ -18,15 +18,11 @@
 #pragma once
 
 #include <butil/macros.h>
-#include <gen_cpp/FrontendService.h>
-#include <gen_cpp/HeartbeatService_types.h>
-#include <gen_cpp/MasterService_types.h>
 
 #include <map>
 #include <string>
 
 #include "common/status.h"
-#include "gutil/macros.h"
 
 namespace doris {
 class TConfirmUnusedRemoteFilesRequest;
diff --git a/be/src/common/configbase.cpp b/be/src/common/configbase.cpp
index 066aa0352d..d965687a23 100644
--- a/be/src/common/configbase.cpp
+++ b/be/src/common/configbase.cpp
@@ -19,7 +19,8 @@
 
 #include <algorithm>
 #include <cctype>
-#include <cerrno>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <cerrno> // IWYU pragma: keep
 #include <cstdlib>
 #include <cstring>
 #include <fstream> // IWYU pragma: keep
@@ -36,6 +37,7 @@
 
 #define __IN_CONFIGBASE_CPP__
 #include "common/config.h" // IWYU pragma: keep
+
 #undef __IN_CONFIGBASE_CPP__
 
 #include "common/status.h"
diff --git a/be/src/common/daemon.cpp b/be/src/common/daemon.cpp
index 6ef0d751f3..5c832cfd77 100644
--- a/be/src/common/daemon.cpp
+++ b/be/src/common/daemon.cpp
@@ -17,7 +17,8 @@
 
 #include "common/daemon.h"
 
-#include <bthread/errno.h>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
 #include <gflags/gflags.h>
 #include <gperftools/malloc_extension.h> // IWYU pragma: keep
 // IWYU pragma: no_include <bits/std_abs.h>
@@ -31,7 +32,6 @@
 // IWYU pragma: no_include <bits/chrono.h>
 #include <chrono> // IWYU pragma: keep
 #include <map>
-#include <memory>
 #include <ostream>
 #include <set>
 #include <string>
@@ -41,6 +41,7 @@
 #include "common/status.h"
 #include "olap/options.h"
 #include "olap/storage_engine.h"
+#include "olap/tablet_manager.h"
 #include "runtime/block_spill_manager.h"
 #include "runtime/exec_env.h"
 #include "runtime/load_channel_mgr.h"
@@ -53,8 +54,10 @@
 #include "util/disk_info.h"
 #include "util/doris_metrics.h"
 #include "util/mem_info.h"
+#include "util/metrics.h"
 #include "util/network_util.h"
 #include "util/perf_counters.h"
+#include "util/system_metrics.h"
 #include "util/thrift_util.h"
 #include "util/time.h"
 
diff --git a/be/src/common/daemon.h b/be/src/common/daemon.h
index ef7ece4620..6fd287a537 100644
--- a/be/src/common/daemon.h
+++ b/be/src/common/daemon.h
@@ -17,7 +17,6 @@
 
 #pragma once
 
-#include <memory>
 #include <vector>
 
 #include "gutil/ref_counted.h"
@@ -27,7 +26,6 @@
 namespace doris {
 
 struct StorePath;
-class Thread;
 
 class Daemon {
 public:
diff --git a/be/src/common/signal_handler.h b/be/src/common/signal_handler.h
index 85da4e7096..3be54d0d39 100644
--- a/be/src/common/signal_handler.h
+++ b/be/src/common/signal_handler.h
@@ -33,14 +33,13 @@
 
 #pragma once
 
+#include <gen_cpp/version.h>
 #include <glog/logging.h>
 #include <gutil/macros.h>
 
 #include <boost/stacktrace.hpp>
 #include <csignal>
 #include <ctime>
-
-#include "gen_cpp/version.h"
 #ifdef HAVE_UCONTEXT_H
 #include <ucontext.h>
 #endif
diff --git a/be/src/common/status.h b/be/src/common/status.h
index d989ba12c3..9f71f167ff 100644
--- a/be/src/common/status.h
+++ b/be/src/common/status.h
@@ -15,7 +15,8 @@
 #include <string_view>
 #include <utility>
 
-#include "common/compiler_util.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #ifdef ENABLE_STACKTRACE
 #include "util/stack_util.h"
 #endif
diff --git a/be/src/exec/arrow/arrow_reader.cpp b/be/src/exec/arrow/arrow_reader.cpp
index 6696d5489c..8e0559ca78 100644
--- a/be/src/exec/arrow/arrow_reader.cpp
+++ b/be/src/exec/arrow/arrow_reader.cpp
@@ -16,9 +16,7 @@
 // under the License.
 #include "exec/arrow/arrow_reader.h"
 
-#include <arrow/buffer.h>
 #include <arrow/record_batch.h>
-#include <opentelemetry/common/threadlocal.h>
 
 #include <algorithm>
 // IWYU pragma: no_include <bits/chrono.h>
@@ -26,6 +24,8 @@
 #include <ostream>
 #include <utility>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/logging.h"
 #include "io/fs/file_reader.h"
 #include "runtime/descriptors.h"
diff --git a/be/src/exec/arrow/arrow_reader.h b/be/src/exec/arrow/arrow_reader.h
index e685e7edaf..ad4c40d26b 100644
--- a/be/src/exec/arrow/arrow_reader.h
+++ b/be/src/exec/arrow/arrow_reader.h
@@ -17,20 +17,8 @@
 
 #pragma once
 
-#include <arrow/api.h>
-#include <arrow/buffer.h>
-#include <arrow/io/api.h>
-#include <arrow/io/file.h>
 #include <arrow/io/interfaces.h>
 #include <arrow/result.h>
-#include <gen_cpp/PaloBrokerService_types.h>
-#include <gen_cpp/PlanNodes_types.h>
-#include <gen_cpp/Types_types.h>
-#include <parquet/api/reader.h>
-#include <parquet/api/writer.h>
-#include <parquet/arrow/reader.h>
-#include <parquet/arrow/writer.h>
-#include <parquet/exception.h>
 #include <parquet/platform.h>
 #include <stddef.h>
 #include <stdint.h>
@@ -47,7 +35,6 @@
 
 #include "common/config.h"
 #include "common/status.h"
-#include "io/fs/file_reader.h"
 #include "io/fs/file_reader_writer_fwd.h"
 #include "vec/exec/format/generic_reader.h"
 
@@ -58,12 +45,8 @@ class RecordBatchReader;
 
 namespace doris {
 
-class ExecEnv;
-class TBrokerRangeDesc;
-class TNetworkAddress;
 class RuntimeState;
 class SlotDescriptor;
-class FileReader;
 class TupleDescriptor;
 
 namespace vectorized {
diff --git a/be/src/exec/arrow/parquet_reader.cpp b/be/src/exec/arrow/parquet_reader.cpp
index f0a1755324..4b924c0275 100644
--- a/be/src/exec/arrow/parquet_reader.cpp
+++ b/be/src/exec/arrow/parquet_reader.cpp
@@ -20,14 +20,12 @@
 #include <arrow/result.h>
 #include <arrow/status.h>
 #include <arrow/type.h>
-#include <opentelemetry/common/threadlocal.h>
 #include <parquet/exception.h>
 #include <parquet/file_reader.h>
 #include <parquet/metadata.h>
 #include <parquet/properties.h>
 #include <parquet/schema.h>
 
-#include <algorithm>
 #include <atomic>
 // IWYU pragma: no_include <bits/chrono.h>
 #include <chrono> // IWYU pragma: keep
@@ -38,6 +36,8 @@
 #include <ostream>
 #include <thread>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/logging.h"
 #include "common/status.h"
 #include "util/string_util.h"
diff --git a/be/src/exec/arrow/parquet_reader.h b/be/src/exec/arrow/parquet_reader.h
index 4f57fdc7b9..ac60a25eae 100644
--- a/be/src/exec/arrow/parquet_reader.h
+++ b/be/src/exec/arrow/parquet_reader.h
@@ -17,34 +17,14 @@
 
 #pragma once
 
-#include <arrow/api.h>
-#include <arrow/buffer.h>
-#include <arrow/io/api.h>
-#include <arrow/io/file.h>
-#include <arrow/io/interfaces.h>
-#include <arrow/status.h>
 #include <arrow/type_fwd.h>
-#include <gen_cpp/PaloBrokerService_types.h>
-#include <gen_cpp/PlanNodes_types.h>
-#include <gen_cpp/Types_types.h>
-#include <parquet/api/reader.h>
-#include <parquet/api/writer.h>
 #include <parquet/arrow/reader.h>
-#include <parquet/arrow/writer.h>
-#include <parquet/exception.h>
 #include <stdint.h>
 
-#include <atomic>
-#include <condition_variable>
-#include <list>
-#include <map>
 #include <memory>
-#include <mutex>
 #include <string>
-#include <thread>
 #include <vector>
 
-#include "common/config.h"
 #include "common/status.h"
 #include "exec/arrow/arrow_reader.h"
 #include "io/fs/file_reader_writer_fwd.h"
@@ -58,13 +38,8 @@ class FileMetaData;
 
 namespace doris {
 
-class ExecEnv;
-class TBrokerRangeDesc;
-class TNetworkAddress;
 class RuntimeState;
 class SlotDescriptor;
-class FileReader;
-class RowGroupReader;
 class TupleDescriptor;
 
 // Reader of parquet file
diff --git a/be/src/exec/base_scanner.cpp b/be/src/exec/base_scanner.cpp
index 20633d2c48..1f9b49a0f8 100644
--- a/be/src/exec/base_scanner.cpp
+++ b/be/src/exec/base_scanner.cpp
@@ -22,16 +22,18 @@
 #include <gen_cpp/Metrics_types.h>
 #include <gen_cpp/PlanNodes_types.h>
 #include <glog/logging.h>
-#include <opentelemetry/common/threadlocal.h>
 #include <parallel_hashmap/phmap.h>
 #include <stddef.h>
 
+#include <algorithm>
 #include <boost/iterator/iterator_facade.hpp>
 #include <iterator>
 #include <map>
 #include <string>
 #include <utility>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/consts.h"
 #include "gutil/casts.h"
 #include "runtime/define_primitive_type.h"
diff --git a/be/src/exec/base_scanner.h b/be/src/exec/base_scanner.h
index 291f7d653d..ee4ae973ef 100644
--- a/be/src/exec/base_scanner.h
+++ b/be/src/exec/base_scanner.h
@@ -33,12 +33,9 @@
 #include "vec/common/schema_util.h"
 #include "vec/core/block.h"
 #include "vec/exprs/vexpr.h"
-#include "vec/exprs/vexpr_context.h"
 
 namespace doris {
 
-class TupleDescriptor;
-class RowDescriptor;
 class RuntimeState;
 class TBrokerRangeDesc;
 class TBrokerScanRangeParams;
@@ -46,7 +43,6 @@ class TNetworkAddress;
 
 namespace vectorized {
 class VExprContext;
-class IColumn;
 
 using MutableColumnPtr = IColumn::MutablePtr;
 } // namespace vectorized
diff --git a/be/src/exec/data_sink.h b/be/src/exec/data_sink.h
index 3ea5da0c3c..ee847da138 100644
--- a/be/src/exec/data_sink.h
+++ b/be/src/exec/data_sink.h
@@ -20,8 +20,6 @@
 
 #pragma once
 
-#include <gen_cpp/DataSinks_types.h>
-#include <gen_cpp/Exprs_types.h>
 #include <opentelemetry/trace/span.h>
 #include <stddef.h>
 // IWYU pragma: no_include <opentelemetry/nostd/shared_ptr.h>
@@ -30,8 +28,6 @@
 #include <vector>
 
 #include "common/status.h"
-#include "runtime/descriptors.h"
-#include "runtime/query_statistics.h"
 #include "util/runtime_profile.h"
 #include "util/telemetry/telemetry.h"
 
diff --git a/be/src/exec/es/es_scan_reader.h b/be/src/exec/es/es_scan_reader.h
index 95b3e37980..a29dced15f 100644
--- a/be/src/exec/es/es_scan_reader.h
+++ b/be/src/exec/es/es_scan_reader.h
@@ -21,7 +21,6 @@
 #include <memory>
 #include <string>
 
-#include "exec/es/es_scroll_parser.h"
 #include "http/http_client.h"
 
 namespace doris {
diff --git a/be/src/exec/es/es_scroll_parser.h b/be/src/exec/es/es_scroll_parser.h
index d4edcb9ed2..be3f9fa8c5 100644
--- a/be/src/exec/es/es_scroll_parser.h
+++ b/be/src/exec/es/es_scroll_parser.h
@@ -24,8 +24,6 @@
 #include <vector>
 
 #include "rapidjson/document.h"
-#include "runtime/descriptors.h"
-#include "vec/core/block.h"
 #include "vec/data_types/data_type.h"
 
 namespace doris {
diff --git a/be/src/exec/exec_node.cpp b/be/src/exec/exec_node.cpp
index aced73a595..68360adba2 100644
--- a/be/src/exec/exec_node.cpp
+++ b/be/src/exec/exec_node.cpp
@@ -22,7 +22,6 @@
 
 #include <gen_cpp/Metrics_types.h>
 #include <gen_cpp/PlanNodes_types.h>
-#include <opentelemetry/common/threadlocal.h>
 #include <thrift/protocol/TDebugProtocol.h>
 
 #include <map>
@@ -30,6 +29,8 @@
 #include <typeinfo>
 #include <utility>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/config.h"
 #include "common/logging.h"
 #include "common/object_pool.h"
@@ -43,7 +44,6 @@
 #include "util/uid_util.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_nullable.h"
-#include "vec/common/pod_array_fwd.h"
 #include "vec/core/block.h"
 #include "vec/core/column_with_type_and_name.h"
 #include "vec/exec/join/vhash_join_node.h"
@@ -71,6 +71,7 @@
 #include "vec/exec/vunion_node.h"
 #include "vec/exprs/vexpr.h"
 #include "vec/exprs/vexpr_context.h"
+#include "vec/runtime/vdatetime_value.h"
 #include "vec/utils/util.hpp"
 
 namespace doris {
diff --git a/be/src/exec/exec_node.h b/be/src/exec/exec_node.h
index d5b9688361..68f84934ad 100644
--- a/be/src/exec/exec_node.h
+++ b/be/src/exec/exec_node.h
@@ -35,31 +35,21 @@
 #include "common/global_types.h"
 #include "common/status.h"
 #include "runtime/descriptors.h"
-#include "runtime/query_statistics.h"
-#include "service/backend_options.h"
-#include "util/blocking_queue.hpp"
 #include "util/runtime_profile.h"
 #include "util/telemetry/telemetry.h"
 #include "vec/core/block.h"
-#include "vec/exprs/vexpr_context.h"
 
 namespace doris {
 class ObjectPool;
-class Counters;
 class RuntimeState;
-class TPlan;
 class MemTracker;
 class QueryStatistics;
 
 namespace vectorized {
-class Block;
-class VExpr;
 class VExprContext;
 } // namespace vectorized
 
 namespace pipeline {
-class PipelineFragmentContext;
-class Pipeline;
 class OperatorBase;
 } // namespace pipeline
 
diff --git a/be/src/exec/olap_common.h b/be/src/exec/olap_common.h
index e160d54fa3..2260dac25c 100644
--- a/be/src/exec/olap_common.h
+++ b/be/src/exec/olap_common.h
@@ -25,12 +25,10 @@
 #include <boost/lexical_cast.hpp>
 #include <cstdint>
 #include <iterator>
-#include <map>
 #include <memory>
 #include <set>
 #include <sstream>
 #include <string>
-#include <type_traits>
 #include <utility>
 #include <variant>
 #include <vector>
diff --git a/be/src/exec/olap_utils.h b/be/src/exec/olap_utils.h
index 610299cc0f..153d4f7963 100644
--- a/be/src/exec/olap_utils.h
+++ b/be/src/exec/olap_utils.h
@@ -17,10 +17,10 @@
 
 #pragma once
 
+#include <gen_cpp/Opcodes_types.h>
 #include <math.h>
 
 #include "common/logging.h"
-#include "gen_cpp/Opcodes_types.h"
 #include "olap/olap_tuple.h"
 #include "runtime/datetime_value.h"
 #include "runtime/primitive_type.h"
diff --git a/be/src/exec/rowid_fetcher.h b/be/src/exec/rowid_fetcher.h
index 19dabdf901..7f789b1c40 100644
--- a/be/src/exec/rowid_fetcher.h
+++ b/be/src/exec/rowid_fetcher.h
@@ -23,7 +23,6 @@
 #include <vector>
 
 #include "common/status.h"
-#include "vec/core/block.h"
 #include "vec/data_types/data_type.h"
 
 namespace doris {
diff --git a/be/src/exec/scan_node.h b/be/src/exec/scan_node.h
index aed701dbe9..6df382a975 100644
--- a/be/src/exec/scan_node.h
+++ b/be/src/exec/scan_node.h
@@ -20,8 +20,6 @@
 
 #pragma once
 
-#include <gen_cpp/PaloInternalService_types.h>
-
 #include <functional>
 #include <string>
 #include <vector>
@@ -32,7 +30,6 @@
 
 namespace doris {
 
-class TScanRange;
 class DescriptorTbl;
 class ObjectPool;
 class RuntimeState;
diff --git a/be/src/exec/schema_scanner.h b/be/src/exec/schema_scanner.h
index bc600bdf58..bc2d7bcfc0 100644
--- a/be/src/exec/schema_scanner.h
+++ b/be/src/exec/schema_scanner.h
@@ -18,7 +18,6 @@
 #pragma once
 
 #include <gen_cpp/Descriptors_types.h>
-#include <gen_cpp/Types_types.h>
 #include <stddef.h>
 #include <stdint.h>
 
@@ -26,11 +25,9 @@
 #include <string>
 #include <vector>
 
-#include "common/object_pool.h"
 #include "common/status.h"
 #include "runtime/define_primitive_type.h"
 #include "util/runtime_profile.h"
-#include "vec/core/block.h"
 
 namespace doris {
 
diff --git a/be/src/exec/schema_scanner/schema_helper.h b/be/src/exec/schema_scanner/schema_helper.h
index d6a340e1e4..72b7a9acf0 100644
--- a/be/src/exec/schema_scanner/schema_helper.h
+++ b/be/src/exec/schema_scanner/schema_helper.h
@@ -17,7 +17,6 @@
 
 #pragma once
 
-#include <gen_cpp/FrontendService_types.h>
 #include <stdint.h>
 
 #include <string>
diff --git a/be/src/exec/schema_scanner/schema_rowsets_scanner.h b/be/src/exec/schema_scanner/schema_rowsets_scanner.h
index 0a8dabc270..7a6b528ec2 100644
--- a/be/src/exec/schema_scanner/schema_rowsets_scanner.h
+++ b/be/src/exec/schema_scanner/schema_rowsets_scanner.h
@@ -19,14 +19,11 @@
 
 #include <cstddef>
 #include <cstdint>
-#include <memory>
 #include <vector>
 
 #include "common/status.h"
 #include "exec/schema_scanner.h"
 #include "olap/rowset/rowset.h"
-#include "olap/rowset/segment_v2/segment.h"
-#include "runtime/runtime_state.h"
 
 namespace doris {
 class RuntimeState;
diff --git a/be/src/exec/schema_scanner/schema_tables_scanner.h b/be/src/exec/schema_scanner/schema_tables_scanner.h
index c295fc6ba2..7cdb363774 100644
--- a/be/src/exec/schema_scanner/schema_tables_scanner.h
+++ b/be/src/exec/schema_scanner/schema_tables_scanner.h
@@ -23,7 +23,6 @@
 
 #include "common/status.h"
 #include "exec/schema_scanner.h"
-#include "vec/core/block.h"
 
 namespace doris {
 class RuntimeState;
diff --git a/be/src/exec/schema_scanner/schema_variables_scanner.h b/be/src/exec/schema_scanner/schema_variables_scanner.h
index 2728b70530..d55f234df1 100644
--- a/be/src/exec/schema_scanner/schema_variables_scanner.h
+++ b/be/src/exec/schema_scanner/schema_variables_scanner.h
@@ -20,8 +20,6 @@
 #include <gen_cpp/FrontendService_types.h>
 #include <gen_cpp/Types_types.h>
 
-#include <map>
-#include <string>
 #include <vector>
 
 #include "common/status.h"
diff --git a/be/src/exec/table_connector.cpp b/be/src/exec/table_connector.cpp
index ef0b9db8c3..0a469defca 100644
--- a/be/src/exec/table_connector.cpp
+++ b/be/src/exec/table_connector.cpp
@@ -17,14 +17,13 @@
 
 #include "exec/table_connector.h"
 
-#include <bthread/errno.h>
-#include <errno.h>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
 #include <gen_cpp/Metrics_types.h>
 #include <gen_cpp/Types_types.h>
 #include <glog/logging.h>
 #include <iconv.h>
 
-#include <cstdlib>
 #include <memory>
 #include <string_view>
 #include <type_traits>
diff --git a/be/src/exec/table_connector.h b/be/src/exec/table_connector.h
index 27d3061e75..8a1cdff7e3 100644
--- a/be/src/exec/table_connector.h
+++ b/be/src/exec/table_connector.h
@@ -21,18 +21,14 @@
 #include <gen_cpp/Types_types.h>
 #include <stdint.h>
 
-#include <boost/format.hpp>
-#include <cstdlib>
 #include <string>
 #include <vector>
 
 #include "common/status.h"
-#include "runtime/descriptors.h"
 #include "runtime/types.h"
 #include "util/runtime_profile.h"
 #include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/data_types/data_type.h"
-#include "vec/exprs/vexpr_context.h"
 
 namespace doris {
 class RuntimeState;
diff --git a/be/src/exec/tablet_info.h b/be/src/exec/tablet_info.h
index 6c9272787d..4840187e56 100644
--- a/be/src/exec/tablet_info.h
+++ b/be/src/exec/tablet_info.h
@@ -32,12 +32,9 @@
 
 #include "common/object_pool.h"
 #include "common/status.h"
-#include "olap/tablet_schema.h"
-#include "runtime/descriptors.h"
 #include "vec/columns/column.h"
 #include "vec/core/block.h"
 #include "vec/core/column_with_type_and_name.h"
-#include "vec/exprs/vexpr_context.h"
 
 namespace doris {
 class MemTracker;
diff --git a/be/src/exec/text_converter.cpp b/be/src/exec/text_converter.cpp
index 8249b42a98..1dcec677bf 100644
--- a/be/src/exec/text_converter.cpp
+++ b/be/src/exec/text_converter.cpp
@@ -21,9 +21,11 @@
 #include <sql.h>
 #include <stdint.h>
 
+#include <algorithm>
 #include <ostream>
 
-#include "common/compiler_util.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "olap/hll.h"
 #include "runtime/decimalv2_value.h"
 #include "runtime/define_primitive_type.h"
diff --git a/be/src/exec/text_converter.h b/be/src/exec/text_converter.h
index 0bf3b3f16e..d70cfba982 100644
--- a/be/src/exec/text_converter.h
+++ b/be/src/exec/text_converter.h
@@ -24,9 +24,6 @@
 namespace doris {
 
 class SlotDescriptor;
-class Status;
-struct StringRef;
-class TupleDescriptor;
 
 // Helper class for dealing with text data, e.g., converting text data to
 // numeric types, etc.
diff --git a/be/src/exprs/bloom_filter_func.h b/be/src/exprs/bloom_filter_func.h
index d127a4c19b..eb8ee79e0d 100644
--- a/be/src/exprs/bloom_filter_func.h
+++ b/be/src/exprs/bloom_filter_func.h
@@ -19,6 +19,7 @@
 
 #include "exprs/block_bloom_filter.hpp"
 #include "exprs/runtime_filter.h"
+#include "olap/rowset/segment_v2/bloom_filter.h" // IWYU pragma: keep
 
 namespace doris {
 
@@ -473,4 +474,4 @@ private:
     typename BloomFilterTypeTraits<type>::FindOp dummy;
 };
 
-} // namespace doris
\ No newline at end of file
+} // namespace doris
diff --git a/be/src/exprs/function_filter.h b/be/src/exprs/function_filter.h
index a9c2f264b4..9802f1b408 100644
--- a/be/src/exprs/function_filter.h
+++ b/be/src/exprs/function_filter.h
@@ -19,6 +19,7 @@
 #include <memory>
 
 #include "udf/udf.h"
+#include "vec/common/string_ref.h" // IWYU pragma: keep
 
 namespace doris {
 
diff --git a/be/src/exprs/json_functions.cpp b/be/src/exprs/json_functions.cpp
index ebfc4b0ae8..107217ed4f 100644
--- a/be/src/exprs/json_functions.cpp
+++ b/be/src/exprs/json_functions.cpp
@@ -32,7 +32,8 @@
 #include <string>
 #include <vector>
 
-#include "common/compiler_util.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/logging.h"
 
 namespace doris {
diff --git a/be/src/exprs/json_functions.h b/be/src/exprs/json_functions.h
index 2de0857219..fcec257142 100644
--- a/be/src/exprs/json_functions.h
+++ b/be/src/exprs/json_functions.h
@@ -17,10 +17,9 @@
 
 #pragma once
 
-#include <fmt/core.h>
 #include <fmt/format.h>
 #include <rapidjson/document.h>
-#include <simdjson.h>
+#include <simdjson.h> // IWYU pragma: keep
 
 #include <sstream>
 #include <string>
@@ -28,7 +27,6 @@
 #include <vector>
 
 #include "common/status.h"
-#include "udf/udf.h"
 
 namespace simdjson {
 namespace fallback {
@@ -49,8 +47,6 @@ enum JsonFunctionType {
     JSON_FUN_UNKNOWN //The last
 };
 
-class OpcodeRegistry;
-
 struct JsonPath {
     std::string key; // key of a json object
     int idx;         // array index of a json array, -1 means not set, -2 means *
diff --git a/be/src/exprs/math_functions.cpp b/be/src/exprs/math_functions.cpp
index e76cdde2a5..bbfdc6053d 100644
--- a/be/src/exprs/math_functions.cpp
+++ b/be/src/exprs/math_functions.cpp
@@ -22,12 +22,14 @@
 
 #include <stdlib.h>
 #include <string.h>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 // IWYU pragma: no_include <bits/std_abs.h>
 #include <cmath>
 #include <limits>
 #include <string_view>
 
-#include "common/compiler_util.h"
 #include "udf/udf.h"
 #include "util/string_parser.hpp"
 
diff --git a/be/src/exprs/runtime_filter.h b/be/src/exprs/runtime_filter.h
index 141df632c5..a76b02080f 100644
--- a/be/src/exprs/runtime_filter.h
+++ b/be/src/exprs/runtime_filter.h
@@ -50,15 +50,12 @@ class IOBufAsZeroCopyInputStream;
 namespace doris {
 class ObjectPool;
 class RuntimePredicateWrapper;
-class MemTracker;
 class PPublishFilterRequest;
 class PMergeFilterRequest;
 class TRuntimeFilterDesc;
 class RowDescriptor;
 class PInFilter;
 class PMinMaxFilter;
-class HashJoinNode;
-class RuntimeProfile;
 class BloomFilterFuncBase;
 class BitmapFilterFuncBase;
 class TNetworkAddress;
diff --git a/be/src/exprs/runtime_filter_slots.h b/be/src/exprs/runtime_filter_slots.h
index 71084bacfb..2d3b0c2c73 100644
--- a/be/src/exprs/runtime_filter_slots.h
+++ b/be/src/exprs/runtime_filter_slots.h
@@ -23,6 +23,7 @@
 #include "vec/columns/column_nullable.h"
 #include "vec/columns/columns_number.h"
 #include "vec/common/assert_cast.h"
+#include "vec/core/block.h" // IWYU pragma: keep
 #include "vec/runtime/shared_hash_table_controller.h"
 
 namespace doris {
diff --git a/be/src/exprs/string_functions.h b/be/src/exprs/string_functions.h
index 5a64790032..8ed917b82d 100644
--- a/be/src/exprs/string_functions.h
+++ b/be/src/exprs/string_functions.h
@@ -22,14 +22,9 @@
 
 #include <re2/re2.h>
 
-#include <iomanip>
-#include <locale>
 #include <memory>
-#include <sstream>
 #include <string>
-#include <string_view>
 
-#include "gutil/strings/numbers.h"
 #include "vec/common/string_ref.h"
 
 namespace doris {
diff --git a/be/src/geo/ByteOrderValues.h b/be/src/geo/ByteOrderValues.h
index c78f6c8856..50915953e0 100644
--- a/be/src/geo/ByteOrderValues.h
+++ b/be/src/geo/ByteOrderValues.h
@@ -19,8 +19,6 @@
 
 #include <sys/types.h>
 
-#include <ctime>
-
 using int64_t = __int64_t;
 using int32_t = __int32_t;
 using uint32_t = __uint32_t;
diff --git a/be/src/geo/geo_tobinary.h b/be/src/geo/geo_tobinary.h
index 8b861a519a..ad4ee9d0a4 100644
--- a/be/src/geo/geo_tobinary.h
+++ b/be/src/geo/geo_tobinary.h
@@ -21,10 +21,6 @@
 
 #include <string>
 
-#include "geo/geo_common.h"
-#include "geo/geo_tobinary_type.h"
-#include "geo/wkt_parse_type.h"
-
 struct ToBinaryContext;
 
 namespace doris {
diff --git a/be/src/geo/geo_types.h b/be/src/geo/geo_types.h
index 7416dccb35..e3d16d67c1 100644
--- a/be/src/geo/geo_types.h
+++ b/be/src/geo/geo_types.h
@@ -21,7 +21,6 @@
 
 #include <memory>
 #include <string>
-#include <vector>
 
 #include "geo/geo_common.h"
 #include "geo/wkt_parse_type.h"
diff --git a/be/src/geo/wkb_parse.h b/be/src/geo/wkb_parse.h
index c27537966b..966e290733 100644
--- a/be/src/geo/wkb_parse.h
+++ b/be/src/geo/wkb_parse.h
@@ -22,9 +22,7 @@
 #include <iosfwd>
 
 #include "geo/geo_common.h"
-#include "geo/geo_types.h"
 #include "geo/wkt_parse_type.h"
-#include "wkb_parse_ctx.h"
 
 struct WkbParseContext;
 
diff --git a/be/src/geo/wkt_parse.h b/be/src/geo/wkt_parse.h
index 0bb999af73..9dfcca98dd 100644
--- a/be/src/geo/wkt_parse.h
+++ b/be/src/geo/wkt_parse.h
@@ -19,8 +19,6 @@
 
 #include <stddef.h>
 
-#include <memory>
-
 #include "geo/geo_common.h"
 
 namespace doris {
diff --git a/be/src/gutil/atomicops.h b/be/src/gutil/atomicops.h
index 7873a39170..4c6d93e28c 100644
--- a/be/src/gutil/atomicops.h
+++ b/be/src/gutil/atomicops.h
@@ -72,13 +72,13 @@
 
 // ThreadSanitizer provides own implementation of atomicops.
 #if defined(THREAD_SANITIZER)
-#include "gutil/atomicops-internals-tsan.h"
+#include "gutil/atomicops-internals-tsan.h" // IWYU pragma: export
 #elif defined(__GNUC__) && (defined(__i386) || defined(__x86_64__))
-#include "gutil/atomicops-internals-x86.h"
+#include "gutil/atomicops-internals-x86.h" // IWYU pragma: export
 #elif defined(__GNUC__) && GCC_VERSION >= 40700
-#include "gutil/atomicops-internals-gcc.h"
+#include "gutil/atomicops-internals-gcc.h" // IWYU pragma: export
 #elif defined(__clang__) && CLANG_VERSION >= 30400
-#include "gutil/atomicops-internals-gcc.h"
+#include "gutil/atomicops-internals-gcc.h" // IWYU pragma: export
 #else
 #error You need to implement atomic operations for this architecture
 #endif
diff --git a/be/src/http/action/check_tablet_segment_action.h b/be/src/http/action/check_tablet_segment_action.h
index a4c6f422ee..0cc26e8199 100644
--- a/be/src/http/action/check_tablet_segment_action.h
+++ b/be/src/http/action/check_tablet_segment_action.h
@@ -20,7 +20,6 @@
 #include <string>
 
 #include "http/http_handler.h"
-#include "util/easy_json.h"
 
 namespace doris {
 class HttpRequest;
diff --git a/be/src/http/action/checksum_action.h b/be/src/http/action/checksum_action.h
index 5afacbfe6d..4ec478acfc 100644
--- a/be/src/http/action/checksum_action.h
+++ b/be/src/http/action/checksum_action.h
@@ -23,7 +23,6 @@
 
 namespace doris {
 
-class ExecEnv;
 class HttpRequest;
 
 class ChecksumAction : public HttpHandler {
diff --git a/be/src/http/action/download_action.h b/be/src/http/action/download_action.h
index 86bee3e2b1..3965461844 100644
--- a/be/src/http/action/download_action.h
+++ b/be/src/http/action/download_action.h
@@ -21,9 +21,7 @@
 #include <vector>
 
 #include "common/status.h"
-#include "exec/scan_node.h"
 #include "http/http_handler.h"
-#include "runtime/descriptors.h"
 
 namespace doris {
 
diff --git a/be/src/http/action/health_action.h b/be/src/http/action/health_action.h
index cdee0f12b4..6724ef0856 100644
--- a/be/src/http/action/health_action.h
+++ b/be/src/http/action/health_action.h
@@ -21,7 +21,6 @@
 
 namespace doris {
 
-class ExecEnv;
 class HttpRequest;
 
 // Get BE health state from http API.
diff --git a/be/src/http/action/meta_action.h b/be/src/http/action/meta_action.h
index c4809cd3b5..fe59ed2744 100644
--- a/be/src/http/action/meta_action.h
+++ b/be/src/http/action/meta_action.h
@@ -24,7 +24,6 @@
 
 namespace doris {
 
-class ExecEnv;
 class HttpRequest;
 
 // Get Meta Info
diff --git a/be/src/http/action/metrics_action.h b/be/src/http/action/metrics_action.h
index 25664a98aa..2051aed53a 100644
--- a/be/src/http/action/metrics_action.h
+++ b/be/src/http/action/metrics_action.h
@@ -21,8 +21,6 @@
 
 namespace doris {
 
-class Webserver;
-class ExecEnv;
 class HttpRequest;
 class MetricRegistry;
 
diff --git a/be/src/http/action/monitor_action.h b/be/src/http/action/monitor_action.h
index 8d3f1c87d7..b4bbe7a09c 100644
--- a/be/src/http/action/monitor_action.h
+++ b/be/src/http/action/monitor_action.h
@@ -25,7 +25,6 @@
 namespace doris {
 
 class HttpRequest;
-class HttpChannel;
 class RestMonitorIface;
 
 class MonitorAction : public HttpHandler {
diff --git a/be/src/http/action/pad_rowset_action.h b/be/src/http/action/pad_rowset_action.h
index 13090dcd8c..26ff6b6854 100644
--- a/be/src/http/action/pad_rowset_action.h
+++ b/be/src/http/action/pad_rowset_action.h
@@ -19,7 +19,6 @@
 
 #include "common/status.h"
 #include "http/http_handler.h"
-#include "http/http_request.h"
 #include "olap/tablet.h"
 
 namespace doris {
diff --git a/be/src/http/action/pprof_actions.cpp b/be/src/http/action/pprof_actions.cpp
index 56ae5e49e9..1d71281273 100644
--- a/be/src/http/action/pprof_actions.cpp
+++ b/be/src/http/action/pprof_actions.cpp
@@ -25,6 +25,7 @@
 #include <fstream>
 #include <memory>
 #include <mutex>
+#include <string>
 
 #include "common/config.h"
 #include "common/object_pool.h"
diff --git a/be/src/http/action/reload_tablet_action.h b/be/src/http/action/reload_tablet_action.h
index aa36a474dd..8ecc946bb0 100644
--- a/be/src/http/action/reload_tablet_action.h
+++ b/be/src/http/action/reload_tablet_action.h
@@ -17,7 +17,6 @@
 
 #pragma once
 
-#include <gen_cpp/AgentService_types.h>
 #include <stdint.h>
 
 #include <string>
diff --git a/be/src/http/action/restore_tablet_action.cpp b/be/src/http/action/restore_tablet_action.cpp
index d710a2b36b..7ff9aa37f3 100644
--- a/be/src/http/action/restore_tablet_action.cpp
+++ b/be/src/http/action/restore_tablet_action.cpp
@@ -21,6 +21,7 @@
 #include <gen_cpp/AgentService_types.h>
 #include <glog/logging.h>
 
+#include <algorithm>
 #include <cstdlib>
 #include <filesystem>
 #include <memory>
diff --git a/be/src/http/action/restore_tablet_action.h b/be/src/http/action/restore_tablet_action.h
index a11d7783b7..2eccb22303 100644
--- a/be/src/http/action/restore_tablet_action.h
+++ b/be/src/http/action/restore_tablet_action.h
@@ -17,7 +17,6 @@
 
 #pragma once
 
-#include <gen_cpp/AgentService_types.h>
 #include <stdint.h>
 
 #include <map>
diff --git a/be/src/http/action/snapshot_action.h b/be/src/http/action/snapshot_action.h
index 0d1399604d..b1b58bee10 100644
--- a/be/src/http/action/snapshot_action.h
+++ b/be/src/http/action/snapshot_action.h
@@ -24,7 +24,6 @@
 
 namespace doris {
 
-class ExecEnv;
 class HttpRequest;
 
 // make snapshot
diff --git a/be/src/http/action/stream_load.h b/be/src/http/action/stream_load.h
index 3c44a77bb4..4319d722d1 100644
--- a/be/src/http/action/stream_load.h
+++ b/be/src/http/action/stream_load.h
@@ -17,15 +17,10 @@
 
 #pragma once
 
-#include <gen_cpp/PlanNodes_types.h>
-
-#include <functional>
 #include <memory>
 #include <string>
 
 #include "http/http_handler.h"
-#include "runtime/client_cache.h"
-#include "runtime/message_body_sink.h"
 #include "util/metrics.h"
 
 namespace doris {
diff --git a/be/src/http/action/tablet_migration_action.h b/be/src/http/action/tablet_migration_action.h
index aee9a65d12..11933e7fc7 100644
--- a/be/src/http/action/tablet_migration_action.h
+++ b/be/src/http/action/tablet_migration_action.h
@@ -17,7 +17,6 @@
 
 #pragma once
 
-#include <gen_cpp/Status_types.h>
 #include <stdint.h>
 
 #include <deque>
@@ -29,10 +28,8 @@
 #include <utility>
 
 #include "common/status.h"
-#include "gutil/stringprintf.h"
-#include "gutil/strings/numbers.h"
+#include "gutil/strings/substitute.h"
 #include "http/http_handler.h"
-#include "olap/data_dir.h"
 #include "olap/tablet.h"
 #include "util/threadpool.h"
 
diff --git a/be/src/http/action/tablets_info_action.cpp b/be/src/http/action/tablets_info_action.cpp
index 656fbdf158..374f7b2136 100644
--- a/be/src/http/action/tablets_info_action.cpp
+++ b/be/src/http/action/tablets_info_action.cpp
@@ -26,7 +26,7 @@
 #include <string>
 #include <vector>
 
-#include "gutil/strings/numbers.h"
+#include "gutil/strings/substitute.h"
 #include "http/http_channel.h"
 #include "http/http_headers.h"
 #include "http/http_request.h"
diff --git a/be/src/http/action/version_action.h b/be/src/http/action/version_action.h
index e78e19f37c..ed9a48ec61 100644
--- a/be/src/http/action/version_action.h
+++ b/be/src/http/action/version_action.h
@@ -22,7 +22,6 @@
 
 namespace doris {
 
-class ExecEnv;
 class HttpRequest;
 
 // Get BE version info from http API.
diff --git a/be/src/http/default_path_handlers.h b/be/src/http/default_path_handlers.h
index 1ff038350b..f9d022e793 100644
--- a/be/src/http/default_path_handlers.h
+++ b/be/src/http/default_path_handlers.h
@@ -17,13 +17,8 @@
 
 #pragma once
 
-#include <stdio.h>
-
-#include <memory>
-
 namespace doris {
 
-class MemTracker;
 class WebPageHandler;
 
 // Adds a set of default path handlers to the webserver to display
diff --git a/be/src/http/ev_http_server.cpp b/be/src/http/ev_http_server.cpp
index ccea181a8f..b0743baee3 100644
--- a/be/src/http/ev_http_server.cpp
+++ b/be/src/http/ev_http_server.cpp
@@ -18,9 +18,10 @@
 #include "http/ev_http_server.h"
 
 #include <arpa/inet.h>
-#include <bthread/errno.h>
 #include <butil/endpoint.h>
 #include <butil/fd_utility.h>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
 #include <event2/event.h>
 #include <event2/http.h>
 #include <event2/http_struct.h>
diff --git a/be/src/http/ev_http_server.h b/be/src/http/ev_http_server.h
index 5ba634b219..bf6b200bdc 100644
--- a/be/src/http/ev_http_server.h
+++ b/be/src/http/ev_http_server.h
@@ -20,7 +20,6 @@
 #include <memory>
 #include <mutex>
 #include <string>
-#include <thread>
 #include <vector>
 
 #include "common/status.h"
diff --git a/be/src/http/http_channel.h b/be/src/http/http_channel.h
index a12eddb535..478f013af8 100644
--- a/be/src/http/http_channel.h
+++ b/be/src/http/http_channel.h
@@ -19,13 +19,10 @@
 
 #include <stddef.h>
 
-#include <cstdint>
 #include <string>
 
 #include "http/http_status.h"
 
-struct mg_connection;
-
 namespace doris {
 
 class HttpRequest;
diff --git a/be/src/http/http_client.h b/be/src/http/http_client.h
index 7ec714587a..c194c29c88 100644
--- a/be/src/http/http_client.h
+++ b/be/src/http/http_client.h
@@ -26,10 +26,7 @@
 #include <string>
 
 #include "common/status.h"
-#include "http/http_headers.h"
 #include "http/http_method.h"
-#include "http/http_response.h"
-#include "http/utils.h"
 
 namespace doris {
 
diff --git a/be/src/http/http_request.h b/be/src/http/http_request.h
index a26be3a22c..a9286410af 100644
--- a/be/src/http/http_request.h
+++ b/be/src/http/http_request.h
@@ -19,17 +19,13 @@
 
 #include <glog/logging.h>
 
-#include <boost/algorithm/string.hpp>
 #include <map>
 #include <memory>
 #include <string>
 
-#include "http/http_common.h"
-#include "http/http_headers.h"
 #include "http/http_method.h"
 #include "util/string_util.h"
 
-struct mg_connection;
 struct evhttp_request;
 
 namespace doris {
diff --git a/be/src/http/utils.h b/be/src/http/utils.h
index f9b09b3017..dd62a9b8a9 100644
--- a/be/src/http/utils.h
+++ b/be/src/http/utils.h
@@ -19,13 +19,10 @@
 
 #include <string>
 
-#include "common/utils.h"
-#include "http/http_common.h"
 #include "http/http_request.h"
 
 namespace doris {
 
-class HttpRequest;
 struct AuthInfo;
 
 std::string encode_basic_auth(const std::string& user, const std::string& passwd);
diff --git a/be/src/http/web_page_handler.h b/be/src/http/web_page_handler.h
index 9ab304d96f..3fab082847 100644
--- a/be/src/http/web_page_handler.h
+++ b/be/src/http/web_page_handler.h
@@ -23,10 +23,8 @@
 #include <sstream>
 #include <string>
 #include <utility>
-#include <vector>
 
 #include "http/http_handler.h"
-#include "util/easy_json.h"
 
 namespace doris {
 
diff --git a/be/src/io/cache/block/block_file_cache.cpp b/be/src/io/cache/block/block_file_cache.cpp
index 0b97facff5..9868c5fee3 100644
--- a/be/src/io/cache/block/block_file_cache.cpp
+++ b/be/src/io/cache/block/block_file_cache.cpp
@@ -20,7 +20,11 @@
 
 #include "io/cache/block/block_file_cache.h"
 
+#include <glog/logging.h>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
 #include <filesystem>
+#include <utility>
 
 #include "io/cache/block/block_file_cache_fwd.h"
 #include "io/cache/block/block_file_cache_settings.h"
diff --git a/be/src/io/cache/block/block_file_cache.h b/be/src/io/cache/block/block_file_cache.h
index 81050c49b4..965a510717 100644
--- a/be/src/io/cache/block/block_file_cache.h
+++ b/be/src/io/cache/block/block_file_cache.h
@@ -20,20 +20,32 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <stdint.h>
+
+#include <cstddef>
 #include <list>
 #include <memory>
+#include <mutex>
+#include <string>
+#include <tuple>
 #include <unordered_map>
+#include <vector>
 
 #include "common/config.h"
+#include "common/status.h"
 #include "io/cache/block/block_file_cache_fwd.h"
+#include "util/hash_util.hpp"
+#include "vec/common/uint128.h"
 
 namespace doris {
 namespace io {
 class FileBlock;
+struct FileCacheSettings;
+
 using FileBlockSPtr = std::shared_ptr<FileBlock>;
 using FileBlocks = std::list<FileBlockSPtr>;
 struct FileBlocksHolder;
-struct ReadSettings;
 
 /**
  * Local cache for remote filesystem files, represented as a set of non-overlapping non-empty file segments.
diff --git a/be/src/io/cache/block/block_file_cache_factory.cpp b/be/src/io/cache/block/block_file_cache_factory.cpp
index 2aee79fc50..5c6c569f21 100644
--- a/be/src/io/cache/block/block_file_cache_factory.cpp
+++ b/be/src/io/cache/block/block_file_cache_factory.cpp
@@ -20,14 +20,21 @@
 
 #include "io/cache/block/block_file_cache_factory.h"
 
-#include <cstddef>
+#include <glog/logging.h>
+
+#include <algorithm>
+#include <ostream>
+#include <utility>
 
 #include "common/config.h"
 #include "io/cache/block/block_file_cache.h"
+#include "io/cache/block/block_file_cache_settings.h"
 #include "io/cache/block/block_lru_file_cache.h"
 #include "io/fs/local_file_system.h"
 
 namespace doris {
+class TUniqueId;
+
 namespace io {
 
 FileCacheFactory& FileCacheFactory::instance() {
diff --git a/be/src/io/cache/block/block_file_cache_factory.h b/be/src/io/cache/block/block_file_cache_factory.h
index e37a6eebf0..62c23335b7 100644
--- a/be/src/io/cache/block/block_file_cache_factory.h
+++ b/be/src/io/cache/block/block_file_cache_factory.h
@@ -20,13 +20,18 @@
 
 #pragma once
 
+#include <memory>
+#include <string>
 #include <vector>
 
+#include "common/status.h"
 #include "io/cache/block/block_file_cache.h"
-#include "io/cache/block/block_file_cache_fwd.h"
-#include "io/cache/block/block_file_cache_settings.h"
+
 namespace doris {
+class TUniqueId;
+
 namespace io {
+struct FileCacheSettings;
 
 enum FileCacheType {
     NORMAL,
diff --git a/be/src/io/cache/block/block_file_cache_profile.cpp b/be/src/io/cache/block/block_file_cache_profile.cpp
index 61b0ef72ce..ef245be350 100644
--- a/be/src/io/cache/block/block_file_cache_profile.cpp
+++ b/be/src/io/cache/block/block_file_cache_profile.cpp
@@ -17,7 +17,9 @@
 
 #include "io/cache/block/block_file_cache_profile.h"
 
+#include <functional>
 #include <memory>
+#include <string>
 
 namespace doris {
 namespace io {
diff --git a/be/src/io/cache/block/block_file_cache_profile.h b/be/src/io/cache/block/block_file_cache_profile.h
index 858a07aed5..7635413cd9 100644
--- a/be/src/io/cache/block/block_file_cache_profile.h
+++ b/be/src/io/cache/block/block_file_cache_profile.h
@@ -17,15 +17,15 @@
 
 #pragma once
 
+#include <gen_cpp/Metrics_types.h>
+
 #include <atomic>
 #include <cstdint>
-#include <cstdlib>
-#include <functional>
 #include <memory>
 #include <mutex>
-#include <string>
 #include <unordered_map>
 
+#include "io/io_common.h"
 #include "olap/olap_common.h"
 #include "util/doris_metrics.h"
 #include "util/metrics.h"
diff --git a/be/src/io/cache/block/block_file_segment.cpp b/be/src/io/cache/block/block_file_segment.cpp
index 020fc85de1..46a1cedc53 100644
--- a/be/src/io/cache/block/block_file_segment.cpp
+++ b/be/src/io/cache/block/block_file_segment.cpp
@@ -20,16 +20,17 @@
 
 #include "io/cache/block/block_file_segment.h"
 
-#include <filesystem>
+#include <glog/logging.h>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
 #include <sstream>
 #include <string>
 #include <thread>
 
 #include "common/status.h"
+#include "io/fs/file_reader.h"
 #include "io/fs/file_writer.h"
 #include "io/fs/local_file_system.h"
-#include "olap/iterators.h"
-#include "vec/common/hex.h"
 
 namespace doris {
 namespace io {
diff --git a/be/src/io/cache/block/block_file_segment.h b/be/src/io/cache/block/block_file_segment.h
index 47732cb70e..7cbc238f93 100644
--- a/be/src/io/cache/block/block_file_segment.h
+++ b/be/src/io/cache/block/block_file_segment.h
@@ -20,20 +20,28 @@
 
 #pragma once
 
+#include <fmt/format.h>
+#include <stddef.h>
+
+#include <atomic>
 #include <condition_variable>
 #include <list>
 #include <memory>
 #include <mutex>
+#include <string>
+#include <utility>
 
 #include "common/status.h"
 #include "io/cache/block/block_file_cache.h"
-#include "io/fs/file_reader.h"
 #include "io/fs/file_writer.h"
+#include "util/slice.h"
 
 namespace doris {
 namespace io {
 
 class FileBlock;
+class FileReader;
+
 using FileBlockSPtr = std::shared_ptr<FileBlock>;
 using FileBlocks = std::list<FileBlockSPtr>;
 
diff --git a/be/src/io/cache/block/block_lru_file_cache.cpp b/be/src/io/cache/block/block_lru_file_cache.cpp
index c13e74664f..431410c42e 100644
--- a/be/src/io/cache/block/block_lru_file_cache.cpp
+++ b/be/src/io/cache/block/block_lru_file_cache.cpp
@@ -20,19 +20,43 @@
 
 #include "io/cache/block/block_lru_file_cache.h"
 
+#include <fmt/format.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <stdint.h>
+
+#include <algorithm>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <cstring>
 #include <filesystem>
+#include <iterator>
+#include <list>
+#include <ostream>
 #include <random>
 #include <system_error>
 #include <utility>
 
 #include "common/status.h"
 #include "io/cache/block/block_file_cache.h"
-#include "io/cache/block/block_file_cache_settings.h"
+#include "io/cache/block/block_file_cache_fwd.h"
+#include "io/fs/file_reader.h"
+#include "io/fs/file_reader_writer_fwd.h"
+#include "io/fs/file_system.h"
+#include "io/fs/file_writer.h"
 #include "io/fs/local_file_system.h"
-#include "olap/iterators.h"
-#include "util/time.h"
+#include "io/fs/path.h"
+#include "util/slice.h"
 #include "vec/common/hex.h"
-#include "vec/common/sip_hash.h"
+
+namespace doris {
+namespace io {
+struct FileCacheSettings;
+} // namespace io
+} // namespace doris
 
 namespace fs = std::filesystem;
 
diff --git a/be/src/io/cache/block/block_lru_file_cache.h b/be/src/io/cache/block/block_lru_file_cache.h
index ac6c09371e..6742a8c0d6 100644
--- a/be/src/io/cache/block/block_lru_file_cache.h
+++ b/be/src/io/cache/block/block_lru_file_cache.h
@@ -20,17 +20,25 @@
 
 #pragma once
 
-#include <chrono>
+#include <cstddef>
 #include <map>
 #include <memory>
+#include <mutex>
 #include <optional>
+#include <string>
 #include <unordered_map>
+#include <utility>
+#include <vector>
 
+#include "common/status.h"
 #include "io/cache/block/block_file_cache.h"
 #include "io/cache/block/block_file_segment.h"
 
 namespace doris {
+class TUniqueId;
+
 namespace io {
+struct FileCacheSettings;
 
 /**
  * Local cache for remote filesystem files, represented as a set of non-overlapping non-empty file segments.
diff --git a/be/src/io/cache/block/cached_remote_file_reader.cpp b/be/src/io/cache/block/cached_remote_file_reader.cpp
index e456811430..1c35375526 100644
--- a/be/src/io/cache/block/cached_remote_file_reader.cpp
+++ b/be/src/io/cache/block/cached_remote_file_reader.cpp
@@ -17,12 +17,26 @@
 
 #include "io/cache/block/cached_remote_file_reader.h"
 
+#include <fmt/format.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <string.h>
+
+#include <algorithm>
+#include <list>
+#include <vector>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/config.h"
 #include "io/cache/block/block_file_cache.h"
 #include "io/cache/block/block_file_cache_factory.h"
+#include "io/cache/block/block_file_segment.h"
 #include "io/fs/file_reader.h"
-#include "olap/iterators.h"
-#include "olap/olap_common.h"
-#include "util/async_io.h"
+#include "io/io_common.h"
+#include "util/bit_util.h"
+#include "util/doris_metrics.h"
+#include "util/runtime_profile.h"
 
 namespace doris {
 namespace io {
diff --git a/be/src/io/cache/block/cached_remote_file_reader.h b/be/src/io/cache/block/cached_remote_file_reader.h
index 95fe00e936..6d9323f5c8 100644
--- a/be/src/io/cache/block/cached_remote_file_reader.h
+++ b/be/src/io/cache/block/cached_remote_file_reader.h
@@ -17,17 +17,25 @@
 
 #pragma once
 
-#include "gutil/macros.h"
+#include <stddef.h>
+#include <stdint.h>
+
+#include <memory>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
 #include "io/cache/block/block_file_cache.h"
-#include "io/cache/block/block_file_cache_fwd.h"
-#include "io/cache/block/block_file_cache_profile.h"
-#include "io/cache/block/block_file_segment.h"
 #include "io/fs/file_reader.h"
+#include "io/fs/file_reader_writer_fwd.h"
+#include "io/fs/file_system.h"
 #include "io/fs/path.h"
-#include "io/fs/s3_file_system.h"
+#include "util/slice.h"
 
 namespace doris {
 namespace io {
+class IOContext;
+struct FileCacheStatistics;
 
 class CachedRemoteFileReader final : public FileReader {
 public:
diff --git a/be/src/io/cache/dummy_file_cache.cpp b/be/src/io/cache/dummy_file_cache.cpp
index 5150fa1db7..14dda6898a 100644
--- a/be/src/io/cache/dummy_file_cache.cpp
+++ b/be/src/io/cache/dummy_file_cache.cpp
@@ -17,9 +17,12 @@
 
 #include "io/cache/dummy_file_cache.h"
 
-#include "gutil/strings/util.h"
+#include <time.h>
+
+#include <memory>
+#include <string>
+
 #include "io/fs/local_file_system.h"
-#include "util/string_util.h"
 
 namespace doris {
 namespace io {
diff --git a/be/src/io/cache/dummy_file_cache.h b/be/src/io/cache/dummy_file_cache.h
index 653d47b1d6..46d21d99a2 100644
--- a/be/src/io/cache/dummy_file_cache.h
+++ b/be/src/io/cache/dummy_file_cache.h
@@ -17,17 +17,23 @@
 
 #pragma once
 
-#include <future>
-#include <memory>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <filesystem>
 #include <queue>
+#include <vector>
 
 #include "common/status.h"
 #include "io/cache/file_cache.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "io/fs/file_system.h"
 #include "io/fs/path.h"
+#include "util/slice.h"
 
 namespace doris {
 namespace io {
+class IOContext;
 
 // Only used for GC
 class DummyFileCache final : public FileCache {
diff --git a/be/src/io/cache/file_cache.cpp b/be/src/io/cache/file_cache.cpp
index 91379a2f1d..d37eba7724 100644
--- a/be/src/io/cache/file_cache.cpp
+++ b/be/src/io/cache/file_cache.cpp
@@ -17,12 +17,25 @@
 
 #include "io/cache/file_cache.h"
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <string.h>
+
+#include <algorithm>
+#include <filesystem>
+#include <list>
+#include <ostream>
+#include <set>
+#include <utility>
+
 #include "common/config.h"
 #include "common/status.h"
 #include "gutil/strings/util.h"
+#include "io/fs/file_system.h"
+#include "io/fs/file_writer.h"
 #include "io/fs/local_file_system.h"
-#include "io/fs/local_file_writer.h"
-#include "olap/iterators.h"
+#include "runtime/exec_env.h"
+#include "util/string_util.h"
 
 namespace doris {
 using namespace ErrorCode;
diff --git a/be/src/io/cache/file_cache.h b/be/src/io/cache/file_cache.h
index 68ffc1a3a1..2b2156ea46 100644
--- a/be/src/io/cache/file_cache.h
+++ b/be/src/io/cache/file_cache.h
@@ -17,17 +17,23 @@
 
 #pragma once
 
+#include <butil/macros.h>
+#include <stddef.h>
+#include <stdint.h>
+
 #include <memory>
-#include <queue>
-#include <shared_mutex>
 #include <string>
+#include <vector>
 
 #include "common/status.h"
 #include "io/fs/file_reader.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "io/fs/path.h"
+#include "util/slice.h"
 
 namespace doris {
 namespace io {
+class IOContext;
 
 const std::string CACHE_DONE_FILE_SUFFIX = "_DONE";
 
diff --git a/be/src/io/cache/file_cache_manager.cpp b/be/src/io/cache/file_cache_manager.cpp
index b14a02c1cb..f581b7d651 100644
--- a/be/src/io/cache/file_cache_manager.cpp
+++ b/be/src/io/cache/file_cache_manager.cpp
@@ -17,14 +17,27 @@
 
 #include "io/cache/file_cache_manager.h"
 
-#include "gutil/strings/util.h"
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <memory>
+#include <mutex>
+#include <ostream>
+#include <utility>
+
+#include "common/config.h"
 #include "io/cache/dummy_file_cache.h"
 #include "io/cache/sub_file_cache.h"
 #include "io/cache/whole_file_cache.h"
+#include "io/fs/file_reader_options.h"
+#include "io/fs/file_system.h"
 #include "io/fs/local_file_system.h"
+#include "olap/data_dir.h"
 #include "olap/rowset/beta_rowset.h"
 #include "olap/storage_engine.h"
-#include "util/string_util.h"
+#include "olap/tablet.h"
+#include "olap/tablet_manager.h"
 
 namespace doris {
 namespace io {
diff --git a/be/src/io/cache/file_cache_manager.h b/be/src/io/cache/file_cache_manager.h
index 6bf49c0dc7..7ac5e652d4 100644
--- a/be/src/io/cache/file_cache_manager.h
+++ b/be/src/io/cache/file_cache_manager.h
@@ -17,17 +17,22 @@
 
 #pragma once
 
-#include <list>
-#include <memory>
-#include <queue>
+#include <stdint.h>
+
+#include <list>
+#include <map>
+#include <queue>
+#include <shared_mutex>
+#include <string>
+#include <vector>
 
-#include "common/config.h"
 #include "common/status.h"
 #include "io/cache/file_cache.h"
-#include "io/fs/file_reader_options.h"
+#include "io/fs/file_reader_writer_fwd.h"
 
 namespace doris {
 namespace io {
+enum class FileCachePolicy : uint8_t;
 
 class GCContextPerDisk {
 public:
diff --git a/be/src/io/cache/sub_file_cache.cpp b/be/src/io/cache/sub_file_cache.cpp
index fc075bf490..d7d09f3448 100644
--- a/be/src/io/cache/sub_file_cache.cpp
+++ b/be/src/io/cache/sub_file_cache.cpp
@@ -17,20 +17,27 @@
 
 #include "io/cache/sub_file_cache.h"
 
+#include <fmt/format.h>
 #include <glog/logging.h>
-#include <sys/types.h>
+#include <time.h>
 
 #include <algorithm>
 #include <cstdlib>
+#include <filesystem>
+#include <future>
+#include <mutex>
+#include <ostream>
+#include <string>
 #include <utility>
 #include <vector>
 
 #include "common/config.h"
 #include "common/status.h"
 #include "io/fs/local_file_system.h"
-#include "olap/iterators.h"
-#include "util/async_io.h"
+#include "io/io_common.h"
+#include "runtime/exec_env.h"
 #include "util/string_util.h"
+#include "util/threadpool.h"
 
 namespace doris {
 using namespace ErrorCode;
diff --git a/be/src/io/cache/sub_file_cache.h b/be/src/io/cache/sub_file_cache.h
index 4766a4b6a6..96dc873a88 100644
--- a/be/src/io/cache/sub_file_cache.h
+++ b/be/src/io/cache/sub_file_cache.h
@@ -17,16 +17,28 @@
 
 #pragma once
 
-#include <future>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <map>
 #include <memory>
+#include <queue>
+#include <shared_mutex>
+#include <utility>
+#include <vector>
 
 #include "common/status.h"
 #include "io/cache/file_cache.h"
+#include "io/fs/file_reader.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "io/fs/file_system.h"
 #include "io/fs/path.h"
+#include "util/slice.h"
 
 namespace doris {
 namespace io {
+class IOContext;
 
 class SubFileCache final : public FileCache {
 public:
diff --git a/be/src/io/cache/whole_file_cache.cpp b/be/src/io/cache/whole_file_cache.cpp
index ef24295e68..59494d9cbd 100644
--- a/be/src/io/cache/whole_file_cache.cpp
+++ b/be/src/io/cache/whole_file_cache.cpp
@@ -17,9 +17,19 @@
 
 #include "io/cache/whole_file_cache.h"
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+
+#include <filesystem>
+#include <future>
+#include <mutex>
+#include <ostream>
+#include <string>
+
 #include "io/fs/local_file_system.h"
-#include "olap/iterators.h"
-#include "util/async_io.h"
+#include "io/io_common.h"
+#include "runtime/exec_env.h"
+#include "util/threadpool.h"
 
 namespace doris {
 using namespace ErrorCode;
diff --git a/be/src/io/cache/whole_file_cache.h b/be/src/io/cache/whole_file_cache.h
index 0036678114..866e6e142e 100644
--- a/be/src/io/cache/whole_file_cache.h
+++ b/be/src/io/cache/whole_file_cache.h
@@ -17,16 +17,23 @@
 
 #pragma once
 
-#include <future>
+#include <stdint.h>
+#include <time.h>
+
 #include <memory>
+#include <shared_mutex>
 
 #include "common/status.h"
 #include "io/cache/file_cache.h"
+#include "io/fs/file_reader.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "io/fs/file_system.h"
 #include "io/fs/path.h"
+#include "util/slice.h"
 
 namespace doris {
 namespace io {
+class IOContext;
 
 class WholeFileCache final : public FileCache {
 public:
diff --git a/be/src/io/file_factory.cpp b/be/src/io/file_factory.cpp
index c539f9ae50..e8fe4b2cec 100644
--- a/be/src/io/file_factory.cpp
+++ b/be/src/io/file_factory.cpp
@@ -17,24 +17,34 @@
 
 #include "io/file_factory.h"
 
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Types_types.h>
+
+#include <utility>
+
 #include "common/config.h"
 #include "common/status.h"
 #include "io/fs/broker_file_system.h"
-#include "io/fs/broker_file_writer.h"
 #include "io/fs/file_reader_options.h"
-#include "io/fs/file_system.h"
 #include "io/fs/hdfs_file_system.h"
-#include "io/fs/hdfs_file_writer.h"
 #include "io/fs/local_file_system.h"
 #include "io/fs/s3_file_system.h"
+#include "io/fs/stream_load_pipe.h"
+#include "io/hdfs_builder.h"
 #include "runtime/exec_env.h"
 #include "runtime/runtime_state.h"
 #include "runtime/stream_load/new_load_stream_mgr.h"
 #include "runtime/stream_load/stream_load_context.h"
-#include "util/runtime_profile.h"
 #include "util/s3_uri.h"
+#include "util/s3_util.h"
+#include "util/uid_util.h"
 
 namespace doris {
+namespace io {
+class FileWriter;
+} // namespace io
+
 io::FileCachePolicy FileFactory::get_cache_policy(RuntimeState* state) {
     if (state != nullptr) {
         if (config::enable_file_cache && state->query_options().enable_file_cache) {
diff --git a/be/src/io/file_factory.h b/be/src/io/file_factory.h
index 01b13cc521..2510b5e30d 100644
--- a/be/src/io/file_factory.h
+++ b/be/src/io/file_factory.h
@@ -16,19 +16,29 @@
 // under the License.
 #pragma once
 
-#include "gen_cpp/PlanNodes_types.h"
-#include "gen_cpp/Types_types.h"
-#include "io/fs/file_reader.h"
-#include "io/fs/file_writer.h"
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <stdint.h>
+
+#include <map>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <vector>
+
+#include "common/status.h"
+#include "io/fs/file_reader_options.h"
+#include "io/fs/file_reader_writer_fwd.h"
 
 namespace doris {
 namespace io {
 class FileSystem;
-class FileReaderOptions;
+class FileWriter;
 } // namespace io
 class ExecEnv;
-class TNetworkAddress;
 class RuntimeProfile;
+class RuntimeState;
 
 struct FileSystemProperties {
     TFileType::type system_type;
diff --git a/be/src/io/fs/broker_file_reader.cpp b/be/src/io/fs/broker_file_reader.cpp
index 40ddfe6b44..ffd4c5154e 100644
--- a/be/src/io/fs/broker_file_reader.cpp
+++ b/be/src/io/fs/broker_file_reader.cpp
@@ -18,13 +18,26 @@
 #include "io/fs/broker_file_reader.h"
 
 #include <gen_cpp/TPaloBrokerService.h>
+#include <string.h>
+#include <thrift/Thrift.h>
+#include <thrift/transport/TTransportException.h>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <ostream>
+#include <string>
+#include <thread>
+#include <utility>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/logging.h"
 #include "common/status.h"
 #include "io/fs/broker_file_system.h"
 #include "util/doris_metrics.h"
 
 namespace doris {
 namespace io {
+class IOContext;
 
 BrokerFileReader::BrokerFileReader(const TNetworkAddress& broker_addr, const Path& path,
                                    size_t file_size, TBrokerFD fd,
diff --git a/be/src/io/fs/broker_file_reader.h b/be/src/io/fs/broker_file_reader.h
index e16e6e336f..a7f126ba0d 100644
--- a/be/src/io/fs/broker_file_reader.h
+++ b/be/src/io/fs/broker_file_reader.h
@@ -19,16 +19,23 @@
 
 #include <gen_cpp/PaloBrokerService_types.h>
 #include <gen_cpp/Types_types.h>
+#include <stddef.h>
 
 #include <atomic>
+#include <memory>
 
+#include "common/status.h"
 #include "io/fs/broker_file_system.h"
 #include "io/fs/file_reader.h"
+#include "io/fs/file_system.h"
+#include "io/fs/path.h"
 #include "runtime/client_cache.h"
+#include "util/slice.h"
+
 namespace doris {
 namespace io {
 
-class BrokerFileSystem;
+class IOContext;
 
 class BrokerFileReader : public FileReader {
 public:
diff --git a/be/src/io/fs/broker_file_system.cpp b/be/src/io/fs/broker_file_system.cpp
index d7e87e9930..cf7f34a9df 100644
--- a/be/src/io/fs/broker_file_system.cpp
+++ b/be/src/io/fs/broker_file_system.cpp
@@ -17,15 +17,33 @@
 
 #include "io/fs/broker_file_system.h"
 
+#include <fmt/format.h>
 #include <gen_cpp/PaloBrokerService_types.h>
 #include <gen_cpp/TPaloBrokerService.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <stddef.h>
+#include <thrift/Thrift.h>
+#include <thrift/transport/TTransportException.h>
 
+#include <algorithm>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <filesystem>
+#include <ostream>
+#include <thread>
+#include <utility>
+
+#include "common/config.h"
 #include "io/fs/broker_file_reader.h"
 #include "io/fs/broker_file_writer.h"
+#include "io/fs/file_reader.h"
+#include "io/fs/file_system.h"
+#include "io/fs/file_writer.h"
 #include "io/fs/local_file_system.h"
 #include "runtime/broker_mgr.h"
 #include "runtime/exec_env.h"
-#include "util/defer_op.h"
+#include "util/slice.h"
 
 namespace doris {
 namespace io {
diff --git a/be/src/io/fs/broker_file_system.h b/be/src/io/fs/broker_file_system.h
index fa39964878..6b04bbc61c 100644
--- a/be/src/io/fs/broker_file_system.h
+++ b/be/src/io/fs/broker_file_system.h
@@ -17,11 +17,25 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include <map>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "common/status.h"
+#include "io/fs/file_reader_writer_fwd.h"
+#include "io/fs/path.h"
 #include "io/fs/remote_file_system.h"
 #include "runtime/client_cache.h"
+
 namespace doris {
+class TNetworkAddress;
 
 namespace io {
+struct FileInfo;
+
 class BrokerFileSystem final : public RemoteFileSystem {
 public:
     static Status create(const TNetworkAddress& broker_addr,
diff --git a/be/src/io/fs/broker_file_writer.cpp b/be/src/io/fs/broker_file_writer.cpp
index 463c44946f..ac34f381d3 100644
--- a/be/src/io/fs/broker_file_writer.cpp
+++ b/be/src/io/fs/broker_file_writer.cpp
@@ -17,13 +17,17 @@
 
 #include "io/fs/broker_file_writer.h"
 
+#include <gen_cpp/PaloBrokerService_types.h>
+#include <gen_cpp/TPaloBrokerService.h>
+#include <gen_cpp/Types_types.h>
+#include <thrift/Thrift.h>
 #include <thrift/protocol/TDebugProtocol.h>
+#include <thrift/transport/TTransportException.h>
 
 #include <sstream>
 
+#include "common/config.h"
 #include "common/logging.h"
-#include "gen_cpp/PaloBrokerService_types.h"
-#include "gen_cpp/TPaloBrokerService.h"
 #include "runtime/broker_mgr.h"
 #include "runtime/client_cache.h"
 #include "runtime/exec_env.h"
diff --git a/be/src/io/fs/broker_file_writer.h b/be/src/io/fs/broker_file_writer.h
index 5705cc18df..95773f0a04 100644
--- a/be/src/io/fs/broker_file_writer.h
+++ b/be/src/io/fs/broker_file_writer.h
@@ -17,21 +17,23 @@
 
 #pragma once
 
+#include <gen_cpp/PaloBrokerService_types.h>
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
 #include <stdint.h>
 
 #include <map>
 #include <string>
 
 #include "common/status.h"
-#include "gen_cpp/PaloBrokerService_types.h"
-#include "gen_cpp/Types_types.h"
+#include "io/fs/file_system.h"
 #include "io/fs/file_writer.h"
+#include "util/hash_util.hpp" // IWYU pragma: keep
+#include "util/slice.h"
 
 namespace doris {
 
 class ExecEnv;
-class TBrokerRangeDesc;
-class TNetworkAddress;
 
 namespace io {
 
diff --git a/be/src/io/fs/buffered_reader.cpp b/be/src/io/fs/buffered_reader.cpp
index 63d4f10f77..0ab588f68f 100644
--- a/be/src/io/fs/buffered_reader.cpp
+++ b/be/src/io/fs/buffered_reader.cpp
@@ -17,15 +17,22 @@
 
 #include "io/fs/buffered_reader.h"
 
-#include <algorithm>
-#include <sstream>
+#include <bvar/reducer.h>
+#include <bvar/window.h>
+#include <string.h>
 
+#include <algorithm>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/config.h"
-#include "olap/iterators.h"
-#include "olap/olap_define.h"
+#include "runtime/exec_env.h"
+#include "util/runtime_profile.h"
+#include "util/threadpool.h"
 
 namespace doris {
 namespace io {
+class IOContext;
 
 // add bvar to capture the download bytes per second by buffered reader
 bvar::Adder<uint64_t> g_bytes_downloaded("buffered_reader", "bytes_downloaded");
diff --git a/be/src/io/fs/buffered_reader.h b/be/src/io/fs/buffered_reader.h
index 434058e32e..670f64a0db 100644
--- a/be/src/io/fs/buffered_reader.h
+++ b/be/src/io/fs/buffered_reader.h
@@ -17,21 +17,28 @@
 
 #pragma once
 
+#include <stddef.h>
 #include <stdint.h>
 
 #include <condition_variable>
 #include <memory>
+#include <mutex>
+#include <string>
+#include <utility>
+#include <vector>
 
-#include "common/config.h"
 #include "common/status.h"
 #include "io/fs/file_reader.h"
-#include "olap/olap_define.h"
-#include "util/runtime_profile.h"
+#include "io/fs/file_reader_writer_fwd.h"
+#include "io/fs/path.h"
+#include "util/slice.h"
 
 namespace doris {
 namespace io {
 
-class PrefetchBufferedReader;
+class FileSystem;
+class IOContext;
+
 struct PrefetchBuffer : std::enable_shared_from_this<PrefetchBuffer> {
     enum class BufferStatus { RESET, PENDING, PREFETCHED, CLOSED };
     PrefetchBuffer() = default;
diff --git a/be/src/io/fs/err_utils.cpp b/be/src/io/fs/err_utils.cpp
index d01c7e7488..5530b25a0d 100644
--- a/be/src/io/fs/err_utils.cpp
+++ b/be/src/io/fs/err_utils.cpp
@@ -17,7 +17,10 @@
 
 #include "io/fs/err_utils.h"
 
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
 #include <fmt/format.h>
+#include <string.h>
 
 #include <sstream>
 
diff --git a/be/src/io/fs/file_reader.cpp b/be/src/io/fs/file_reader.cpp
index fba531630f..eff62210d5 100644
--- a/be/src/io/fs/file_reader.cpp
+++ b/be/src/io/fs/file_reader.cpp
@@ -17,8 +17,10 @@
 
 #include "io/fs/file_reader.h"
 
+#include <bthread/bthread.h>
+#include <glog/logging.h>
+
 #include "io/fs/file_system.h"
-#include "io/io_common.h"
 #include "util/async_io.h"
 
 namespace doris {
diff --git a/be/src/io/fs/file_reader.h b/be/src/io/fs/file_reader.h
index c429b36ed3..d4f7c5d530 100644
--- a/be/src/io/fs/file_reader.h
+++ b/be/src/io/fs/file_reader.h
@@ -17,9 +17,12 @@
 
 #pragma once
 
+#include <butil/macros.h>
+#include <stddef.h>
+
+#include <memory>
+
 #include "common/status.h"
-#include "gutil/macros.h"
-#include "io/fs/file_reader_writer_fwd.h"
 #include "io/fs/path.h"
 #include "util/slice.h"
 
diff --git a/be/src/io/fs/file_reader_options.h b/be/src/io/fs/file_reader_options.h
index c12ce727bf..ca50bcb1dd 100644
--- a/be/src/io/fs/file_reader_options.h
+++ b/be/src/io/fs/file_reader_options.h
@@ -17,7 +17,9 @@
 
 #pragma once
 
-#include "io/cache/block/block_file_cache.h"
+#include <stdint.h>
+
+#include <string>
 
 namespace doris {
 namespace io {
diff --git a/be/src/io/fs/file_system.cpp b/be/src/io/fs/file_system.cpp
index 1f34500035..d399fbbc57 100644
--- a/be/src/io/fs/file_system.cpp
+++ b/be/src/io/fs/file_system.cpp
@@ -17,8 +17,7 @@
 
 #include "io/fs/file_system.h"
 
-#include "olap/olap_define.h"
-#include "util/async_io.h"
+#include "util/async_io.h" // IWYU pragma: keep
 
 namespace doris {
 namespace io {
diff --git a/be/src/io/fs/file_system.h b/be/src/io/fs/file_system.h
index 3dbf741a05..30ac5f5a4a 100644
--- a/be/src/io/fs/file_system.h
+++ b/be/src/io/fs/file_system.h
@@ -17,17 +17,26 @@
 
 #pragma once
 
+#include <bthread/bthread.h>
+#include <butil/macros.h>
+#include <glog/logging.h>
+#include <stdint.h>
+
+#include <filesystem>
+#include <functional>
 #include <memory>
+#include <string>
+#include <utility>
+#include <vector>
 
 #include "common/status.h"
-#include "gutil/macros.h"
 #include "io/fs/file_reader_options.h"
 #include "io/fs/file_reader_writer_fwd.h"
 #include "io/fs/path.h"
-#include "io/io_common.h"
 
 namespace doris {
 namespace io {
+class FileSystem;
 
 #ifndef FILESYSTEM_M
 #define FILESYSTEM_M(stmt)                    \
diff --git a/be/src/io/fs/fs_utils.cpp b/be/src/io/fs/fs_utils.cpp
index 066dbaa63a..7f12c6767e 100644
--- a/be/src/io/fs/fs_utils.cpp
+++ b/be/src/io/fs/fs_utils.cpp
@@ -17,7 +17,12 @@
 
 #include "io/fs/fs_utils.h"
 
+#include <stddef.h>
+
+#include <memory>
+
 #include "io/fs/file_reader.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "io/fs/file_system.h"
 
 namespace doris {
diff --git a/be/src/io/fs/fs_utils.h b/be/src/io/fs/fs_utils.h
index 866b3cfaea..39744d25a0 100644
--- a/be/src/io/fs/fs_utils.h
+++ b/be/src/io/fs/fs_utils.h
@@ -17,8 +17,15 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <stdint.h>
+
+#include <ostream>
+#include <string>
+
 #include "common/status.h"
 #include "io/fs/file_system.h"
+#include "io/fs/path.h"
 
 namespace doris {
 namespace io {
diff --git a/be/src/io/fs/hdfs_file_reader.cpp b/be/src/io/fs/hdfs_file_reader.cpp
index ddd035213b..901d51bd9c 100644
--- a/be/src/io/fs/hdfs_file_reader.cpp
+++ b/be/src/io/fs/hdfs_file_reader.cpp
@@ -17,6 +17,16 @@
 
 #include "io/fs/hdfs_file_reader.h"
 
+#include <stdint.h>
+
+#include <algorithm>
+#include <filesystem>
+#include <ostream>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/logging.h"
 #include "io/fs/err_utils.h"
 #include "io/fs/hdfs_file_system.h"
 #include "service/backend_options.h"
diff --git a/be/src/io/fs/hdfs_file_system.cpp b/be/src/io/fs/hdfs_file_system.cpp
index 0d33bd30c9..c338390af4 100644
--- a/be/src/io/fs/hdfs_file_system.cpp
+++ b/be/src/io/fs/hdfs_file_system.cpp
@@ -17,16 +17,32 @@
 
 #include "io/fs/hdfs_file_system.h"
 
+#include <fcntl.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <limits.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <filesystem>
+#include <map>
+#include <mutex>
+#include <ostream>
+#include <unordered_map>
+#include <utility>
+
+#include "common/config.h"
 #include "gutil/hash/hash.h"
-#include "io/cache/block/cached_remote_file_reader.h"
+#include "gutil/integral_types.h"
 #include "io/fs/err_utils.h"
+#include "io/fs/file_reader.h"
+#include "io/fs/file_system.h"
+#include "io/fs/file_writer.h"
 #include "io/fs/hdfs_file_reader.h"
 #include "io/fs/hdfs_file_writer.h"
 #include "io/fs/local_file_system.h"
 #include "io/hdfs_builder.h"
-#include "service/backend_options.h"
 #include "util/hdfs_util.h"
-#include "util/stack_util.h"
+#include "util/slice.h"
 
 namespace doris {
 namespace io {
diff --git a/be/src/io/fs/hdfs_file_system.h b/be/src/io/fs/hdfs_file_system.h
index 9cec56b86c..4928f193c3 100644
--- a/be/src/io/fs/hdfs_file_system.h
+++ b/be/src/io/fs/hdfs_file_system.h
@@ -17,15 +17,27 @@
 
 #pragma once
 
-#include <gen_cpp/PlanNodes_types.h>
+#include <glog/logging.h>
+#include <stdint.h>
 
 #include <atomic>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <memory>
+#include <string>
+#include <vector>
 
+#include "common/status.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "io/fs/hdfs.h"
+#include "io/fs/path.h"
 #include "io/fs/remote_file_system.h"
 
 namespace doris {
+class THdfsParams;
+
 namespace io {
+struct FileInfo;
 
 class HdfsFileSystemHandle {
 public:
diff --git a/be/src/io/fs/hdfs_file_writer.cpp b/be/src/io/fs/hdfs_file_writer.cpp
index 723856fa25..5eaaa8a404 100644
--- a/be/src/io/fs/hdfs_file_writer.cpp
+++ b/be/src/io/fs/hdfs_file_writer.cpp
@@ -17,7 +17,13 @@
 
 #include "io/fs/hdfs_file_writer.h"
 
+#include <fcntl.h>
+#include <stdint.h>
+
 #include <filesystem>
+#include <ostream>
+#include <string>
+#include <utility>
 
 #include "common/logging.h"
 #include "io/fs/err_utils.h"
diff --git a/be/src/io/fs/hdfs_file_writer.h b/be/src/io/fs/hdfs_file_writer.h
index c9b3ff9def..371cae6e0c 100644
--- a/be/src/io/fs/hdfs_file_writer.h
+++ b/be/src/io/fs/hdfs_file_writer.h
@@ -19,16 +19,11 @@
 
 #include <stddef.h>
 
-#include <map>
-#include <string>
-
 #include "common/status.h"
-#include "gen_cpp/PlanNodes_types.h"
 #include "io/fs/file_system.h"
 #include "io/fs/file_writer.h"
 #include "io/fs/hdfs.h"
 #include "io/fs/path.h"
-#include "io/hdfs_builder.h"
 #include "util/slice.h"
 
 namespace doris {
diff --git a/be/src/io/fs/local_file_reader.cpp b/be/src/io/fs/local_file_reader.cpp
index c62633e8cb..bed46e0c28 100644
--- a/be/src/io/fs/local_file_reader.cpp
+++ b/be/src/io/fs/local_file_reader.cpp
@@ -17,14 +17,28 @@
 
 #include "io/fs/local_file_reader.h"
 
-#include <atomic>
+#include <bthread/bthread.h>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <unistd.h>
 
+#include <algorithm>
+#include <atomic>
+#include <cstring>
+#include <string>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "io/fs/err_utils.h"
 #include "util/async_io.h"
 #include "util/doris_metrics.h"
 
 namespace doris {
 namespace io {
+class IOContext;
 
 LocalFileReader::LocalFileReader(Path path, size_t file_size, int fd,
                                  std::shared_ptr<LocalFileSystem> fs)
diff --git a/be/src/io/fs/local_file_reader.h b/be/src/io/fs/local_file_reader.h
index 1f74e7b4be..a42c0e3c07 100644
--- a/be/src/io/fs/local_file_reader.h
+++ b/be/src/io/fs/local_file_reader.h
@@ -17,14 +17,21 @@
 
 #pragma once
 
-#include <atomic>
+#include <stddef.h>
 
+#include <atomic>
+#include <memory>
+
+#include "common/status.h"
 #include "io/fs/file_reader.h"
+#include "io/fs/file_system.h"
 #include "io/fs/local_file_system.h"
 #include "io/fs/path.h"
+#include "util/slice.h"
 
 namespace doris {
 namespace io {
+class IOContext;
 
 class LocalFileReader final : public FileReader {
 public:
diff --git a/be/src/io/fs/local_file_system.cpp b/be/src/io/fs/local_file_system.cpp
index 0b7747caf4..9b5c8eec85 100644
--- a/be/src/io/fs/local_file_system.cpp
+++ b/be/src/io/fs/local_file_system.cpp
@@ -17,18 +17,33 @@
 
 #include "io/fs/local_file_system.h"
 
+#include <fcntl.h>
+#include <fmt/format.h>
+#include <glog/logging.h>
 #include <openssl/md5.h>
 #include <sys/mman.h>
+#include <sys/stat.h>
+#include <unistd.h>
 
+#include <filesystem>
+#include <iomanip>
+#include <istream>
+#include <system_error>
+#include <utility>
+
+#include "gutil/macros.h"
 #include "io/fs/err_utils.h"
 #include "io/fs/file_system.h"
+#include "io/fs/file_writer.h"
 #include "io/fs/local_file_reader.h"
 #include "io/fs/local_file_writer.h"
 #include "runtime/thread_context.h"
-#include "util/async_io.h"
+#include "util/async_io.h" // IWYU pragma: keep
+#include "util/defer_op.h"
 
 namespace doris {
 namespace io {
+class FileReaderOptions;
 
 std::shared_ptr<LocalFileSystem> LocalFileSystem::create(Path path, std::string id) {
     return std::shared_ptr<LocalFileSystem>(new LocalFileSystem(std::move(path), std::move(id)));
diff --git a/be/src/io/fs/local_file_system.h b/be/src/io/fs/local_file_system.h
index 15b4c875c2..bcebd48a65 100644
--- a/be/src/io/fs/local_file_system.h
+++ b/be/src/io/fs/local_file_system.h
@@ -17,10 +17,22 @@
 
 #pragma once
 
+#include <stdint.h>
+#include <time.h>
+
+#include <functional>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "common/status.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "io/fs/file_system.h"
+#include "io/fs/path.h"
 
 namespace doris {
 namespace io {
+class FileReaderOptions;
 
 class LocalFileSystem final : public FileSystem {
 public:
diff --git a/be/src/io/fs/local_file_writer.cpp b/be/src/io/fs/local_file_writer.cpp
index c141471d98..62536dd687 100644
--- a/be/src/io/fs/local_file_writer.cpp
+++ b/be/src/io/fs/local_file_writer.cpp
@@ -17,12 +17,23 @@
 
 #include "io/fs/local_file_writer.h"
 
-#include <fmt/core.h>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
+#include <fcntl.h>
+#include <glog/logging.h>
+#include <limits.h>
+#include <stdint.h>
 #include <sys/uio.h>
+#include <unistd.h>
 
-#include <cerrno>
+#include <algorithm>
+#include <cstring>
+#include <memory>
+#include <ostream>
+#include <utility>
 
-#include "common/compiler_util.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/status.h"
 #include "gutil/macros.h"
 #include "io/fs/file_writer.h"
diff --git a/be/src/io/fs/local_file_writer.h b/be/src/io/fs/local_file_writer.h
index e4745893dd..11b2f16434 100644
--- a/be/src/io/fs/local_file_writer.h
+++ b/be/src/io/fs/local_file_writer.h
@@ -19,9 +19,11 @@
 
 #include <cstddef>
 
+#include "common/status.h"
 #include "io/fs/file_system.h"
 #include "io/fs/file_writer.h"
-#include "io/fs/local_file_system.h"
+#include "io/fs/path.h"
+#include "util/slice.h"
 
 namespace doris {
 namespace io {
diff --git a/be/src/io/fs/remote_file_system.cpp b/be/src/io/fs/remote_file_system.cpp
index 9e4e777639..53f7ced5f9 100644
--- a/be/src/io/fs/remote_file_system.cpp
+++ b/be/src/io/fs/remote_file_system.cpp
@@ -17,11 +17,18 @@
 
 #include "io/fs/remote_file_system.h"
 
+#include <glog/logging.h>
+
+#include <algorithm>
+
+#include "common/config.h"
 #include "gutil/strings/stringpiece.h"
 #include "io/cache/block/cached_remote_file_reader.h"
+#include "io/cache/file_cache.h"
 #include "io/cache/file_cache_manager.h"
+#include "io/fs/file_reader.h"
 #include "io/fs/file_reader_options.h"
-#include "util/async_io.h"
+#include "util/async_io.h" // IWYU pragma: keep
 
 namespace doris {
 namespace io {
diff --git a/be/src/io/fs/remote_file_system.h b/be/src/io/fs/remote_file_system.h
index 8626d0840a..44ecde4041 100644
--- a/be/src/io/fs/remote_file_system.h
+++ b/be/src/io/fs/remote_file_system.h
@@ -17,10 +17,21 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include <memory>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "io/fs/file_system.h"
+#include "io/fs/path.h"
 
 namespace doris {
 namespace io {
+class FileReaderOptions;
 
 class RemoteFileSystem : public FileSystem {
 public:
diff --git a/be/src/io/fs/s3_file_reader.cpp b/be/src/io/fs/s3_file_reader.cpp
index b60e038d39..77a3bfcd22 100644
--- a/be/src/io/fs/s3_file_reader.cpp
+++ b/be/src/io/fs/s3_file_reader.cpp
@@ -17,15 +17,25 @@
 
 #include "io/fs/s3_file_reader.h"
 
+#include <aws/core/utils/Outcome.h>
 #include <aws/s3/S3Client.h>
+#include <aws/s3/S3Errors.h>
 #include <aws/s3/model/GetObjectRequest.h>
+#include <aws/s3/model/GetObjectResult.h>
+#include <fmt/format.h>
+#include <glog/logging.h>
 
+#include <algorithm>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "io/fs/s3_common.h"
-#include "util/async_io.h"
 #include "util/doris_metrics.h"
 
 namespace doris {
 namespace io {
+class IOContext;
 
 S3FileReader::S3FileReader(Path path, size_t file_size, std::string key, std::string bucket,
                            std::shared_ptr<S3FileSystem> fs)
diff --git a/be/src/io/fs/s3_file_reader.h b/be/src/io/fs/s3_file_reader.h
index c21084ae8d..1f53fb5814 100644
--- a/be/src/io/fs/s3_file_reader.h
+++ b/be/src/io/fs/s3_file_reader.h
@@ -17,13 +17,22 @@
 
 #pragma once
 
-#include "gutil/macros.h"
+#include <stddef.h>
+
+#include <atomic>
+#include <memory>
+#include <string>
+
+#include "common/status.h"
 #include "io/fs/file_reader.h"
+#include "io/fs/file_system.h"
 #include "io/fs/path.h"
 #include "io/fs/s3_file_system.h"
+#include "util/slice.h"
 
 namespace doris {
 namespace io {
+class IOContext;
 
 class S3FileReader final : public FileReader {
 public:
diff --git a/be/src/io/fs/s3_file_system.cpp b/be/src/io/fs/s3_file_system.cpp
index 3236e3f826..c7efcc414a 100644
--- a/be/src/io/fs/s3_file_system.cpp
+++ b/be/src/io/fs/s3_file_system.cpp
@@ -17,25 +17,57 @@
 
 #include "io/fs/s3_file_system.h"
 
+#include <aws/core/client/AWSError.h>
+#include <aws/core/http/HttpResponse.h>
+#include <aws/core/utils/Outcome.h>
+#include <aws/core/utils/memory/stl/AWSAllocator.h>
+#include <aws/core/utils/memory/stl/AWSMap.h>
+#include <aws/core/utils/memory/stl/AWSStreamFwd.h>
+#include <aws/core/utils/memory/stl/AWSString.h>
+#include <aws/core/utils/memory/stl/AWSStringStream.h>
+#include <aws/core/utils/memory/stl/AWSVector.h>
 #include <aws/core/utils/threading/Executor.h>
 #include <aws/s3/S3Client.h>
+#include <aws/s3/S3Errors.h>
 #include <aws/s3/model/CopyObjectRequest.h>
+#include <aws/s3/model/CopyObjectResult.h>
+#include <aws/s3/model/Delete.h>
 #include <aws/s3/model/DeleteObjectRequest.h>
+#include <aws/s3/model/DeleteObjectResult.h>
 #include <aws/s3/model/DeleteObjectsRequest.h>
+#include <aws/s3/model/DeleteObjectsResult.h>
+#include <aws/s3/model/Error.h>
+#include <aws/s3/model/GetObjectRequest.h>
+#include <aws/s3/model/GetObjectResult.h>
 #include <aws/s3/model/HeadObjectRequest.h>
+#include <aws/s3/model/HeadObjectResult.h>
 #include <aws/s3/model/ListObjectsV2Request.h>
+#include <aws/s3/model/ListObjectsV2Result.h>
+#include <aws/s3/model/Object.h>
+#include <aws/s3/model/ObjectIdentifier.h>
 #include <aws/s3/model/PutObjectRequest.h>
+#include <aws/s3/model/PutObjectResult.h>
+#include <aws/transfer/TransferHandle.h>
 #include <aws/transfer/TransferManager.h>
-#include <opentelemetry/common/threadlocal.h>
+#include <fmt/format.h>
+#include <stddef.h>
 
+#include <algorithm>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
 #include <filesystem>
-#include <fstream>
+#include <fstream> // IWYU pragma: keep
 #include <memory>
+#include <sstream>
 
 #include "common/config.h"
+#include "common/logging.h"
 #include "common/status.h"
-#include "gutil/strings/stringpiece.h"
-#include "io/cache/block/cached_remote_file_reader.h"
+#include "io/fs/file_system.h"
+#include "io/fs/file_writer.h"
 #include "io/fs/remote_file_system.h"
 #include "io/fs/s3_file_reader.h"
 #include "io/fs/s3_file_writer.h"
diff --git a/be/src/io/fs/s3_file_system.h b/be/src/io/fs/s3_file_system.h
index 0ae687e44d..0c7e3d17e1 100644
--- a/be/src/io/fs/s3_file_system.h
+++ b/be/src/io/fs/s3_file_system.h
@@ -17,8 +17,18 @@
 
 #pragma once
 
-#include <mutex>
+#include <stdint.h>
 
+#include <filesystem>
+#include <memory>
+#include <mutex>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
+#include "io/fs/file_reader_writer_fwd.h"
+#include "io/fs/path.h"
 #include "io/fs/remote_file_system.h"
 #include "util/s3_util.h"
 
@@ -31,6 +41,7 @@ class PooledThreadExecutor;
 
 namespace doris {
 namespace io {
+struct FileInfo;
 
 // File system for S3 compatible object storage
 // When creating S3FileSystem, all required info should be set in S3Conf,
diff --git a/be/src/io/fs/s3_file_writer.cpp b/be/src/io/fs/s3_file_writer.cpp
index f613b14e17..989f9c0c2e 100644
--- a/be/src/io/fs/s3_file_writer.cpp
+++ b/be/src/io/fs/s3_file_writer.cpp
@@ -17,29 +17,41 @@
 
 #include "io/fs/s3_file_writer.h"
 
-#include <aws/core/Aws.h>
+#include <aws/core/http/HttpResponse.h>
+#include <aws/core/utils/Array.h>
 #include <aws/core/utils/HashingUtils.h>
+#include <aws/core/utils/Outcome.h>
+#include <aws/core/utils/memory/stl/AWSAllocator.h>
 #include <aws/s3/S3Client.h>
+#include <aws/s3/S3Errors.h>
 #include <aws/s3/model/AbortMultipartUploadRequest.h>
+#include <aws/s3/model/AbortMultipartUploadResult.h>
 #include <aws/s3/model/CompleteMultipartUploadRequest.h>
+#include <aws/s3/model/CompleteMultipartUploadResult.h>
+#include <aws/s3/model/CompletedMultipartUpload.h>
+#include <aws/s3/model/CompletedPart.h>
 #include <aws/s3/model/CreateMultipartUploadRequest.h>
-#include <aws/s3/model/DeleteObjectRequest.h>
-#include <aws/s3/model/DeleteObjectsRequest.h>
-#include <aws/s3/model/GetObjectRequest.h>
+#include <aws/s3/model/CreateMultipartUploadResult.h>
 #include <aws/s3/model/UploadPartRequest.h>
-#include <fmt/core.h>
-#include <sys/uio.h>
+#include <aws/s3/model/UploadPartResult.h>
+#include <glog/logging.h>
 
-#include <cerrno>
+#include <sstream>
+#include <utility>
 
-#include "common/compiler_util.h"
 #include "common/status.h"
-#include "gutil/macros.h"
 #include "io/fs/file_writer.h"
 #include "io/fs/path.h"
-#include "io/fs/s3_file_system.h"
 #include "util/doris_metrics.h"
 
+namespace Aws {
+namespace S3 {
+namespace Model {
+class DeleteObjectRequest;
+} // namespace Model
+} // namespace S3
+} // namespace Aws
+
 using Aws::S3::Model::AbortMultipartUploadRequest;
 using Aws::S3::Model::CompletedPart;
 using Aws::S3::Model::CompletedMultipartUpload;
diff --git a/be/src/io/fs/s3_file_writer.h b/be/src/io/fs/s3_file_writer.h
index 87aa58680f..d8dc9cc800 100644
--- a/be/src/io/fs/s3_file_writer.h
+++ b/be/src/io/fs/s3_file_writer.h
@@ -17,12 +17,19 @@
 
 #pragma once
 
+#include <aws/core/utils/memory/stl/AWSStringStream.h>
+
 #include <cstddef>
 #include <list>
+#include <memory>
+#include <string>
 
+#include "common/status.h"
+#include "io/fs/file_system.h"
 #include "io/fs/file_writer.h"
-#include "io/fs/s3_file_system.h"
+#include "io/fs/path.h"
 #include "util/s3_util.h"
+#include "util/slice.h"
 
 namespace Aws::S3 {
 namespace Model {
diff --git a/be/src/io/fs/stream_load_pipe.cpp b/be/src/io/fs/stream_load_pipe.cpp
index f0e1be2c6e..25099c8796 100644
--- a/be/src/io/fs/stream_load_pipe.cpp
+++ b/be/src/io/fs/stream_load_pipe.cpp
@@ -17,12 +17,22 @@
 
 #include "stream_load_pipe.h"
 
-#include "olap/iterators.h"
+#include <glog/logging.h>
+
+#include <algorithm>
+#include <ostream>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "runtime/exec_env.h"
 #include "runtime/thread_context.h"
 #include "util/bit_util.h"
 
 namespace doris {
 namespace io {
+class IOContext;
+
 StreamLoadPipe::StreamLoadPipe(size_t max_buffered_bytes, size_t min_chunk_size,
                                int64_t total_length, bool use_proto)
         : _buffered_bytes(0),
diff --git a/be/src/io/fs/stream_load_pipe.h b/be/src/io/fs/stream_load_pipe.h
index 33e72c57a1..e06e7dfc6e 100644
--- a/be/src/io/fs/stream_load_pipe.h
+++ b/be/src/io/fs/stream_load_pipe.h
@@ -18,16 +18,26 @@
 #pragma once
 
 #include <gen_cpp/internal_service.pb.h>
+#include <stddef.h>
+#include <stdint.h>
 
 #include <condition_variable>
 #include <deque>
+#include <memory>
+#include <mutex>
+#include <string>
 
+#include "common/status.h"
 #include "io/fs/file_reader.h"
 #include "io/fs/file_system.h"
+#include "io/fs/path.h"
 #include "runtime/message_body_sink.h"
+#include "util/byte_buffer.h"
+#include "util/slice.h"
 
 namespace doris {
 namespace io {
+class IOContext;
 
 const size_t kMaxPipeBufferedBytes = 4 * 1024 * 1024;
 
diff --git a/be/src/io/hdfs_builder.cpp b/be/src/io/hdfs_builder.cpp
index 3ce48eede7..ca629e4401 100644
--- a/be/src/io/hdfs_builder.cpp
+++ b/be/src/io/hdfs_builder.cpp
@@ -18,14 +18,18 @@
 #include "io/hdfs_builder.h"
 
 #include <fmt/format.h>
+#include <gen_cpp/PlanNodes_types.h>
 
+#include <cstdlib>
 #include <fstream>
+#include <utility>
+#include <vector>
 
 #include "agent/utils.h"
 #include "common/logging.h"
+#include "io/fs/hdfs.h"
 #include "util/string_util.h"
 #include "util/uid_util.h"
-#include "util/url_coding.h"
 
 namespace doris {
 
diff --git a/be/src/io/hdfs_builder.h b/be/src/io/hdfs_builder.h
index 7d448cb1cb..d3f7c35017 100644
--- a/be/src/io/hdfs_builder.h
+++ b/be/src/io/hdfs_builder.h
@@ -17,9 +17,15 @@
 
 #pragma once
 
+#include <gen_cpp/PlanNodes_types.h>
+
+#include <map>
+#include <string>
+
 #include "common/status.h"
-#include "gen_cpp/PlanNodes_types.h"
-#include "io/fs/hdfs.h"
+#include "io/fs/hdfs.h" // IWYU pragma: keep
+
+struct hdfsBuilder;
 
 namespace doris {
 
diff --git a/be/src/io/io_common.h b/be/src/io/io_common.h
index d99b4bb1fb..b3b95853f0 100644
--- a/be/src/io/io_common.h
+++ b/be/src/io/io_common.h
@@ -17,7 +17,7 @@
 
 #pragma once
 
-#include "gen_cpp/Types_types.h"
+#include <gen_cpp/Types_types.h>
 
 namespace doris {
 
diff --git a/be/src/olap/base_compaction.cpp b/be/src/olap/base_compaction.cpp
index d67f8a2d72..c1fcbe3f33 100644
--- a/be/src/olap/base_compaction.cpp
+++ b/be/src/olap/base_compaction.cpp
@@ -17,7 +17,21 @@
 
 #include "olap/base_compaction.h"
 
+#include <gen_cpp/olap_file.pb.h>
+#include <stdint.h>
+#include <time.h>
+
+#include <memory>
+#include <mutex>
+#include <ostream>
+
+#include "common/config.h"
+#include "common/logging.h"
+#include "olap/olap_define.h"
+#include "olap/rowset/rowset_meta.h"
+#include "runtime/thread_context.h"
 #include "util/doris_metrics.h"
+#include "util/thread.h"
 #include "util/trace.h"
 
 namespace doris {
diff --git a/be/src/olap/base_compaction.h b/be/src/olap/base_compaction.h
index da53a89d95..ecee304449 100644
--- a/be/src/olap/base_compaction.h
+++ b/be/src/olap/base_compaction.h
@@ -17,7 +17,16 @@
 
 #pragma once
 
+#include <butil/macros.h>
+
+#include <string>
+#include <vector>
+
+#include "common/status.h"
+#include "io/io_common.h"
 #include "olap/compaction.h"
+#include "olap/rowset/rowset.h"
+#include "olap/tablet.h"
 
 namespace doris {
 
diff --git a/be/src/olap/base_tablet.cpp b/be/src/olap/base_tablet.cpp
index ee2e9539ab..3af9615168 100644
--- a/be/src/olap/base_tablet.cpp
+++ b/be/src/olap/base_tablet.cpp
@@ -17,11 +17,18 @@
 
 #include "olap/base_tablet.h"
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+
+#include <ostream>
+#include <vector>
+
 #include "gutil/strings/substitute.h"
 #include "olap/data_dir.h"
+#include "olap/olap_define.h"
+#include "olap/rowset/rowset_meta.h"
 #include "olap/tablet_schema_cache.h"
 #include "util/doris_metrics.h"
-#include "util/path_util.h"
 
 namespace doris {
 using namespace ErrorCode;
diff --git a/be/src/olap/base_tablet.h b/be/src/olap/base_tablet.h
index 05707e81a6..164215989b 100644
--- a/be/src/olap/base_tablet.h
+++ b/be/src/olap/base_tablet.h
@@ -17,13 +17,17 @@
 
 #pragma once
 
+#include <butil/macros.h>
+#include <gen_cpp/olap_file.pb.h>
+#include <stdint.h>
+
 #include <memory>
 #include <string>
 
-#include "olap/olap_define.h"
+#include "common/status.h"
+#include "olap/olap_common.h"
 #include "olap/tablet_meta.h"
 #include "olap/tablet_schema.h"
-#include "olap/utils.h"
 #include "util/metrics.h"
 
 namespace doris {
diff --git a/be/src/olap/block_column_predicate.cpp b/be/src/olap/block_column_predicate.cpp
index ff7e0ba694..1bcde94de3 100644
--- a/be/src/olap/block_column_predicate.cpp
+++ b/be/src/olap/block_column_predicate.cpp
@@ -17,9 +17,17 @@
 
 #include "block_column_predicate.h"
 
-#include "olap/rowset/segment_v2/bloom_filter.h"
+#include <string.h>
+
+namespace roaring {
+class Roaring;
+} // namespace roaring
 
 namespace doris {
+class WrapperField;
+namespace segment_v2 {
+class InvertedIndexIterator;
+} // namespace segment_v2
 
 uint16_t SingleColumnBlockPredicate::evaluate(vectorized::MutableColumns& block, uint16_t* sel,
                                               uint16_t selected_size) const {
diff --git a/be/src/olap/block_column_predicate.h b/be/src/olap/block_column_predicate.h
index 18c957e3c4..467962e809 100644
--- a/be/src/olap/block_column_predicate.h
+++ b/be/src/olap/block_column_predicate.h
@@ -17,15 +17,33 @@
 
 #pragma once
 
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <ostream>
+#include <set>
+#include <string>
+#include <utility>
 #include <vector>
 
+#include "common/status.h"
 #include "olap/column_predicate.h"
+#include "olap/olap_common.h"
+#include "vec/columns/column.h"
+
+namespace roaring {
+class Roaring;
+} // namespace roaring
 
 namespace doris {
+class WrapperField;
 
 namespace segment_v2 {
 class BloomFilter;
-}
+class InvertedIndexIterator;
+} // namespace segment_v2
 
 // Block Column Predicate support do column predicate and support OR and AND predicate
 // Block Column Predicate will replace column predicate as a unified external vectorized interface
diff --git a/be/src/olap/cold_data_compaction.cpp b/be/src/olap/cold_data_compaction.cpp
index 2a27d92acb..6757691d84 100644
--- a/be/src/olap/cold_data_compaction.cpp
+++ b/be/src/olap/cold_data_compaction.cpp
@@ -17,10 +17,25 @@
 
 #include "olap/cold_data_compaction.h"
 
-#include "common/compiler_util.h"
+#include <stdint.h>
+
+#include <algorithm>
+#include <memory>
+#include <mutex>
+#include <shared_mutex>
+#include <utility>
+#include <vector>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/config.h"
 #include "olap/compaction.h"
-#include "olap/rowset/beta_rowset.h"
+#include "olap/olap_common.h"
 #include "olap/rowset/rowset.h"
+#include "olap/tablet_meta.h"
+#include "runtime/thread_context.h"
+#include "util/thread.h"
+#include "util/uid_util.h"
 
 namespace doris {
 using namespace ErrorCode;
diff --git a/be/src/olap/cold_data_compaction.h b/be/src/olap/cold_data_compaction.h
index b9e8465273..0b2eddf234 100644
--- a/be/src/olap/cold_data_compaction.h
+++ b/be/src/olap/cold_data_compaction.h
@@ -17,7 +17,13 @@
 
 #pragma once
 
+#include <string>
+
+#include "common/status.h"
+#include "io/io_common.h"
 #include "olap/compaction.h"
+#include "olap/merger.h"
+#include "olap/tablet.h"
 
 namespace doris {
 
diff --git a/be/src/olap/compaction.cpp b/be/src/olap/compaction.cpp
index cacf3a9b66..608c89f6ae 100644
--- a/be/src/olap/compaction.cpp
+++ b/be/src/olap/compaction.cpp
@@ -17,14 +17,39 @@
 
 #include "olap/compaction.h"
 
+#include <fmt/format.h>
+#include <gen_cpp/olap_file.pb.h>
+#include <glog/logging.h>
+
+#include <algorithm>
+#include <cstdlib>
+#include <list>
+#include <map>
+#include <mutex>
+#include <ostream>
+#include <set>
+#include <shared_mutex>
+#include <utility>
+
+#include "common/config.h"
 #include "common/status.h"
-#include "gutil/strings/substitute.h"
+#include "io/fs/file_system.h"
+#include "io/fs/remote_file_system.h"
+#include "olap/cumulative_compaction_policy.h"
+#include "olap/data_dir.h"
+#include "olap/olap_define.h"
 #include "olap/rowset/beta_rowset.h"
 #include "olap/rowset/rowset.h"
 #include "olap/rowset/rowset_meta.h"
+#include "olap/rowset/rowset_writer.h"
 #include "olap/rowset/rowset_writer_context.h"
+#include "olap/storage_engine.h"
+#include "olap/storage_policy.h"
 #include "olap/tablet.h"
+#include "olap/tablet_meta.h"
 #include "olap/task/engine_checksum_task.h"
+#include "olap/utils.h"
+#include "runtime/memory/mem_tracker_limiter.h"
 #include "util/time.h"
 #include "util/trace.h"
 
diff --git a/be/src/olap/compaction.h b/be/src/olap/compaction.h
index cf35ee8aa6..17acb45710 100644
--- a/be/src/olap/compaction.h
+++ b/be/src/olap/compaction.h
@@ -17,21 +17,27 @@
 
 #pragma once
 
+#include <butil/macros.h>
+#include <stdint.h>
+
+#include <memory>
+#include <string>
 #include <vector>
 
+#include "common/status.h"
+#include "io/io_common.h"
 #include "olap/merger.h"
 #include "olap/olap_common.h"
-#include "olap/olap_define.h"
-#include "olap/storage_engine.h"
+#include "olap/rowid_conversion.h"
+#include "olap/rowset/rowset.h"
+#include "olap/rowset/rowset_reader.h"
 #include "olap/tablet.h"
-#include "olap/tablet_meta.h"
-#include "olap/utils.h"
-#include "rowset/rowset_id_generator.h"
+#include "olap/tablet_schema.h"
 
 namespace doris {
 
-class DataDir;
-class Merger;
+class MemTrackerLimiter;
+class RowsetWriter;
 
 // This class is a base class for compaction.
 // The entrance of this class is compact()
diff --git a/be/src/olap/compaction_permit_limiter.cpp b/be/src/olap/compaction_permit_limiter.cpp
index 2d6309ae6c..0986730059 100644
--- a/be/src/olap/compaction_permit_limiter.cpp
+++ b/be/src/olap/compaction_permit_limiter.cpp
@@ -17,6 +17,9 @@
 
 #include "olap/compaction_permit_limiter.h"
 
+#include "common/config.h"
+#include "util/doris_metrics.h"
+
 namespace doris {
 
 CompactionPermitLimiter::CompactionPermitLimiter() : _used_permits(0) {}
diff --git a/be/src/olap/compaction_permit_limiter.h b/be/src/olap/compaction_permit_limiter.h
index 21c819830d..c00e0aae27 100644
--- a/be/src/olap/compaction_permit_limiter.h
+++ b/be/src/olap/compaction_permit_limiter.h
@@ -17,13 +17,12 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include <atomic>
 #include <condition_variable>
 #include <mutex>
 
-#include "common/config.h"
-#include "olap/utils.h"
-#include "util/doris_metrics.h"
-
 namespace doris {
 
 /*
diff --git a/be/src/olap/comparison_predicate.h b/be/src/olap/comparison_predicate.h
index 0371b27282..c74b06eaf3 100644
--- a/be/src/olap/comparison_predicate.h
+++ b/be/src/olap/comparison_predicate.h
@@ -21,6 +21,7 @@
 
 #include "olap/column_predicate.h"
 #include "olap/rowset/segment_v2/bloom_filter.h"
+#include "olap/rowset/segment_v2/inverted_index_cache.h" // IWYU pragma: keep
 #include "olap/rowset/segment_v2/inverted_index_reader.h"
 #include "olap/wrapper_field.h"
 #include "vec/columns/column_dictionary.h"
diff --git a/be/src/olap/cumulative_compaction.cpp b/be/src/olap/cumulative_compaction.cpp
index 745b316cab..303b58eefe 100644
--- a/be/src/olap/cumulative_compaction.cpp
+++ b/be/src/olap/cumulative_compaction.cpp
@@ -17,6 +17,19 @@
 
 #include "olap/cumulative_compaction.h"
 
+#include <stddef.h>
+#include <stdint.h>
+
+#include <memory>
+#include <mutex>
+#include <ostream>
+
+#include "common/config.h"
+#include "common/logging.h"
+#include "olap/cumulative_compaction_policy.h"
+#include "olap/olap_define.h"
+#include "olap/rowset/rowset_meta.h"
+#include "runtime/thread_context.h"
 #include "util/doris_metrics.h"
 #include "util/time.h"
 #include "util/trace.h"
diff --git a/be/src/olap/cumulative_compaction.h b/be/src/olap/cumulative_compaction.h
index fdfc9dbe46..d74542a2ff 100644
--- a/be/src/olap/cumulative_compaction.h
+++ b/be/src/olap/cumulative_compaction.h
@@ -17,10 +17,17 @@
 
 #pragma once
 
-#include <string>
+#include <butil/macros.h>
 
+#include <string>
+#include <vector>
+
+#include "common/status.h"
+#include "io/io_common.h"
 #include "olap/compaction.h"
-#include "olap/cumulative_compaction_policy.h"
+#include "olap/olap_common.h"
+#include "olap/rowset/rowset.h"
+#include "olap/tablet.h"
 
 namespace doris {
 
diff --git a/be/src/olap/cumulative_compaction_policy.cpp b/be/src/olap/cumulative_compaction_policy.cpp
index b38271fd1a..d5d86ad142 100644
--- a/be/src/olap/cumulative_compaction_policy.cpp
+++ b/be/src/olap/cumulative_compaction_policy.cpp
@@ -17,10 +17,15 @@
 
 #include "olap/cumulative_compaction_policy.h"
 
-#include <boost/algorithm/string.hpp>
+#include <algorithm>
+#include <list>
+#include <ostream>
 #include <string>
 
-#include "util/time.h"
+#include "common/logging.h"
+#include "olap/olap_common.h"
+#include "olap/tablet.h"
+#include "olap/tablet_meta.h"
 
 namespace doris {
 
diff --git a/be/src/olap/cumulative_compaction_policy.h b/be/src/olap/cumulative_compaction_policy.h
index 20a9d64b61..a12d1d8e8b 100644
--- a/be/src/olap/cumulative_compaction_policy.h
+++ b/be/src/olap/cumulative_compaction_policy.h
@@ -17,17 +17,21 @@
 
 #pragma once
 
-#include <string>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "common/config.h"
 #include "olap/rowset/rowset.h"
 #include "olap/rowset/rowset_meta.h"
-#include "olap/tablet.h"
-#include "olap/tablet_meta.h"
-#include "olap/utils.h"
 
 namespace doris {
 
 class Tablet;
+struct Version;
 
 const static std::string CUMULATIVE_SIZE_BASED_POLICY = "SIZE_BASED";
 
diff --git a/be/src/olap/data_dir.cpp b/be/src/olap/data_dir.cpp
index c1f6e807d7..b83ce43635 100644
--- a/be/src/olap/data_dir.cpp
+++ b/be/src/olap/data_dir.cpp
@@ -17,42 +17,50 @@
 
 #include "olap/data_dir.h"
 
-#include "olap/storage_policy.h"
-
-#ifndef __APPLE__
-#include <mntent.h>
-#include <sys/statfs.h>
-#endif
-
-#include <ctype.h>
+#include <fmt/format.h>
+#include <gen_cpp/Types_types.h>
 #include <gen_cpp/olap_file.pb.h>
 #include <stdio.h>
-#include <sys/file.h>
-#include <utime.h>
 
 #include <atomic>
-#include <boost/algorithm/string/classification.hpp>
-#include <boost/algorithm/string/predicate.hpp>
-#include <boost/algorithm/string/split.hpp>
-#include <boost/algorithm/string/trim.hpp>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <filesystem>
 #include <memory>
+#include <new>
 #include <set>
 #include <sstream>
 #include <string>
+#include <thread>
+#include <utility>
 
+#include "common/config.h"
+#include "common/logging.h"
 #include "gutil/strings/substitute.h"
+#include "io/fs/file_reader_writer_fwd.h"
+#include "io/fs/file_writer.h"
 #include "io/fs/fs_utils.h"
 #include "io/fs/local_file_system.h"
 #include "io/fs/path.h"
+#include "io/fs/remote_file_system.h"
 #include "olap/olap_define.h"
+#include "olap/olap_meta.h"
 #include "olap/rowset/beta_rowset.h"
+#include "olap/rowset/rowset.h"
+#include "olap/rowset/rowset_id_generator.h"
+#include "olap/rowset/rowset_meta.h"
 #include "olap/rowset/rowset_meta_manager.h"
 #include "olap/storage_engine.h"
+#include "olap/storage_policy.h"
+#include "olap/tablet.h"
+#include "olap/tablet_manager.h"
 #include "olap/tablet_meta_manager.h"
+#include "olap/txn_manager.h"
 #include "olap/utils.h" // for check_dir_existed
 #include "service/backend_options.h"
-#include "util/errno.h"
+#include "util/doris_metrics.h"
 #include "util/string_util.h"
+#include "util/uid_util.h"
 
 using strings::Substitute;
 
diff --git a/be/src/olap/data_dir.h b/be/src/olap/data_dir.h
index f9aa644b91..9c6adab9c5 100644
--- a/be/src/olap/data_dir.h
+++ b/be/src/olap/data_dir.h
@@ -17,28 +17,32 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+
+#include <atomic>
 #include <condition_variable>
 #include <cstdint>
+#include <memory>
 #include <mutex>
 #include <set>
 #include <shared_mutex>
 #include <string>
+#include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/Types_types.h"
-#include "gen_cpp/olap_file.pb.h"
 #include "io/fs/file_system.h"
 #include "io/fs/fs_utils.h"
 #include "olap/olap_common.h"
-#include "olap/rowset/rowset_id_generator.h"
 #include "util/metrics.h"
 
 namespace doris {
 
 class Tablet;
 class TabletManager;
-class TabletMeta;
 class TxnManager;
+class OlapMeta;
+class RowsetIdGenerator;
 
 // A DataDir used to manage data in same path.
 // Now, After DataDir was created, it will never be deleted for easy implementation.
diff --git a/be/src/olap/delete_handler.cpp b/be/src/olap/delete_handler.cpp
index 655670e9a8..9beafe634f 100644
--- a/be/src/olap/delete_handler.cpp
+++ b/be/src/olap/delete_handler.cpp
@@ -17,20 +17,24 @@
 
 #include "olap/delete_handler.h"
 
-#include <errno.h>
-#include <json2pb/pb_to_json.h>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/olap_file.pb.h>
 #include <thrift/protocol/TDebugProtocol.h>
 
+#include <algorithm>
 #include <limits>
 #include <regex>
 #include <sstream>
 #include <string>
+#include <utility>
 #include <vector>
 
-#include "gen_cpp/olap_file.pb.h"
+#include "common/config.h"
+#include "common/logging.h"
+#include "olap/block_column_predicate.h"
+#include "olap/column_predicate.h"
 #include "olap/olap_common.h"
 #include "olap/predicate_creator.h"
-#include "olap/tablet.h"
 #include "olap/utils.h"
 
 using apache::thrift::ThriftDebugString;
diff --git a/be/src/olap/delete_handler.h b/be/src/olap/delete_handler.h
index f18b7c69bd..7239f4326b 100644
--- a/be/src/olap/delete_handler.h
+++ b/be/src/olap/delete_handler.h
@@ -17,23 +17,25 @@
 
 #pragma once
 
+#include <butil/macros.h>
+#include <stdint.h>
+
+#include <memory>
 #include <string>
+#include <unordered_map>
 #include <vector>
 
-#include "gen_cpp/AgentService_types.h"
-#include "gen_cpp/olap_file.pb.h"
-#include "olap/block_column_predicate.h"
-#include "olap/column_predicate.h"
-#include "olap/olap_define.h"
+#include "common/status.h"
 #include "olap/rowset/rowset_meta.h"
 #include "olap/tablet_schema.h"
+#include "vec/common/arena.h"
 
 namespace doris {
 
-class RowCursor;
-class Tablet;
-class TabletReader;
-class TabletSchema;
+class AndBlockColumnPredicate;
+class ColumnPredicate;
+class DeletePredicatePB;
+class TCondition;
 
 // Represent a delete condition.
 struct DeleteConditions {
diff --git a/be/src/olap/delta_writer.cpp b/be/src/olap/delta_writer.cpp
index 5f339a5471..b846a5479f 100644
--- a/be/src/olap/delta_writer.cpp
+++ b/be/src/olap/delta_writer.cpp
@@ -17,21 +17,49 @@
 
 #include "olap/delta_writer.h"
 
+#include <brpc/controller.h>
+#include <butil/errno.h>
+#include <fmt/format.h>
+#include <gen_cpp/internal_service.pb.h>
+#include <gen_cpp/olap_file.pb.h>
+
+#include <filesystem>
+#include <ostream>
+#include <string>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/config.h"
+#include "common/logging.h"
 #include "common/status.h"
 #include "exec/tablet_info.h"
+#include "gutil/strings/numbers.h"
+#include "io/fs/file_writer.h" // IWYU pragma: keep
 #include "olap/data_dir.h"
 #include "olap/memtable.h"
 #include "olap/memtable_flush_executor.h"
+#include "olap/olap_define.h"
 #include "olap/rowset/beta_rowset.h"
 #include "olap/rowset/beta_rowset_writer.h"
+#include "olap/rowset/rowset_meta.h"
+#include "olap/rowset/rowset_writer.h"
 #include "olap/rowset/rowset_writer_context.h"
+#include "olap/rowset/segment_v2/segment.h"
 #include "olap/schema.h"
 #include "olap/schema_change.h"
 #include "olap/storage_engine.h"
+#include "olap/tablet_manager.h"
+#include "olap/txn_manager.h"
+#include "runtime/exec_env.h"
 #include "runtime/load_channel_mgr.h"
+#include "runtime/memory/mem_tracker.h"
 #include "service/backend_options.h"
 #include "util/brpc_client_cache.h"
 #include "util/ref_count_closure.h"
+#include "util/stopwatch.hpp"
+#include "util/time.h"
+#include "vec/core/block.h"
 
 namespace doris {
 using namespace ErrorCode;
diff --git a/be/src/olap/delta_writer.h b/be/src/olap/delta_writer.h
index 0b133e6ac9..03b30160fe 100644
--- a/be/src/olap/delta_writer.h
+++ b/be/src/olap/delta_writer.h
@@ -17,10 +17,26 @@
 
 #pragma once
 
-#include "gen_cpp/internal_service.pb.h"
-#include "olap/rowset/rowset_writer.h"
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/internal_service.pb.h>
+#include <gen_cpp/types.pb.h>
+#include <stdint.h>
+
+#include <atomic>
+#include <memory>
+#include <mutex>
+#include <shared_mutex>
+#include <unordered_set>
+#include <vector>
+
+#include "common/status.h"
+#include "olap/olap_common.h"
+#include "olap/rowset/rowset.h"
 #include "olap/tablet.h"
+#include "olap/tablet_meta.h"
+#include "olap/tablet_schema.h"
 #include "util/spinlock.h"
+#include "util/uid_util.h"
 
 namespace doris {
 
@@ -31,6 +47,12 @@ class Schema;
 class StorageEngine;
 class TupleDescriptor;
 class SlotDescriptor;
+class OlapTableSchemaParam;
+class RowsetWriter;
+
+namespace vectorized {
+class Block;
+} // namespace vectorized
 
 enum WriteType { LOAD = 1, LOAD_DELETE = 2, DELETE = 3 };
 
diff --git a/be/src/olap/hll.cpp b/be/src/olap/hll.cpp
index 8f7a61e712..bfbe954217 100644
--- a/be/src/olap/hll.cpp
+++ b/be/src/olap/hll.cpp
@@ -17,12 +17,14 @@
 
 #include "olap/hll.h"
 
-#include <algorithm>
+#include <cmath>
 #include <map>
+#include <new>
+#include <ostream>
 
 #include "common/logging.h"
 #include "util/coding.h"
-#include "vec/common/string_ref.h"
+#include "util/slice.h"
 
 using std::map;
 using std::nothrow;
diff --git a/be/src/olap/hll.h b/be/src/olap/hll.h
index f1ad81100b..7c88abfbb5 100644
--- a/be/src/olap/hll.h
+++ b/be/src/olap/hll.h
@@ -17,18 +17,17 @@
 
 #pragma once
 
-#include <math.h>
-#include <stdio.h>
+#include <stdint.h>
+#include <string.h>
 
 #include <map>
-#include <set>
 #include <string>
+#include <utility>
 
 #ifdef __x86_64__
 #include <immintrin.h>
 #endif
 
-#include "gutil/macros.h"
 #include "vec/common/hash_table/phmap_fwd_decl.h"
 
 namespace doris {
diff --git a/be/src/olap/in_list_predicate.h b/be/src/olap/in_list_predicate.h
index e207bf28d5..abecd69077 100644
--- a/be/src/olap/in_list_predicate.h
+++ b/be/src/olap/in_list_predicate.h
@@ -25,6 +25,7 @@
 #include "olap/column_predicate.h"
 #include "olap/olap_common.h"
 #include "olap/rowset/segment_v2/bloom_filter.h"
+#include "olap/rowset/segment_v2/inverted_index_cache.h" // IWYU pragma: keep
 #include "olap/rowset/segment_v2/inverted_index_reader.h"
 #include "olap/wrapper_field.h"
 #include "runtime/define_primitive_type.h"
diff --git a/be/src/olap/itoken_extractor.cpp b/be/src/olap/itoken_extractor.cpp
index 90ad29d291..fefe86ec27 100644
--- a/be/src/olap/itoken_extractor.cpp
+++ b/be/src/olap/itoken_extractor.cpp
@@ -17,6 +17,8 @@
 
 #include "itoken_extractor.h"
 
+#include <stdint.h>
+
 #include "util/simd/vstring_function.h"
 
 namespace doris {
diff --git a/be/src/olap/key_coder.cpp b/be/src/olap/key_coder.cpp
index 450c805a8d..803b353375 100644
--- a/be/src/olap/key_coder.cpp
+++ b/be/src/olap/key_coder.cpp
@@ -17,7 +17,9 @@
 
 #include "olap/key_coder.h"
 
+#include <cstddef>
 #include <unordered_map>
+#include <utility>
 
 namespace doris {
 
diff --git a/be/src/olap/key_coder.h b/be/src/olap/key_coder.h
index 411517bd05..87c2748688 100644
--- a/be/src/olap/key_coder.h
+++ b/be/src/olap/key_coder.h
@@ -17,13 +17,23 @@
 
 #pragma once
 
+#include <glog/logging.h>
+#include <limits.h>
+#include <stdint.h>
+#include <string.h>
+
+#include <algorithm>
+#include <ostream>
 #include <string>
 #include <type_traits>
 
 #include "common/status.h"
 #include "gutil/endian.h"
 #include "gutil/strings/substitute.h"
+#include "olap/decimal12.h"
+#include "olap/olap_common.h"
 #include "olap/types.h"
+#include "util/slice.h"
 
 namespace doris {
 
diff --git a/be/src/olap/like_column_predicate.cpp b/be/src/olap/like_column_predicate.cpp
index 049a419fbf..ef1374d445 100644
--- a/be/src/olap/like_column_predicate.cpp
+++ b/be/src/olap/like_column_predicate.cpp
@@ -17,8 +17,10 @@
 
 #include "olap/like_column_predicate.h"
 
-#include "olap/field.h"
+#include "runtime/define_primitive_type.h"
 #include "udf/udf.h"
+#include "vec/columns/columns_number.h"
+#include "vec/columns/predicate_column.h"
 #include "vec/common/string_ref.h"
 
 namespace doris {
diff --git a/be/src/olap/like_column_predicate.h b/be/src/olap/like_column_predicate.h
index 11e4c9f07f..ddbe892303 100644
--- a/be/src/olap/like_column_predicate.h
+++ b/be/src/olap/like_column_predicate.h
@@ -16,14 +16,36 @@
 // under the License.
 #pragma once
 
+#include <glog/logging.h>
+#include <stdint.h>
+
+#include <boost/iterator/iterator_facade.hpp>
+#include <functional>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
 #include "olap/column_predicate.h"
-#include "udf/udf.h"
+#include "olap/rowset/segment_v2/bloom_filter.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_dictionary.h"
+#include "vec/columns/column_nullable.h"
 #include "vec/common/string_ref.h"
 #include "vec/core/types.h"
 #include "vec/functions/like.h"
 
+namespace roaring {
+class Roaring;
+} // namespace roaring
+
 namespace doris {
+class FunctionContext;
+
+namespace segment_v2 {
+class BitmapIndexIterator;
+} // namespace segment_v2
 
 class LikeColumnPredicate : public ColumnPredicate {
 public:
diff --git a/be/src/olap/lru_cache.cpp b/be/src/olap/lru_cache.cpp
index 2132db5b2d..000315f773 100644
--- a/be/src/olap/lru_cache.cpp
+++ b/be/src/olap/lru_cache.cpp
@@ -4,17 +4,14 @@
 
 #include "olap/lru_cache.h"
 
-#include <rapidjson/document.h>
-#include <stdio.h>
 #include <stdlib.h>
 
+#include <mutex>
+#include <new>
 #include <sstream>
 #include <string>
 
 #include "gutil/bits.h"
-#include "olap/olap_common.h"
-#include "olap/olap_define.h"
-#include "olap/utils.h"
 #include "runtime/thread_context.h"
 #include "util/doris_metrics.h"
 
diff --git a/be/src/olap/lru_cache.h b/be/src/olap/lru_cache.h
index 599f386e1b..654b9e8964 100644
--- a/be/src/olap/lru_cache.h
+++ b/be/src/olap/lru_cache.h
@@ -4,18 +4,22 @@
 
 #pragma once
 
+#include <assert.h>
+#include <butil/macros.h>
+#include <glog/logging.h>
 #include <gtest/gtest_prod.h>
-#include <rapidjson/document.h>
 #include <stdint.h>
+#include <stdlib.h>
 #include <string.h>
 
+#include <atomic>
 #include <functional>
-#include <queue>
+#include <memory>
+#include <set>
 #include <string>
-#include <vector>
+#include <utility>
 
-#include "olap/olap_common.h"
-#include "runtime/memory/mem_tracker.h"
+#include "runtime/memory/mem_tracker_limiter.h"
 #include "runtime/thread_context.h"
 #include "util/doris_metrics.h"
 #include "util/lock.h"
@@ -49,7 +53,6 @@ namespace doris {
     } while (0)
 
 class Cache;
-class CacheKey;
 
 enum LRUCacheType {
     SIZE,  // The capacity of cache is based on the size of cache entry.
diff --git a/be/src/olap/match_predicate.cpp b/be/src/olap/match_predicate.cpp
index 22d56cd1ff..6837d78f4e 100644
--- a/be/src/olap/match_predicate.cpp
+++ b/be/src/olap/match_predicate.cpp
@@ -17,14 +17,16 @@
 
 #include "olap/match_predicate.h"
 
-#include <string.h>
-
-#include <memory>
-#include <sstream>
+#include <roaring/roaring.hh>
 
 #include "exec/olap_utils.h"
-#include "exprs/string_functions.h"
+#include "olap/field.h"
+#include "olap/olap_common.h"
+#include "olap/rowset/segment_v2/inverted_index_cache.h"
+#include "olap/rowset/segment_v2/inverted_index_reader.h"
 #include "olap/schema.h"
+#include "olap/types.h"
+#include "olap/utils.h"
 #include "vec/common/string_ref.h"
 
 namespace doris {
diff --git a/be/src/olap/match_predicate.h b/be/src/olap/match_predicate.h
index ff41fb00eb..ad793b90f8 100644
--- a/be/src/olap/match_predicate.h
+++ b/be/src/olap/match_predicate.h
@@ -18,17 +18,30 @@
 #ifndef DORIS_BE_SRC_QUERY_EXPRS_MATCH_PREDICATE_H
 #define DORIS_BE_SRC_QUERY_EXPRS_MATCH_PREDICATE_H
 
-#include <memory>
+#include <glog/logging.h>
+#include <stdint.h>
+
+#include <ostream>
 #include <string>
 
-#include "gen_cpp/Exprs_types.h"
+#include "common/status.h"
 #include "olap/column_predicate.h"
-#include "runtime/string_search.hpp"
+#include "olap/schema.h"
+
+namespace roaring {
+class Roaring;
+} // namespace roaring
 
 namespace doris {
 
 enum class MatchType;
 
+namespace segment_v2 {
+class BitmapIndexIterator;
+class InvertedIndexIterator;
+enum class InvertedIndexQueryType;
+} // namespace segment_v2
+
 class MatchPredicate : public ColumnPredicate {
 public:
     MatchPredicate(uint32_t column_id, const std::string& value, MatchType match_type);
diff --git a/be/src/olap/memtable.cpp b/be/src/olap/memtable.cpp
index db18782dc3..9084cc617a 100644
--- a/be/src/olap/memtable.cpp
+++ b/be/src/olap/memtable.cpp
@@ -17,18 +17,41 @@
 
 #include "olap/memtable.h"
 
+#include <fmt/format.h>
+#include <gen_cpp/olap_file.pb.h>
+
+#include <algorithm>
+#include <limits>
+#include <shared_mutex>
+#include <string>
+#include <utility>
+
+#include "common/config.h"
+#include "common/consts.h"
 #include "common/logging.h"
+#include "olap/olap_define.h"
 #include "olap/rowset/beta_rowset.h"
 #include "olap/rowset/rowset_writer.h"
+#include "olap/rowset/segment_v2/segment.h"
 #include "olap/schema.h"
 #include "olap/schema_change.h"
+#include "olap/tablet_schema.h"
+#include "runtime/descriptors.h"
+#include "runtime/exec_env.h"
 #include "runtime/load_channel_mgr.h"
+#include "runtime/thread_context.h"
 #include "util/doris_metrics.h"
+#include "util/runtime_profile.h"
+#include "util/stopwatch.hpp"
 #include "vec/aggregate_functions/aggregate_function_reader.h"
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_object.h"
-#include "vec/core/columns_with_type_and_name.h"
-#include "vec/core/field.h"
+#include "vec/columns/column_string.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/data_types/data_type.h"
+#include "vec/json/path_in_data.h"
 #include "vec/jsonb/serialize.h"
 
 namespace doris {
diff --git a/be/src/olap/memtable.h b/be/src/olap/memtable.h
index e47e696bb0..f50b78c0b1 100644
--- a/be/src/olap/memtable.h
+++ b/be/src/olap/memtable.h
@@ -17,16 +17,22 @@
 
 #pragma once
 
+#include <stddef.h>
+#include <stdint.h>
+
+#include <memory>
 #include <ostream>
+#include <vector>
 
 #include "common/object_pool.h"
-#include "olap/olap_define.h"
+#include "common/status.h"
+#include "olap/olap_common.h"
 #include "olap/skiplist.h"
 #include "olap/tablet.h"
+#include "olap/tablet_meta.h"
 #include "runtime/memory/mem_tracker.h"
 #include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/common/arena.h"
-#include "vec/common/string_ref.h"
 #include "vec/core/block.h"
 
 namespace doris {
@@ -36,6 +42,7 @@ class Schema;
 class SlotDescriptor;
 class TabletSchema;
 class TupleDescriptor;
+enum KeysType : int;
 
 class MemTable {
 public:
diff --git a/be/src/olap/memtable_flush_executor.cpp b/be/src/olap/memtable_flush_executor.cpp
index b40ca93d9d..9c5b92ed10 100644
--- a/be/src/olap/memtable_flush_executor.cpp
+++ b/be/src/olap/memtable_flush_executor.cpp
@@ -17,11 +17,16 @@
 
 #include "olap/memtable_flush_executor.h"
 
-#include <functional>
+#include <gen_cpp/olap_file.pb.h>
+#include <stddef.h>
 
+#include <algorithm>
+#include <ostream>
+
+#include "common/config.h"
+#include "common/logging.h"
 #include "olap/memtable.h"
-#include "runtime/thread_context.h"
-#include "util/scoped_cleanup.h"
+#include "util/stopwatch.hpp"
 #include "util/time.h"
 
 namespace doris {
diff --git a/be/src/olap/memtable_flush_executor.h b/be/src/olap/memtable_flush_executor.h
index ff5e1e08e8..7608fa4d19 100644
--- a/be/src/olap/memtable_flush_executor.h
+++ b/be/src/olap/memtable_flush_executor.h
@@ -17,21 +17,22 @@
 
 #pragma once
 
+#include <stdint.h>
+
 #include <atomic>
+#include <iosfwd>
 #include <memory>
+#include <utility>
 #include <vector>
 
-#include "gen_cpp/olap_file.pb.h"
-#include "olap/olap_define.h"
+#include "common/status.h"
 #include "util/threadpool.h"
 
 namespace doris {
 
 class DataDir;
-class DeltaWriter;
-class ExecEnv;
 class MemTable;
-class MemTrackerLimiter;
+enum RowsetTypePB : int;
 
 // the statistic of a certain flush handler.
 // use atomic because it may be updated by multi threads
diff --git a/be/src/olap/merger.cpp b/be/src/olap/merger.cpp
index a12cc4cc16..437d9ba9eb 100644
--- a/be/src/olap/merger.cpp
+++ b/be/src/olap/merger.cpp
@@ -17,15 +17,39 @@
 
 #include "olap/merger.h"
 
+#include <gen_cpp/olap_file.pb.h>
+#include <gen_cpp/types.pb.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <iterator>
 #include <memory>
+#include <numeric>
+#include <ostream>
+#include <shared_mutex>
+#include <string>
+#include <utility>
 #include <vector>
 
+#include "common/config.h"
+#include "common/logging.h"
+#include "olap/olap_common.h"
 #include "olap/olap_define.h"
-#include "olap/row_cursor.h"
+#include "olap/reader.h"
+#include "olap/rowid_conversion.h"
+#include "olap/rowset/rowset.h"
+#include "olap/rowset/rowset_meta.h"
+#include "olap/rowset/rowset_writer.h"
+#include "olap/rowset/segment_v2/segment_writer.h"
 #include "olap/storage_engine.h"
 #include "olap/tablet.h"
+#include "olap/utils.h"
+#include "util/slice.h"
 #include "util/trace.h"
+#include "vec/core/block.h"
 #include "vec/olap/block_reader.h"
+#include "vec/olap/vertical_block_reader.h"
+#include "vec/olap/vertical_merge_iterator.h"
 
 namespace doris {
 
diff --git a/be/src/olap/merger.h b/be/src/olap/merger.h
index 467c126304..37291c548c 100644
--- a/be/src/olap/merger.h
+++ b/be/src/olap/merger.h
@@ -17,20 +17,29 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include <vector>
+
+#include "common/status.h"
 #include "io/io_common.h"
-#include "olap/olap_define.h"
-#include "olap/rowid_conversion.h"
-#include "olap/rowset/rowset_writer.h"
-#include "olap/rowset/segment_v2/segment_writer.h"
+#include "olap/rowset/rowset_reader.h"
 #include "olap/tablet.h"
-#include "vec/olap/vertical_block_reader.h"
-#include "vec/olap/vertical_merge_iterator.h"
+#include "olap/tablet_schema.h"
 
 namespace doris {
+class KeyBoundsPB;
+class RowIdConversion;
+class RowsetWriter;
+
+namespace segment_v2 {
+class SegmentWriter;
+} // namespace segment_v2
 
 namespace vectorized {
 class RowSourcesBuffer;
-};
+class VerticalBlockReader;
+}; // namespace vectorized
 
 class Merger {
 public:
diff --git a/be/src/olap/null_predicate.cpp b/be/src/olap/null_predicate.cpp
index e9a9bf9705..e4130f0d32 100644
--- a/be/src/olap/null_predicate.cpp
+++ b/be/src/olap/null_predicate.cpp
@@ -17,9 +17,16 @@
 
 #include "olap/null_predicate.h"
 
-#include "olap/field.h"
+#include <string.h>
+
+#include <roaring/roaring.hh>
+
+#include "olap/rowset/segment_v2/bitmap_index_reader.h"
+#include "olap/rowset/segment_v2/inverted_index_cache.h"
+#include "olap/rowset/segment_v2/inverted_index_reader.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_nullable.h"
-#include "vec/common/string_ref.h"
+#include "vec/runtime/vdatetime_value.h"
 
 using namespace doris::vectorized;
 
diff --git a/be/src/olap/null_predicate.h b/be/src/olap/null_predicate.h
index e878907228..ed81cc6f1b 100644
--- a/be/src/olap/null_predicate.h
+++ b/be/src/olap/null_predicate.h
@@ -17,15 +17,31 @@
 
 #pragma once
 
+#include <glog/logging.h>
 #include <stdint.h>
 
-#include <roaring/roaring.hh>
+#include <ostream>
+#include <string>
+#include <utility>
 
+#include "common/status.h"
 #include "olap/column_predicate.h"
 #include "olap/rowset/segment_v2/bloom_filter.h"
+#include "olap/schema.h"
 #include "olap/wrapper_field.h"
 
+namespace roaring {
+class Roaring;
+} // namespace roaring
+
 namespace doris {
+namespace segment_v2 {
+class BitmapIndexIterator;
+class InvertedIndexIterator;
+} // namespace segment_v2
+namespace vectorized {
+class IColumn;
+} // namespace vectorized
 
 class NullPredicate : public ColumnPredicate {
 public:
diff --git a/be/src/olap/olap_common.h b/be/src/olap/olap_common.h
index afb2bdcf22..ad81578a6a 100644
--- a/be/src/olap/olap_common.h
+++ b/be/src/olap/olap_common.h
@@ -17,6 +17,7 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
 #include <netinet/in.h>
 
 #include <cstdint>
@@ -31,7 +32,6 @@
 #include <unordered_map>
 #include <unordered_set>
 
-#include "gen_cpp/Types_types.h"
 #include "io/io_common.h"
 #include "olap/olap_define.h"
 #include "util/hash_util.hpp"
diff --git a/be/src/olap/olap_meta.cpp b/be/src/olap/olap_meta.cpp
index 92b85b1d5b..35ea9b8d8e 100644
--- a/be/src/olap/olap_meta.cpp
+++ b/be/src/olap/olap_meta.cpp
@@ -17,9 +17,16 @@
 
 #include "olap/olap_meta.h"
 
+#include <rocksdb/iterator.h>
+#include <rocksdb/status.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <memory>
 #include <sstream>
 #include <vector>
 
+#include "common/config.h"
 #include "common/logging.h"
 #include "olap/olap_define.h"
 #include "rocksdb/db.h"
diff --git a/be/src/olap/olap_meta.h b/be/src/olap/olap_meta.h
index d2557415b3..a8f1df8d22 100644
--- a/be/src/olap/olap_meta.h
+++ b/be/src/olap/olap_meta.h
@@ -18,11 +18,15 @@
 #pragma once
 
 #include <functional>
-#include <map>
 #include <string>
+#include <vector>
 
-#include "olap/olap_define.h"
-#include "rocksdb/db.h"
+#include "common/status.h"
+
+namespace rocksdb {
+class ColumnFamilyHandle;
+class DB;
+} // namespace rocksdb
 
 namespace doris {
 
diff --git a/be/src/olap/olap_server.cpp b/be/src/olap/olap_server.cpp
index 55b6ef737b..3f21adb229 100644
--- a/be/src/olap/olap_server.cpp
+++ b/be/src/olap/olap_server.cpp
@@ -15,28 +15,55 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <gperftools/profiler.h>
-#include <sys/socket.h>
-#include <unistd.h>
+#include <gen_cpp/Types_types.h>
+#include <stdint.h>
 
-#include <boost/algorithm/string.hpp>
+#include <algorithm>
+#include <atomic>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
 #include <cmath>
+#include <condition_variable>
 #include <ctime>
+#include <functional>
+#include <map>
+#include <memory>
+#include <mutex>
+#include <ostream>
 #include <random>
 #include <string>
+#include <type_traits>
+#include <unordered_set>
+#include <utility>
+#include <vector>
 
 #include "common/config.h"
+#include "common/logging.h"
 #include "common/status.h"
-#include "gutil/strings/substitute.h"
+#include "gutil/ref_counted.h"
 #include "io/cache/file_cache_manager.h"
+#include "io/fs/file_writer.h" // IWYU pragma: keep
+#include "io/fs/path.h"
 #include "olap/cold_data_compaction.h"
-#include "olap/cumulative_compaction.h"
+#include "olap/compaction_permit_limiter.h"
+#include "olap/cumulative_compaction_policy.h"
+#include "olap/data_dir.h"
 #include "olap/olap_common.h"
-#include "olap/olap_define.h"
 #include "olap/rowset/beta_rowset_writer.h"
+#include "olap/rowset/segcompaction.h"
 #include "olap/storage_engine.h"
+#include "olap/tablet.h"
+#include "olap/tablet_manager.h"
+#include "olap/tablet_meta.h"
+#include "olap/tablet_schema.h"
 #include "service/point_query_executor.h"
+#include "util/countdown_latch.h"
+#include "util/doris_metrics.h"
+#include "util/priority_thread_pool.hpp"
+#include "util/thread.h"
+#include "util/threadpool.h"
 #include "util/time.h"
+#include "util/uid_util.h"
 
 using std::string;
 
diff --git a/be/src/olap/options.cpp b/be/src/olap/options.cpp
index 1226ecd415..80af2321e4 100644
--- a/be/src/olap/options.cpp
+++ b/be/src/olap/options.cpp
@@ -17,16 +17,23 @@
 
 #include "olap/options.h"
 
+#include <ctype.h>
 #include <rapidjson/document.h>
+#include <rapidjson/encodings.h>
+#include <rapidjson/rapidjson.h>
+#include <stdlib.h>
 
 #include <algorithm>
+#include <memory>
+#include <ostream>
 
 #include "common/config.h"
 #include "common/logging.h"
 #include "common/status.h"
 #include "gutil/strings/split.h"
-#include "gutil/strings/substitute.h"
+#include "gutil/strings/strip.h"
 #include "io/fs/local_file_system.h"
+#include "olap/olap_define.h"
 #include "olap/utils.h"
 #include "util/path_util.h"
 
diff --git a/be/src/olap/options.h b/be/src/olap/options.h
index c6ed214ceb..84634a92cd 100644
--- a/be/src/olap/options.h
+++ b/be/src/olap/options.h
@@ -17,11 +17,15 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <stdint.h>
+
 #include <string>
+#include <utility>
 #include <vector>
 
+#include "common/status.h"
 #include "io/cache/block/block_file_cache_settings.h"
-#include "olap/olap_define.h"
 #include "util/uid_util.h"
 
 namespace doris {
diff --git a/be/src/olap/page_cache.cpp b/be/src/olap/page_cache.cpp
index 2813f85dd3..53f9378d17 100644
--- a/be/src/olap/page_cache.cpp
+++ b/be/src/olap/page_cache.cpp
@@ -17,7 +17,9 @@
 
 #include "olap/page_cache.h"
 
-#include "runtime/thread_context.h"
+#include <glog/logging.h>
+
+#include <ostream>
 
 namespace doris {
 
diff --git a/be/src/olap/page_cache.h b/be/src/olap/page_cache.h
index 6313e931e0..cab0da2dd7 100644
--- a/be/src/olap/page_cache.h
+++ b/be/src/olap/page_cache.h
@@ -17,14 +17,17 @@
 
 #pragma once
 
+#include <butil/macros.h>
+#include <gen_cpp/segment_v2.pb.h>
+#include <stddef.h>
+#include <stdint.h>
+
 #include <memory>
 #include <string>
 #include <utility>
 
-#include "gen_cpp/segment_v2.pb.h" // for cache allocation
-#include "gutil/macros.h"          // for DISALLOW_COPY_AND_ASSIGN
 #include "olap/lru_cache.h"
-#include "runtime/memory/mem_tracker.h"
+#include "util/slice.h"
 
 namespace doris {
 
diff --git a/be/src/olap/primary_key_index.cpp b/be/src/olap/primary_key_index.cpp
index f9656ee897..e7f11aa3d1 100644
--- a/be/src/olap/primary_key_index.cpp
+++ b/be/src/olap/primary_key_index.cpp
@@ -17,9 +17,17 @@
 
 #include "olap/primary_key_index.h"
 
+#include <gen_cpp/segment_v2.pb.h>
+
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/config.h"
-#include "io/fs/file_reader.h"
+#include "olap/olap_common.h"
+#include "olap/rowset/segment_v2/bloom_filter_index_reader.h"
 #include "olap/rowset/segment_v2/encoding_info.h"
+#include "olap/types.h"
 
 namespace doris {
 
diff --git a/be/src/olap/primary_key_index.h b/be/src/olap/primary_key_index.h
index c80f0f132c..65cc64f0cd 100644
--- a/be/src/olap/primary_key_index.h
+++ b/be/src/olap/primary_key_index.h
@@ -17,11 +17,15 @@
 
 #pragma once
 
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <memory>
+
 #include "common/status.h"
-#include "gen_cpp/segment_v2.pb.h"
-#include "io/fs/file_writer.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "olap/rowset/segment_v2/bloom_filter.h"
-#include "olap/rowset/segment_v2/bloom_filter_index_reader.h"
 #include "olap/rowset/segment_v2/bloom_filter_index_writer.h"
 #include "olap/rowset/segment_v2/indexed_column_reader.h"
 #include "olap/rowset/segment_v2/indexed_column_writer.h"
@@ -29,6 +33,14 @@
 #include "util/slice.h"
 
 namespace doris {
+class TypeInfo;
+
+namespace io {
+class FileWriter;
+} // namespace io
+namespace segment_v2 {
+class PrimaryKeyIndexMetaPB;
+} // namespace segment_v2
 
 // Build index for primary key.
 // The primary key index is designed in a similar way like RocksDB
diff --git a/be/src/olap/push_handler.cpp b/be/src/olap/push_handler.cpp
index 6ed44ea253..81a46753a3 100644
--- a/be/src/olap/push_handler.cpp
+++ b/be/src/olap/push_handler.cpp
@@ -17,20 +17,43 @@
 
 #include "olap/push_handler.h"
 
-#include <algorithm>
-#include <filesystem>
-#include <iostream>
-#include <sstream>
+#include <gen_cpp/AgentService_types.h>
+#include <gen_cpp/Descriptors_types.h>
+#include <gen_cpp/MasterService_types.h>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/olap_file.pb.h>
+#include <gen_cpp/types.pb.h>
 
-#include "common/object_pool.h"
+#include <algorithm>
+#include <iostream>
+#include <mutex>
+#include <new>
+#include <queue>
+#include <shared_mutex>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/config.h"
+#include "common/logging.h"
 #include "common/status.h"
-#include "olap/rowset/rowset_id_generator.h"
-#include "olap/rowset/rowset_meta_manager.h"
-#include "olap/schema_change.h"
+#include "olap/delete_handler.h"
+#include "olap/olap_define.h"
+#include "olap/rowset/rowset_meta.h"
+#include "olap/rowset/rowset_writer.h"
+#include "olap/rowset/rowset_writer_context.h"
+#include "olap/schema.h"
 #include "olap/storage_engine.h"
 #include "olap/tablet.h"
+#include "olap/tablet_manager.h"
 #include "olap/tablet_schema.h"
+#include "olap/txn_manager.h"
+#include "runtime/descriptors.h"
 #include "runtime/exec_env.h"
+#include "util/runtime_profile.h"
+#include "util/time.h"
+#include "vec/core/block.h"
 #include "vec/exec/vparquet_scanner.h"
 
 namespace doris {
diff --git a/be/src/olap/push_handler.h b/be/src/olap/push_handler.h
index 8546fbf749..9f6a91a3d5 100644
--- a/be/src/olap/push_handler.h
+++ b/be/src/olap/push_handler.h
@@ -17,24 +17,36 @@
 
 #pragma once
 
-#include <map>
+#include <butil/macros.h>
+#include <gen_cpp/AgentService_types.h>
+#include <gen_cpp/Exprs_types.h>
+#include <stdint.h>
+
+#include <memory>
 #include <string>
 #include <vector>
 
+#include "common/object_pool.h"
+#include "common/status.h"
 #include "exec/base_scanner.h"
-#include "gen_cpp/AgentService_types.h"
-#include "gen_cpp/MasterService_types.h"
-#include "gen_cpp/PaloInternalService_types.h"
-#include "olap/merger.h"
 #include "olap/olap_common.h"
-#include "olap/row_cursor.h"
 #include "olap/rowset/rowset.h"
+#include "olap/tablet.h"
 #include "olap/tablet_schema.h"
+#include "runtime/runtime_state.h"
 
 namespace doris {
 
-struct ColumnMapping;
-class RowCursor;
+class DescriptorTbl;
+class RuntimeProfile;
+class Schema;
+class TBrokerScanRange;
+class TDescriptorTable;
+class TTabletInfo;
+
+namespace vectorized {
+class Block;
+} // namespace vectorized
 
 class PushHandler {
 public:
diff --git a/be/src/olap/reader.cpp b/be/src/olap/reader.cpp
index b21691af2d..9886db76d6 100644
--- a/be/src/olap/reader.cpp
+++ b/be/src/olap/reader.cpp
@@ -17,20 +17,42 @@
 
 #include "olap/reader.h"
 
+#include <gen_cpp/olap_file.pb.h>
+#include <gen_cpp/segment_v2.pb.h>
+#include <thrift/protocol/TDebugProtocol.h>
+
+#include <algorithm>
+#include <functional>
+#include <iterator>
+#include <numeric>
+#include <ostream>
+#include <shared_mutex>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/config.h"
+#include "common/logging.h"
 #include "common/status.h"
+#include "exprs/bitmapfilter_predicate.h"
+#include "exprs/bloom_filter_func.h"
 #include "exprs/create_predicate_function.h"
 #include "exprs/hybrid_set.h"
-#include "gen_cpp/segment_v2.pb.h"
-#include "olap/bloom_filter_predicate.h"
-#include "olap/comparison_predicate.h"
-#include "olap/in_list_predicate.h"
+#include "olap/column_predicate.h"
 #include "olap/itoken_extractor.h"
 #include "olap/like_column_predicate.h"
 #include "olap/olap_common.h"
+#include "olap/olap_define.h"
 #include "olap/predicate_creator.h"
 #include "olap/row_cursor.h"
+#include "olap/rowset/segment_v2/bloom_filter.h"
 #include "olap/schema.h"
 #include "olap/tablet.h"
+#include "olap/tablet_meta.h"
+#include "runtime/query_fragments_ctx.h"
+#include "runtime/runtime_predicate.h"
+#include "runtime/runtime_state.h"
+#include "vec/common/arena.h"
+#include "vec/core/block.h"
 
 namespace doris {
 using namespace ErrorCode;
diff --git a/be/src/olap/reader.h b/be/src/olap/reader.h
index 5d0cc71126..394ca247b8 100644
--- a/be/src/olap/reader.h
+++ b/be/src/olap/reader.h
@@ -17,31 +17,50 @@
 
 #pragma once
 
-#include <thrift/protocol/TDebugProtocol.h>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <stddef.h>
+#include <stdint.h>
 
-#include "exprs/bitmapfilter_predicate.h"
+#include <memory>
+#include <set>
+#include <string>
+#include <unordered_set>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
 #include "exprs/function_filter.h"
-#include "exprs/hybrid_set.h"
+#include "gutil/strings/substitute.h"
 #include "io/io_common.h"
 #include "olap/delete_handler.h"
+#include "olap/iterators.h"
+#include "olap/olap_common.h"
+#include "olap/olap_tuple.h"
 #include "olap/row_cursor.h"
+#include "olap/rowset/rowset.h"
+#include "olap/rowset/rowset_meta.h"
 #include "olap/rowset/rowset_reader.h"
+#include "olap/rowset/rowset_reader_context.h"
 #include "olap/tablet.h"
 #include "olap/tablet_schema.h"
-#include "util/runtime_profile.h"
-#include "vec/common/arena.h"
-#include "vec/exprs/vexpr_context.h"
 
 namespace doris {
 
-class Tablet;
-class RowCursor;
 class RuntimeState;
+class BitmapFilterFuncBase;
+class BloomFilterFuncBase;
+class ColumnPredicate;
+class DeleteBitmap;
+class HybridSetBase;
+class RuntimeProfile;
 
 namespace vectorized {
 class VCollectIterator;
 class Block;
 class VExpr;
+class Arena;
+class VExprContext;
 } // namespace vectorized
 
 // Used to compare row with input scan key. Scan key only contains key columns,
diff --git a/be/src/olap/row_cursor.cpp b/be/src/olap/row_cursor.cpp
index ec0fd094ff..48756e2366 100644
--- a/be/src/olap/row_cursor.cpp
+++ b/be/src/olap/row_cursor.cpp
@@ -17,10 +17,18 @@
 
 #include "olap/row_cursor.h"
 
-#include <algorithm>
-#include <unordered_set>
+#include <glog/logging.h>
+#include <stdlib.h>
 
-#include "util/stack_util.h"
+#include <algorithm>
+#include <new>
+#include <numeric>
+#include <ostream>
+
+#include "olap/field.h"
+#include "olap/olap_common.h"
+#include "olap/olap_define.h"
+#include "util/slice.h"
 
 using std::min;
 using std::nothrow;
diff --git a/be/src/olap/row_cursor.h b/be/src/olap/row_cursor.h
index 0a0f5f2913..7c19e7ad55 100644
--- a/be/src/olap/row_cursor.h
+++ b/be/src/olap/row_cursor.h
@@ -17,15 +17,19 @@
 
 #pragma once
 
+#include <butil/macros.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <memory>
 #include <string>
 #include <vector>
 
-#include "olap/field.h"
-#include "olap/olap_common.h"
-#include "olap/olap_define.h"
+#include "common/status.h"
 #include "olap/olap_tuple.h"
 #include "olap/row_cursor_cell.h"
 #include "olap/schema.h"
+#include "olap/tablet_schema.h"
 
 namespace doris {
 class Field;
diff --git a/be/src/olap/rowset/beta_rowset.cpp b/be/src/olap/rowset/beta_rowset.cpp
index 513a04a8d1..9e9769bcd4 100644
--- a/be/src/olap/rowset/beta_rowset.cpp
+++ b/be/src/olap/rowset/beta_rowset.cpp
@@ -17,16 +17,23 @@
 
 #include "olap/rowset/beta_rowset.h"
 
-#include <fmt/core.h>
-#include <glog/logging.h>
-#include <stdio.h>  // for remove()
-#include <unistd.h> // for link()
+#include <ctype.h>
+#include <fmt/format.h>
 
+#include <algorithm>
+#include <ostream>
+#include <utility>
+
+#include "common/config.h"
+#include "common/logging.h"
 #include "common/status.h"
-#include "gutil/strings/substitute.h"
 #include "io/cache/file_cache_manager.h"
-#include "io/fs/fs_utils.h"
-#include "io/fs/s3_file_system.h"
+#include "io/fs/file_reader_options.h"
+#include "io/fs/file_system.h"
+#include "io/fs/local_file_system.h"
+#include "io/fs/path.h"
+#include "io/fs/remote_file_system.h"
+#include "olap/olap_common.h"
 #include "olap/olap_define.h"
 #include "olap/rowset/beta_rowset_reader.h"
 #include "olap/rowset/segment_v2/inverted_index_cache.h"
diff --git a/be/src/olap/rowset/beta_rowset.h b/be/src/olap/rowset/beta_rowset.h
index 5401c60860..0ef0c456ae 100644
--- a/be/src/olap/rowset/beta_rowset.h
+++ b/be/src/olap/rowset/beta_rowset.h
@@ -18,22 +18,29 @@
 #ifndef DORIS_SRC_OLAP_ROWSET_BETA_ROWSET_H_
 #define DORIS_SRC_OLAP_ROWSET_BETA_ROWSET_H_
 
-#include <cstdint>
-#include <string>
+#include <stddef.h>
 
-#include "olap/olap_common.h"
-#include "olap/olap_define.h"
+#include <cstdint>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "common/status.h"
 #include "olap/rowset/rowset.h"
 #include "olap/rowset/rowset_meta.h"
 #include "olap/rowset/rowset_reader.h"
 #include "olap/rowset/segment_v2/segment.h"
+#include "olap/tablet_schema.h"
 
 namespace doris {
 
-class BetaRowsetReader;
-class RowsetFactory;
-
 class BetaRowset;
+
+namespace io {
+class RemoteFileSystem;
+} // namespace io
+struct RowsetId;
+
 using BetaRowsetSharedPtr = std::shared_ptr<BetaRowset>;
 
 class BetaRowset : public Rowset {
diff --git a/be/src/olap/rowset/beta_rowset_reader.cpp b/be/src/olap/rowset/beta_rowset_reader.cpp
index 0a4e7ccdb4..e54527e962 100644
--- a/be/src/olap/rowset/beta_rowset_reader.cpp
+++ b/be/src/olap/rowset/beta_rowset_reader.cpp
@@ -17,12 +17,30 @@
 
 #include "beta_rowset_reader.h"
 
+#include <stddef.h>
+
+#include <algorithm>
+#include <ostream>
+#include <roaring/roaring.hh>
+#include <set>
+#include <string>
+#include <unordered_map>
 #include <utility>
 
+#include "common/logging.h"
+#include "io/io_common.h"
+#include "olap/block_column_predicate.h"
+#include "olap/column_predicate.h"
 #include "olap/delete_handler.h"
+#include "olap/olap_define.h"
 #include "olap/row_cursor.h"
+#include "olap/rowset/rowset_meta.h"
+#include "olap/rowset/rowset_reader_context.h"
+#include "olap/rowset/segment_v2/segment.h"
 #include "olap/schema.h"
 #include "olap/tablet_meta.h"
+#include "olap/tablet_schema.h"
+#include "util/runtime_profile.h"
 #include "vec/core/block.h"
 #include "vec/olap/vgeneric_iterators.h"
 
diff --git a/be/src/olap/rowset/beta_rowset_reader.h b/be/src/olap/rowset/beta_rowset_reader.h
index 10cfcb8b0f..de2ecb44d4 100644
--- a/be/src/olap/rowset/beta_rowset_reader.h
+++ b/be/src/olap/rowset/beta_rowset_reader.h
@@ -17,13 +17,27 @@
 
 #pragma once
 
+#include <gen_cpp/olap_file.pb.h>
+#include <stdint.h>
+
+#include <memory>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
 #include "olap/iterators.h"
-#include "olap/row_cursor.h"
+#include "olap/olap_common.h"
 #include "olap/rowset/beta_rowset.h"
+#include "olap/rowset/rowset.h"
 #include "olap/rowset/rowset_reader.h"
 #include "olap/segment_loader.h"
+#include "vec/core/block.h"
 
 namespace doris {
+class RuntimeProfile;
+class Schema;
+struct RowLocation;
+struct RowsetReaderContext;
 
 class BetaRowsetReader : public RowsetReader {
 public:
diff --git a/be/src/olap/rowset/beta_rowset_writer.cpp b/be/src/olap/rowset/beta_rowset_writer.cpp
index 8b9b04a82c..744cf8066c 100644
--- a/be/src/olap/rowset/beta_rowset_writer.cpp
+++ b/be/src/olap/rowset/beta_rowset_writer.cpp
@@ -17,35 +17,46 @@
 
 #include "olap/rowset/beta_rowset_writer.h"
 
+#include <assert.h>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
+#include <stdio.h>
+
 #include <ctime> // time
+#include <filesystem>
 #include <memory>
 #include <sstream>
+#include <utility>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/config.h"
 #include "common/logging.h"
+#include "gutil/integral_types.h"
 #include "gutil/strings/substitute.h"
+#include "io/fs/file_reader_options.h"
+#include "io/fs/file_system.h"
 #include "io/fs/file_writer.h"
-#include "olap/memtable.h"
-#include "olap/merger.h"
+#include "olap/data_dir.h"
 #include "olap/olap_define.h"
-#include "olap/row_cursor.h" // RowCursor
 #include "olap/rowset/beta_rowset.h"
 #include "olap/rowset/rowset_factory.h"
 #include "olap/rowset/segment_v2/inverted_index_cache.h"
 #include "olap/rowset/segment_v2/inverted_index_desc.h"
+#include "olap/rowset/segment_v2/segment.h"
 #include "olap/rowset/segment_v2/segment_writer.h"
 #include "olap/storage_engine.h"
-#include "runtime/exec_env.h"
-#include "runtime/memory/mem_tracker_limiter.h"
+#include "olap/tablet.h"
+#include "olap/tablet_schema.h"
 #include "segcompaction.h"
+#include "util/slice.h"
+#include "util/time.h"
 #include "vec/common/schema_util.h" // LocalSchemaChangeRecorder
-#include "vec/jsonb/serialize.h"
+#include "vec/core/block.h"
 
 namespace doris {
 using namespace ErrorCode;
 
-class StorageEngine;
-
 BetaRowsetWriter::BetaRowsetWriter()
         : _rowset_meta(nullptr),
           _num_segment(0),
diff --git a/be/src/olap/rowset/beta_rowset_writer.h b/be/src/olap/rowset/beta_rowset_writer.h
index 25c05318e9..c6fa740a58 100644
--- a/be/src/olap/rowset/beta_rowset_writer.h
+++ b/be/src/olap/rowset/beta_rowset_writer.h
@@ -17,22 +17,41 @@
 
 #pragma once
 
+#include <fmt/format.h>
+#include <gen_cpp/olap_file.pb.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <atomic>
+#include <condition_variable>
+#include <map>
+#include <memory>
+#include <mutex>
+#include <string>
+#include <unordered_set>
+#include <vector>
+
+#include "common/status.h"
+#include "io/fs/file_reader_writer_fwd.h"
+#include "olap/olap_common.h"
+#include "olap/rowset/rowset.h"
+#include "olap/rowset/rowset_meta.h"
 #include "olap/rowset/rowset_writer.h"
+#include "olap/rowset/rowset_writer_context.h"
 #include "segcompaction.h"
 #include "segment_v2/segment.h"
-#include "vec/columns/column.h"
-#include "vec/olap/vertical_block_reader.h"
-#include "vec/olap/vgeneric_iterators.h"
+#include "util/spinlock.h"
 
 namespace doris {
+namespace vectorized {
+class Block;
+} // namespace vectorized
+
 namespace segment_v2 {
 class SegmentWriter;
 } // namespace segment_v2
 
-namespace io {
-class FileWriter;
-} // namespace io
-
 using SegCompactionCandidates = std::vector<segment_v2::SegmentSharedPtr>;
 using SegCompactionCandidatesSharedPtr = std::shared_ptr<SegCompactionCandidates>;
 namespace vectorized::schema_util {
diff --git a/be/src/olap/rowset/rowset.cpp b/be/src/olap/rowset/rowset.cpp
index d29c02ed95..d9aa8fb05e 100644
--- a/be/src/olap/rowset/rowset.cpp
+++ b/be/src/olap/rowset/rowset.cpp
@@ -17,8 +17,10 @@
 
 #include "olap/rowset/rowset.h"
 
+#include <gen_cpp/olap_file.pb.h>
+
+#include "olap/olap_define.h"
 #include "olap/tablet_schema.h"
-#include "olap/tablet_schema_cache.h"
 #include "util/time.h"
 
 namespace doris {
diff --git a/be/src/olap/rowset/rowset.h b/be/src/olap/rowset/rowset.h
index c2ae5f1b05..198dd8a0cd 100644
--- a/be/src/olap/rowset/rowset.h
+++ b/be/src/olap/rowset/rowset.h
@@ -17,26 +17,36 @@
 
 #pragma once
 
+#include <butil/macros.h>
+#include <fmt/format.h>
+#include <gen_cpp/olap_file.pb.h>
+#include <gen_cpp/types.pb.h>
+#include <stddef.h>
+#include <stdint.h>
+
 #include <atomic>
 #include <memory>
 #include <mutex>
+#include <ostream>
+#include <string>
 #include <vector>
 
-#include "gen_cpp/olap_file.pb.h"
-#include "gutil/macros.h"
-#include "io/fs/remote_file_system.h"
+#include "common/logging.h"
+#include "common/status.h"
+#include "olap/olap_common.h"
 #include "olap/rowset/rowset_meta.h"
 #include "olap/tablet_schema.h"
 #include "util/lock.h"
 
 namespace doris {
 
-class DataDir;
-class OlapTuple;
-class RowCursor;
 class Rowset;
+
+namespace io {
+class RemoteFileSystem;
+} // namespace io
+
 using RowsetSharedPtr = std::shared_ptr<Rowset>;
-class RowsetFactory;
 class RowsetReader;
 
 // the rowset state transfer graph:
diff --git a/be/src/olap/rowset/rowset_factory.cpp b/be/src/olap/rowset/rowset_factory.cpp
index 78f6b2326c..e66004da70 100644
--- a/be/src/olap/rowset/rowset_factory.cpp
+++ b/be/src/olap/rowset/rowset_factory.cpp
@@ -17,12 +17,15 @@
 
 #include "olap/rowset/rowset_factory.h"
 
+#include <gen_cpp/olap_file.pb.h>
+
 #include <memory>
 
 #include "beta_rowset.h"
-#include "gen_cpp/olap_file.pb.h"
+#include "io/fs/file_writer.h" // IWYU pragma: keep
 #include "olap/rowset/beta_rowset_writer.h"
 #include "olap/rowset/rowset_writer.h"
+#include "olap/rowset/rowset_writer_context.h"
 #include "olap/rowset/vertical_beta_rowset_writer.h"
 
 namespace doris {
diff --git a/be/src/olap/rowset/rowset_factory.h b/be/src/olap/rowset/rowset_factory.h
index d0adc6f457..9c3543c184 100644
--- a/be/src/olap/rowset/rowset_factory.h
+++ b/be/src/olap/rowset/rowset_factory.h
@@ -18,9 +18,13 @@
 #ifndef DORIS_BE_SRC_OLAP_ROWSET_FACTORY_H
 #define DORIS_BE_SRC_OLAP_ROWSET_FACTORY_H
 
-#include "gen_cpp/olap_file.pb.h"
-#include "olap/data_dir.h"
+#include <memory>
+#include <string>
+
+#include "common/status.h"
 #include "olap/rowset/rowset.h"
+#include "olap/rowset/rowset_meta.h"
+#include "olap/tablet_schema.h"
 
 namespace doris {
 
diff --git a/be/src/olap/rowset/rowset_meta.h b/be/src/olap/rowset/rowset_meta.h
index 9697055864..f9491e1553 100644
--- a/be/src/olap/rowset/rowset_meta.h
+++ b/be/src/olap/rowset/rowset_meta.h
@@ -18,12 +18,13 @@
 #ifndef DORIS_BE_SRC_OLAP_ROWSET_ROWSET_META_H
 #define DORIS_BE_SRC_OLAP_ROWSET_ROWSET_META_H
 
+#include <gen_cpp/olap_file.pb.h>
+
 #include <memory>
 #include <string>
 #include <vector>
 
 #include "common/logging.h"
-#include "gen_cpp/olap_file.pb.h"
 #include "google/protobuf/util/message_differencer.h"
 #include "io/fs/file_system.h"
 #include "io/fs/local_file_system.h"
diff --git a/be/src/olap/rowset/rowset_meta_manager.cpp b/be/src/olap/rowset/rowset_meta_manager.cpp
index 84e8be217f..3373db47d8 100644
--- a/be/src/olap/rowset/rowset_meta_manager.cpp
+++ b/be/src/olap/rowset/rowset_meta_manager.cpp
@@ -17,17 +17,18 @@
 
 #include "olap/rowset/rowset_meta_manager.h"
 
+#include <gen_cpp/olap_file.pb.h>
+
 #include <boost/algorithm/string/trim.hpp>
 #include <fstream>
-#include <sstream>
+#include <memory>
+#include <new>
 #include <string>
 #include <vector>
 
 #include "common/logging.h"
-#include "json2pb/json_to_pb.h"
-#include "json2pb/pb_to_json.h"
 #include "olap/olap_define.h"
-#include "olap/storage_engine.h"
+#include "olap/olap_meta.h"
 #include "olap/utils.h"
 
 namespace doris {
diff --git a/be/src/olap/rowset/rowset_meta_manager.h b/be/src/olap/rowset/rowset_meta_manager.h
index cca2304ce3..f6f193caae 100644
--- a/be/src/olap/rowset/rowset_meta_manager.h
+++ b/be/src/olap/rowset/rowset_meta_manager.h
@@ -18,11 +18,18 @@
 #ifndef DORIS_BE_SRC_OLAP_ROWSET_ROWSET_META_MANAGER_H
 #define DORIS_BE_SRC_OLAP_ROWSET_ROWSET_META_MANAGER_H
 
+#include <functional>
 #include <string>
 
-#include "olap/olap_meta.h"
+#include "common/status.h"
+#include "olap/olap_common.h"
 #include "olap/rowset/rowset_meta.h"
 
+namespace doris {
+class OlapMeta;
+class RowsetMetaPB;
+} // namespace doris
+
 using std::string;
 
 namespace doris {
diff --git a/be/src/olap/rowset/rowset_reader.h b/be/src/olap/rowset/rowset_reader.h
index 3d6b2c6fad..dadfd0e05c 100644
--- a/be/src/olap/rowset/rowset_reader.h
+++ b/be/src/olap/rowset/rowset_reader.h
@@ -18,10 +18,11 @@
 #ifndef DORIS_BE_SRC_OLAP_ROWSET_ROWSET_READER_H
 #define DORIS_BE_SRC_OLAP_ROWSET_ROWSET_READER_H
 
+#include <gen_cpp/olap_file.pb.h>
+
 #include <memory>
 #include <unordered_map>
 
-#include "gen_cpp/olap_file.pb.h"
 #include "olap/iterators.h"
 #include "olap/rowset/rowset.h"
 #include "olap/rowset/rowset_reader_context.h"
diff --git a/be/src/olap/rowset/rowset_tree.cpp b/be/src/olap/rowset/rowset_tree.cpp
index a9cc6a81e0..8ac153435f 100644
--- a/be/src/olap/rowset/rowset_tree.cpp
+++ b/be/src/olap/rowset/rowset_tree.cpp
@@ -21,8 +21,10 @@
 
 #include "olap/rowset/rowset_tree.h"
 
+#include <gen_cpp/olap_file.pb.h>
 #include <glog/logging.h>
 
+#include <algorithm>
 #include <cstddef>
 #include <functional>
 #include <iterator>
@@ -36,6 +38,7 @@
 #include "olap/rowset/rowset.h"
 #include "olap/rowset/rowset_meta.h"
 #include "util/interval_tree-inl.h"
+#include "util/interval_tree.h"
 #include "util/slice.h"
 
 using std::shared_ptr;
diff --git a/be/src/olap/rowset/rowset_tree.h b/be/src/olap/rowset/rowset_tree.h
index 8228bc26d7..5ebabca0f1 100644
--- a/be/src/olap/rowset/rowset_tree.h
+++ b/be/src/olap/rowset/rowset_tree.h
@@ -21,15 +21,17 @@
 
 #pragma once
 
-#include <boost/optional/optional.hpp>
 #include <cstdint>
 #include <functional>
 #include <memory>
+#include <optional>
 #include <unordered_map>
+#include <utility>
 #include <vector>
 
 #include "common/status.h"
 #include "gutil/map-util.h"
+#include "olap/olap_common.h"
 #include "olap/rowset/rowset.h"
 #include "util/slice.h"
 
@@ -37,7 +39,6 @@ namespace doris {
 
 template <class Traits>
 class IntervalTree;
-
 struct RowsetIntervalTraits;
 struct RowsetWithBounds;
 
diff --git a/be/src/olap/rowset/rowset_writer.h b/be/src/olap/rowset/rowset_writer.h
index f4d55ac051..6249f01637 100644
--- a/be/src/olap/rowset/rowset_writer.h
+++ b/be/src/olap/rowset/rowset_writer.h
@@ -17,8 +17,9 @@
 
 #pragma once
 
-#include "gen_cpp/olap_file.pb.h"
-#include "gen_cpp/types.pb.h"
+#include <gen_cpp/olap_file.pb.h>
+#include <gen_cpp/types.pb.h>
+
 #include "gutil/macros.h"
 #include "olap/column_mapping.h"
 #include "olap/rowset/rowset.h"
diff --git a/be/src/olap/rowset/rowset_writer_context.h b/be/src/olap/rowset/rowset_writer_context.h
index fc33a521bb..8d4fabfc88 100644
--- a/be/src/olap/rowset/rowset_writer_context.h
+++ b/be/src/olap/rowset/rowset_writer_context.h
@@ -17,7 +17,8 @@
 
 #pragma once
 
-#include "gen_cpp/olap_file.pb.h"
+#include <gen_cpp/olap_file.pb.h>
+
 #include "io/fs/file_system.h"
 #include "olap/tablet.h"
 #include "olap/tablet_schema.h"
diff --git a/be/src/olap/rowset/segcompaction.cpp b/be/src/olap/rowset/segcompaction.cpp
index cc151afdac..8eb60eae7f 100644
--- a/be/src/olap/rowset/segcompaction.cpp
+++ b/be/src/olap/rowset/segcompaction.cpp
@@ -17,30 +17,50 @@
 
 #include "segcompaction.h"
 
-#include <fmt/core.h>
+#include <fmt/format.h>
+#include <gen_cpp/olap_file.pb.h>
+#include <limits.h>
 
+#include <algorithm>
+#include <atomic>
+#include <condition_variable>
 #include <filesystem>
-#include <iostream>
+#include <map>
 #include <mutex>
-#include <thread>
+#include <sstream>
+#include <string>
+#include <utility>
 
 #include "beta_rowset_writer.h"
-#include "common/config.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/logging.h"
+#include "gutil/stringprintf.h"
 #include "gutil/strings/substitute.h"
+#include "io/fs/file_system.h"
 #include "io/fs/file_writer.h"
-#include "olap/memtable.h"
+#include "io/io_common.h"
+#include "olap/data_dir.h"
+#include "olap/iterators.h"
 #include "olap/merger.h"
+#include "olap/olap_common.h"
 #include "olap/olap_define.h"
-#include "olap/row_cursor.h" // RowCursor
+#include "olap/reader.h"
 #include "olap/rowset/beta_rowset.h"
+#include "olap/rowset/rowset_meta.h"
+#include "olap/rowset/rowset_writer_context.h"
 #include "olap/rowset/segment_v2/inverted_index_cache.h"
 #include "olap/rowset/segment_v2/inverted_index_desc.h"
+#include "olap/rowset/segment_v2/segment.h"
+#include "olap/rowset/segment_v2/segment_writer.h"
+#include "olap/schema.h"
 #include "olap/storage_engine.h"
-#include "runtime/exec_env.h"
+#include "olap/tablet_schema.h"
 #include "runtime/memory/mem_tracker_limiter.h"
-#include "vec/common/schema_util.h" // LocalSchemaChangeRecorder
-#include "vec/jsonb/serialize.h"
+#include "runtime/thread_context.h"
+#include "util/time.h"
+#include "vec/olap/vertical_block_reader.h"
+#include "vec/olap/vertical_merge_iterator.h"
 
 namespace doris {
 using namespace ErrorCode;
diff --git a/be/src/olap/rowset/segcompaction.h b/be/src/olap/rowset/segcompaction.h
index 871f864b9b..33fcf4ba28 100644
--- a/be/src/olap/rowset/segcompaction.h
+++ b/be/src/olap/rowset/segcompaction.h
@@ -18,18 +18,27 @@
 #pragma once
 #include <stdint.h>
 
-#include <ctime> // time
 #include <memory>
-#include <sstream>
+#include <vector>
 
 #include "common/status.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "olap/merger.h"
-#include "olap/rowset/rowset_writer.h"
+#include "olap/tablet.h"
 #include "segment_v2/segment.h"
-#include "segment_v2/segment_writer.h"
-#include "vec/olap/vertical_block_reader.h"
 
 namespace doris {
+class Schema;
+
+namespace segment_v2 {
+class SegmentWriter;
+} // namespace segment_v2
+namespace vectorized {
+class RowSourcesBuffer;
+class VerticalBlockReader;
+} // namespace vectorized
+struct OlapReaderStatistics;
+
 using SegCompactionCandidates = std::vector<segment_v2::SegmentSharedPtr>;
 using SegCompactionCandidatesSharedPtr = std::shared_ptr<SegCompactionCandidates>;
 
diff --git a/be/src/olap/rowset/segment_v2/binary_dict_page.cpp b/be/src/olap/rowset/segment_v2/binary_dict_page.cpp
index 4ac002400e..94057fe5da 100644
--- a/be/src/olap/rowset/segment_v2/binary_dict_page.cpp
+++ b/be/src/olap/rowset/segment_v2/binary_dict_page.cpp
@@ -17,11 +17,26 @@
 
 #include "olap/rowset/segment_v2/binary_dict_page.h"
 
+#include <gen_cpp/segment_v2.pb.h>
+
+#include <algorithm>
+#include <ostream>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/logging.h"
+#include "gutil/casts.h"
+#include "gutil/port.h"
 #include "gutil/strings/substitute.h" // for Substitute
-#include "util/slice.h"               // for Slice
+#include "olap/rowset/segment_v2/bitshuffle_page.h"
+#include "util/coding.h"
+#include "util/slice.h" // for Slice
 #include "vec/columns/column.h"
 
 namespace doris {
+struct StringRef;
+
 namespace segment_v2 {
 
 using strings::Substitute;
diff --git a/be/src/olap/rowset/segment_v2/binary_dict_page.h b/be/src/olap/rowset/segment_v2/binary_dict_page.h
index b570b042dc..0ab50ce634 100644
--- a/be/src/olap/rowset/segment_v2/binary_dict_page.h
+++ b/be/src/olap/rowset/segment_v2/binary_dict_page.h
@@ -17,21 +17,33 @@
 
 #pragma once
 
-#include <functional>
-#include <memory>
-#include <string>
+#include <parallel_hashmap/phmap.h>
+#include <stddef.h>
+#include <stdint.h>
 
-#include "gen_cpp/segment_v2.pb.h"
+#include <memory>
+#include <vector>
+
+#include "common/status.h"
 #include "gutil/hash/string_hash.h"
 #include "olap/olap_common.h"
 #include "olap/rowset/segment_v2/binary_plain_page.h"
-#include "olap/rowset/segment_v2/bitshuffle_page.h"
 #include "olap/rowset/segment_v2/common.h"
 #include "olap/rowset/segment_v2/options.h"
+#include "olap/rowset/segment_v2/page_builder.h"
+#include "olap/rowset/segment_v2/page_decoder.h"
+#include "util/faststring.h"
+#include "util/slice.h"
 #include "vec/common/arena.h"
+#include "vec/data_types/data_type.h"
 
 namespace doris {
+struct StringRef;
+
 namespace segment_v2 {
+enum EncodingTypePB : int;
+template <FieldType Type>
+class BitShufflePageDecoder;
 
 enum { BINARY_DICT_PAGE_HEADER_SIZE = 4 };
 
diff --git a/be/src/olap/rowset/segment_v2/binary_prefix_page.cpp b/be/src/olap/rowset/segment_v2/binary_prefix_page.cpp
index 90029ea09c..e3e893a0dd 100644
--- a/be/src/olap/rowset/segment_v2/binary_prefix_page.cpp
+++ b/be/src/olap/rowset/segment_v2/binary_prefix_page.cpp
@@ -19,13 +19,11 @@
 
 #include <stddef.h>
 #include <stdint.h>
-#include <string.h>
 
-#include <map>
-#include <string>
+#include <algorithm>
 #include <vector>
 
-#include "common/logging.h"
+#include "gutil/port.h"
 #include "gutil/strings/substitute.h"
 #include "util/coding.h"
 #include "util/faststring.h"
diff --git a/be/src/olap/rowset/segment_v2/binary_prefix_page.h b/be/src/olap/rowset/segment_v2/binary_prefix_page.h
index 2930a8e616..b6ca02190f 100644
--- a/be/src/olap/rowset/segment_v2/binary_prefix_page.h
+++ b/be/src/olap/rowset/segment_v2/binary_prefix_page.h
@@ -17,20 +17,20 @@
 
 #pragma once
 
+#include <glog/logging.h>
 #include <stddef.h>
 #include <stdint.h>
-#include <string.h>
 
-#include <map>
-#include <string>
 #include <vector>
 
+#include "common/status.h"
 #include "olap/rowset/segment_v2/options.h"
 #include "olap/rowset/segment_v2/page_builder.h"
 #include "olap/rowset/segment_v2/page_decoder.h"
 #include "util/coding.h"
 #include "util/faststring.h"
 #include "util/slice.h"
+#include "vec/columns/column.h"
 
 namespace doris {
 namespace segment_v2 {
diff --git a/be/src/olap/rowset/segment_v2/bitmap_index_reader.cpp b/be/src/olap/rowset/segment_v2/bitmap_index_reader.cpp
index c101e3e7da..0e2c26ff30 100644
--- a/be/src/olap/rowset/segment_v2/bitmap_index_reader.cpp
+++ b/be/src/olap/rowset/segment_v2/bitmap_index_reader.cpp
@@ -17,7 +17,16 @@
 
 #include "olap/rowset/segment_v2/bitmap_index_reader.h"
 
+#include <gen_cpp/segment_v2.pb.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <roaring/roaring.hh>
+
 #include "olap/types.h"
+#include "vec/columns/column.h"
+#include "vec/common/string_ref.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_factory.hpp"
 
 namespace doris {
diff --git a/be/src/olap/rowset/segment_v2/bitmap_index_reader.h b/be/src/olap/rowset/segment_v2/bitmap_index_reader.h
index e5fc28adcd..c448c9f5cc 100644
--- a/be/src/olap/rowset/segment_v2/bitmap_index_reader.h
+++ b/be/src/olap/rowset/segment_v2/bitmap_index_reader.h
@@ -17,23 +17,28 @@
 
 #pragma once
 
-#include <roaring/roaring.hh>
+#include <stdint.h>
+
+#include <memory>
+#include <utility>
 
 #include "common/status.h"
-#include "gen_cpp/segment_v2.pb.h"
-#include "io/fs/file_reader.h"
+#include "io/fs/file_reader_writer_fwd.h"
+#include "olap/olap_common.h"
 #include "olap/rowset/segment_v2/common.h"
 #include "olap/rowset/segment_v2/indexed_column_reader.h"
+#include "olap/types.h"
+
+namespace roaring {
+class Roaring;
+} // namespace roaring
 
 namespace doris {
 
-class TypeInfo;
-
 namespace segment_v2 {
 
 class BitmapIndexIterator;
-class IndexedColumnReader;
-class IndexedColumnIterator;
+class BitmapIndexPB;
 
 class BitmapIndexReader {
 public:
diff --git a/be/src/olap/rowset/segment_v2/bitmap_index_writer.cpp b/be/src/olap/rowset/segment_v2/bitmap_index_writer.cpp
index 957da3dc3d..8523740920 100644
--- a/be/src/olap/rowset/segment_v2/bitmap_index_writer.cpp
+++ b/be/src/olap/rowset/segment_v2/bitmap_index_writer.cpp
@@ -17,9 +17,15 @@
 
 #include "olap/rowset/segment_v2/bitmap_index_writer.h"
 
+#include <gen_cpp/segment_v2.pb.h>
+
+#include <algorithm>
 #include <map>
 #include <roaring/roaring.hh>
+#include <string>
+#include <vector>
 
+#include "olap/olap_common.h"
 #include "olap/rowset/segment_v2/common.h"
 #include "olap/rowset/segment_v2/encoding_info.h"
 #include "olap/rowset/segment_v2/indexed_column_writer.h"
diff --git a/be/src/olap/rowset/segment_v2/bitmap_index_writer.h b/be/src/olap/rowset/segment_v2/bitmap_index_writer.h
index 93142728a4..cf7de5efe9 100644
--- a/be/src/olap/rowset/segment_v2/bitmap_index_writer.h
+++ b/be/src/olap/rowset/segment_v2/bitmap_index_writer.h
@@ -17,12 +17,13 @@
 
 #pragma once
 
+#include <butil/macros.h>
+#include <stdint.h>
+
 #include <cstddef>
 #include <memory>
 
 #include "common/status.h"
-#include "gen_cpp/segment_v2.pb.h"
-#include "gutil/macros.h"
 
 namespace doris {
 
@@ -33,6 +34,7 @@ class FileWriter;
 }
 
 namespace segment_v2 {
+class ColumnIndexMetaPB;
 
 class BitmapIndexWriter {
 public:
diff --git a/be/src/olap/rowset/segment_v2/bitshuffle_page.cpp b/be/src/olap/rowset/segment_v2/bitshuffle_page.cpp
index 3afa0f4422..cd6f1f4e91 100644
--- a/be/src/olap/rowset/segment_v2/bitshuffle_page.cpp
+++ b/be/src/olap/rowset/segment_v2/bitshuffle_page.cpp
@@ -17,8 +17,6 @@
 
 #include "olap/rowset/segment_v2/bitshuffle_page.h"
 
-#include "olap/rowset/segment_v2/common.h"
-
 namespace doris {
 
 namespace segment_v2 {
diff --git a/be/src/olap/rowset/segment_v2/bitshuffle_page.h b/be/src/olap/rowset/segment_v2/bitshuffle_page.h
index 767ff35cfe..657a34e7c1 100644
--- a/be/src/olap/rowset/segment_v2/bitshuffle_page.h
+++ b/be/src/olap/rowset/segment_v2/bitshuffle_page.h
@@ -18,13 +18,15 @@
 #pragma once
 
 #include <glog/logging.h>
-#include <sys/types.h>
 
 #include <algorithm>
 #include <cstdint>
 #include <cstring>
 #include <ostream>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/status.h"
 #include "gutil/port.h"
 #include "olap/olap_common.h"
 #include "olap/rowset/segment_v2/bitshuffle_wrapper.h"
@@ -33,12 +35,12 @@
 #include "olap/rowset/segment_v2/page_builder.h"
 #include "olap/rowset/segment_v2/page_decoder.h"
 #include "olap/types.h"
-#include "runtime/memory/chunk_allocator.h"
+#include "util/alignment.h"
 #include "util/coding.h"
 #include "util/faststring.h"
 #include "util/slice.h"
-#include "vec/columns/column_nullable.h"
-#include "vec/runtime/vdatetime_value.h"
+#include "vec/columns/column.h"
+#include "vec/data_types/data_type.h"
 
 namespace doris {
 namespace segment_v2 {
diff --git a/be/src/olap/rowset/segment_v2/block_split_bloom_filter.cpp b/be/src/olap/rowset/segment_v2/block_split_bloom_filter.cpp
index da842d491b..871da736d5 100644
--- a/be/src/olap/rowset/segment_v2/block_split_bloom_filter.cpp
+++ b/be/src/olap/rowset/segment_v2/block_split_bloom_filter.cpp
@@ -17,7 +17,7 @@
 
 #include "olap/rowset/segment_v2/block_split_bloom_filter.h"
 
-#include "util/debug_util.h"
+#include <glog/logging.h>
 
 namespace doris {
 namespace segment_v2 {
diff --git a/be/src/olap/rowset/segment_v2/block_split_bloom_filter.h b/be/src/olap/rowset/segment_v2/block_split_bloom_filter.h
index 93591adc1d..f68ddd7e74 100644
--- a/be/src/olap/rowset/segment_v2/block_split_bloom_filter.h
+++ b/be/src/olap/rowset/segment_v2/block_split_bloom_filter.h
@@ -17,6 +17,8 @@
 
 #pragma once
 
+#include <stdint.h>
+
 #include "olap/rowset/segment_v2/bloom_filter.h"
 
 namespace doris {
diff --git a/be/src/olap/rowset/segment_v2/bloom_filter.cpp b/be/src/olap/rowset/segment_v2/bloom_filter.cpp
index e11fe6ca1d..12a7f2395e 100644
--- a/be/src/olap/rowset/segment_v2/bloom_filter.cpp
+++ b/be/src/olap/rowset/segment_v2/bloom_filter.cpp
@@ -15,14 +15,17 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <cstdint>
-#include <functional>
+#include "olap/rowset/segment_v2/bloom_filter.h"
 
-#include "gen_cpp/segment_v2.pb.h"
-#include "gutil/strings/substitute.h"
+#include <gen_cpp/segment_v2.pb.h>
+#include <math.h>
+
+#include <cstdint>
+#include <memory>
+
+#include "common/status.h"
 #include "olap/rowset/segment_v2/block_split_bloom_filter.h"
 #include "olap/rowset/segment_v2/ngram_bloom_filter.h"
-#include "olap/utils.h"
 
 namespace doris {
 namespace segment_v2 {
diff --git a/be/src/olap/rowset/segment_v2/bloom_filter.h b/be/src/olap/rowset/segment_v2/bloom_filter.h
index c235ec39b9..969d7b8aa9 100644
--- a/be/src/olap/rowset/segment_v2/bloom_filter.h
+++ b/be/src/olap/rowset/segment_v2/bloom_filter.h
@@ -17,14 +17,15 @@
 
 #pragma once
 
+#include <gen_cpp/segment_v2.pb.h>
+#include <glog/logging.h>
+#include <string.h>
+
 #include <cstdint>
 #include <functional>
 #include <memory>
 
 #include "common/status.h"
-#include "gen_cpp/segment_v2.pb.h"
-#include "gutil/strings/substitute.h"
-#include "olap/utils.h"
 #include "util/murmur_hash3.h"
 
 namespace doris {
diff --git a/be/src/olap/rowset/segment_v2/bloom_filter_index_reader.cpp b/be/src/olap/rowset/segment_v2/bloom_filter_index_reader.cpp
index cecbd8c4ca..e3ae352dd7 100644
--- a/be/src/olap/rowset/segment_v2/bloom_filter_index_reader.cpp
+++ b/be/src/olap/rowset/segment_v2/bloom_filter_index_reader.cpp
@@ -17,8 +17,14 @@
 
 #include "olap/rowset/segment_v2/bloom_filter_index_reader.h"
 
+#include <gen_cpp/segment_v2.pb.h>
+#include <glog/logging.h>
+
 #include "olap/rowset/segment_v2/bloom_filter.h"
 #include "olap/types.h"
+#include "vec/columns/column.h"
+#include "vec/common/string_ref.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_factory.hpp"
 
 namespace doris {
diff --git a/be/src/olap/rowset/segment_v2/bloom_filter_index_reader.h b/be/src/olap/rowset/segment_v2/bloom_filter_index_reader.h
index bb9a02c244..452a7ea2ae 100644
--- a/be/src/olap/rowset/segment_v2/bloom_filter_index_reader.h
+++ b/be/src/olap/rowset/segment_v2/bloom_filter_index_reader.h
@@ -17,26 +17,25 @@
 
 #pragma once
 
-#include <map>
+#include <stddef.h>
+
 #include <memory>
+#include <utility>
 
 #include "common/status.h"
-#include "gen_cpp/segment_v2.pb.h"
-#include "io/fs/file_reader.h"
+#include "io/fs/file_reader_writer_fwd.h"
+#include "olap/olap_common.h"
 #include "olap/rowset/segment_v2/common.h"
 #include "olap/rowset/segment_v2/indexed_column_reader.h"
-#include "olap/rowset/segment_v2/row_ranges.h"
+#include "olap/types.h"
 
 namespace doris {
 
-class TypeInfo;
-
 namespace segment_v2 {
 
 class BloomFilterIndexIterator;
-class IndexedColumnReader;
-class IndexedColumnIterator;
 class BloomFilter;
+class BloomFilterIndexPB;
 
 class BloomFilterIndexReader {
 public:
diff --git a/be/src/olap/rowset/segment_v2/bloom_filter_index_writer.cpp b/be/src/olap/rowset/segment_v2/bloom_filter_index_writer.cpp
index 588f2eb997..34ff479733 100644
--- a/be/src/olap/rowset/segment_v2/bloom_filter_index_writer.cpp
+++ b/be/src/olap/rowset/segment_v2/bloom_filter_index_writer.cpp
@@ -17,16 +17,21 @@
 
 #include "olap/rowset/segment_v2/bloom_filter_index_writer.h"
 
-#include <map>
-#include <roaring/roaring.hh>
+#include <gen_cpp/segment_v2.pb.h>
+#include <string.h>
 
+#include <algorithm>
+#include <set>
+#include <string>
+#include <utility>
+
+#include "olap/olap_common.h"
 #include "olap/rowset/segment_v2/bloom_filter.h" // for BloomFilterOptions, BloomFilter
-#include "olap/rowset/segment_v2/common.h"
-#include "olap/rowset/segment_v2/encoding_info.h"
 #include "olap/rowset/segment_v2/indexed_column_writer.h"
 #include "olap/types.h"
-#include "util/faststring.h"
+#include "runtime/decimalv2_value.h"
 #include "util/slice.h"
+#include "util/types.h"
 
 namespace doris {
 namespace segment_v2 {
diff --git a/be/src/olap/rowset/segment_v2/bloom_filter_index_writer.h b/be/src/olap/rowset/segment_v2/bloom_filter_index_writer.h
index 2cd0b8e340..c4f20d0f6b 100644
--- a/be/src/olap/rowset/segment_v2/bloom_filter_index_writer.h
+++ b/be/src/olap/rowset/segment_v2/bloom_filter_index_writer.h
@@ -17,13 +17,16 @@
 
 #pragma once
 
+#include <butil/macros.h>
+#include <stdint.h>
+
 #include <cstddef>
 #include <memory>
+#include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/segment_v2.pb.h"
-#include "gutil/macros.h"
 #include "olap/itoken_extractor.h"
+#include "olap/rowset/segment_v2/bloom_filter.h"
 #include "vec/common/arena.h"
 
 namespace doris {
@@ -36,7 +39,7 @@ class FileWriter;
 
 namespace segment_v2 {
 
-struct BloomFilterOptions;
+class ColumnIndexMetaPB;
 
 class BloomFilterIndexWriter {
 public:
diff --git a/be/src/olap/rowset/segment_v2/column_reader.cpp b/be/src/olap/rowset/segment_v2/column_reader.cpp
index a4a437c353..c4219916c9 100644
--- a/be/src/olap/rowset/segment_v2/column_reader.cpp
+++ b/be/src/olap/rowset/segment_v2/column_reader.cpp
@@ -17,21 +17,51 @@
 
 #include "olap/rowset/segment_v2/column_reader.h"
 
+#include <assert.h>
+#include <gen_cpp/segment_v2.pb.h>
+
+#include <algorithm>
+#include <ostream>
+#include <set>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "io/fs/file_reader.h"
+#include "olap/block_column_predicate.h"
+#include "olap/column_predicate.h"
+#include "olap/decimal12.h"
+#include "olap/inverted_index_parser.h"
 #include "olap/rowset/segment_v2/binary_dict_page.h" // for BinaryDictPageDecoder
+#include "olap/rowset/segment_v2/binary_plain_page.h"
+#include "olap/rowset/segment_v2/bitmap_index_reader.h"
+#include "olap/rowset/segment_v2/bloom_filter.h"
 #include "olap/rowset/segment_v2/bloom_filter_index_reader.h"
 #include "olap/rowset/segment_v2/encoding_info.h" // for EncodingInfo
-#include "olap/rowset/segment_v2/page_handle.h"   // for PageHandle
+#include "olap/rowset/segment_v2/inverted_index_reader.h"
+#include "olap/rowset/segment_v2/page_decoder.h"
+#include "olap/rowset/segment_v2/page_handle.h" // for PageHandle
 #include "olap/rowset/segment_v2/page_io.h"
 #include "olap/rowset/segment_v2/page_pointer.h" // for PagePointer
-#include "olap/types.h"                          // for TypeInfo
+#include "olap/rowset/segment_v2/row_ranges.h"
+#include "olap/rowset/segment_v2/zone_map_index.h"
+#include "olap/tablet_schema.h"
+#include "olap/types.h" // for TypeInfo
 #include "olap/wrapper_field.h"
+#include "runtime/decimalv2_value.h"
+#include "util/binary_cast.hpp"
+#include "util/bitmap.h"
 #include "util/block_compression.h"
 #include "util/rle_encoding.h" // for RleDecoder
+#include "util/slice.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
 #include "vec/columns/column_map.h"
+#include "vec/columns/column_nullable.h"
 #include "vec/columns/column_struct.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/string_ref.h"
 #include "vec/core/types.h"
 #include "vec/runtime/vdatetime_value.h" //for VecDateTime
 
diff --git a/be/src/olap/rowset/segment_v2/column_reader.h b/be/src/olap/rowset/segment_v2/column_reader.h
index e0e7bc4a6f..19c43ed427 100644
--- a/be/src/olap/rowset/segment_v2/column_reader.h
+++ b/be/src/olap/rowset/segment_v2/column_reader.h
@@ -17,48 +17,61 @@
 
 #pragma once
 
+#include <gen_cpp/segment_v2.pb.h>
+#include <sys/types.h>
+
 #include <cstddef> // for size_t
 #include <cstdint> // for uint32_t
 #include <memory>  // for unique_ptr
+#include <mutex>
+#include <string>
+#include <utility>
+#include <vector>
 
-#include "bloom_filter_index_reader.h"
+#include "common/config.h"
 #include "common/logging.h"
-#include "common/status.h"         // for Status
-#include "gen_cpp/segment_v2.pb.h" // for ColumnMetaPB
-#include "io/fs/file_reader.h"
+#include "common/status.h" // for Status
+#include "io/fs/file_reader_writer_fwd.h"
 #include "io/io_common.h"
-#include "olap/block_column_predicate.h"
-#include "olap/column_predicate.h"
-#include "olap/iterators.h"
-#include "olap/rowset/segment_v2/bitmap_index_reader.h" // for BitmapIndexReader
+#include "olap/olap_common.h"
 #include "olap/rowset/segment_v2/common.h"
-#include "olap/rowset/segment_v2/inverted_index_reader.h" // for InvertedIndexReader
-#include "olap/rowset/segment_v2/ordinal_page_index.h"    // for OrdinalPageIndexIterator
-#include "olap/rowset/segment_v2/page_handle.h"           // for PageHandle
-#include "olap/rowset/segment_v2/parsed_page.h"           // for ParsedPage
-#include "olap/rowset/segment_v2/row_ranges.h"            // for RowRanges
-#include "olap/rowset/segment_v2/zone_map_index.h"
-#include "olap/tablet_schema.h"
+#include "olap/rowset/segment_v2/ordinal_page_index.h" // for OrdinalPageIndexIterator
+#include "olap/rowset/segment_v2/page_handle.h"        // for PageHandle
+#include "olap/rowset/segment_v2/page_pointer.h"
+#include "olap/rowset/segment_v2/parsed_page.h" // for ParsedPage
+#include "olap/types.h"
+#include "olap/utils.h"
 #include "util/once.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h" // ColumnArray
+#include "vec/data_types/data_type.h"
 
 namespace doris {
 
-class TypeInfo;
 class BlockCompressionCodec;
 class WrapperField;
+class AndBlockColumnPredicate;
+class ColumnPredicate;
+class TabletIndex;
 
-namespace fs {
-class ReadableBlock;
-}
+namespace io {
+class FileReader;
+} // namespace io
+struct Slice;
+struct StringRef;
 
 namespace segment_v2 {
 
 class EncodingInfo;
-class PageHandle;
-struct PagePointer;
 class ColumnIterator;
 class BloomFilterIndexReader;
+class BitmapIndexIterator;
+class BitmapIndexReader;
+class InvertedIndexIterator;
+class InvertedIndexReader;
+class PageDecoder;
+class RowRanges;
+class ZoneMapIndexReader;
 
 struct ColumnReaderOptions {
     // whether verify checksum when read page
diff --git a/be/src/olap/rowset/segment_v2/column_writer.cpp b/be/src/olap/rowset/segment_v2/column_writer.cpp
index ab542a49dd..0ee3608f84 100644
--- a/be/src/olap/rowset/segment_v2/column_writer.cpp
+++ b/be/src/olap/rowset/segment_v2/column_writer.cpp
@@ -17,11 +17,17 @@
 
 #include "olap/rowset/segment_v2/column_writer.h"
 
-#include <cstddef>
+#include <assert.h>
+#include <gen_cpp/segment_v2.pb.h>
 
+#include <algorithm>
+#include <filesystem>
+
+#include "common/config.h"
 #include "common/logging.h"
 #include "gutil/strings/substitute.h"
 #include "io/fs/file_writer.h"
+#include "olap/olap_common.h"
 #include "olap/rowset/segment_v2/bitmap_index_writer.h"
 #include "olap/rowset/segment_v2/bloom_filter.h"
 #include "olap/rowset/segment_v2/bloom_filter_index_writer.h"
@@ -31,10 +37,15 @@
 #include "olap/rowset/segment_v2/ordinal_page_index.h"
 #include "olap/rowset/segment_v2/page_builder.h"
 #include "olap/rowset/segment_v2/page_io.h"
+#include "olap/rowset/segment_v2/page_pointer.h"
 #include "olap/rowset/segment_v2/zone_map_index.h"
+#include "olap/tablet_schema.h"
+#include "olap/types.h"
+#include "runtime/collection_value.h"
 #include "util/block_compression.h"
 #include "util/faststring.h"
 #include "util/rle_encoding.h"
+#include "vec/core/types.h"
 
 namespace doris {
 namespace segment_v2 {
diff --git a/be/src/olap/rowset/segment_v2/column_writer.h b/be/src/olap/rowset/segment_v2/column_writer.h
index 2aed982126..1c235ccda7 100644
--- a/be/src/olap/rowset/segment_v2/column_writer.h
+++ b/be/src/olap/rowset/segment_v2/column_writer.h
@@ -17,22 +17,29 @@
 
 #pragma once
 
-#include <memory> // for unique_ptr
+#include <gen_cpp/segment_v2.pb.h>
+#include <stddef.h>
+#include <stdint.h>
 
-#include "common/status.h"         // for Status
-#include "gen_cpp/segment_v2.pb.h" // for EncodingTypePB
-#include "olap/field.h"            // for Field
-#include "olap/inverted_index_parser.h"
+#include <algorithm>
+#include <memory> // for unique_ptr
+#include <ostream>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "common/status.h" // for Status
+#include "olap/field.h"    // for Field
 #include "olap/rowset/segment_v2/common.h"
-#include "olap/rowset/segment_v2/page_pointer.h" // for PagePointer
-#include "olap/tablet_schema.h"                  // for TabletColumn
-#include "util/bitmap.h"                         // for BitmapChange
-#include "util/slice.h"                          // for OwnedSlice
+#include "olap/rowset/segment_v2/inverted_index_writer.h"
+#include "util/bitmap.h" // for BitmapChange
+#include "util/slice.h"  // for OwnedSlice
 
 namespace doris {
 
-class TypeInfo;
 class BlockCompressionCodec;
+class TabletColumn;
+class TabletIndex;
 
 namespace io {
 class FileWriter;
@@ -69,7 +76,6 @@ struct ColumnWriterOptions {
 };
 
 class BitmapIndexWriter;
-class InvertedIndexColumnWriter;
 class EncodingInfo;
 class NullBitmapBuilder;
 class OrdinalIndexWriter;
diff --git a/be/src/olap/rowset/segment_v2/empty_segment_iterator.cpp b/be/src/olap/rowset/segment_v2/empty_segment_iterator.cpp
index e9a51ed46a..95491196f9 100644
--- a/be/src/olap/rowset/segment_v2/empty_segment_iterator.cpp
+++ b/be/src/olap/rowset/segment_v2/empty_segment_iterator.cpp
@@ -18,6 +18,10 @@
 #include "olap/rowset/segment_v2/empty_segment_iterator.h"
 
 namespace doris {
+namespace vectorized {
+class Block;
+} // namespace vectorized
+
 namespace segment_v2 {
 
 EmptySegmentIterator::EmptySegmentIterator(const doris::Schema& schema) : _schema(schema) {}
diff --git a/be/src/olap/rowset/segment_v2/empty_segment_iterator.h b/be/src/olap/rowset/segment_v2/empty_segment_iterator.h
index 0bd7014c61..46d4e9bb8f 100644
--- a/be/src/olap/rowset/segment_v2/empty_segment_iterator.h
+++ b/be/src/olap/rowset/segment_v2/empty_segment_iterator.h
@@ -19,10 +19,13 @@
 
 #include "common/status.h"
 #include "olap/iterators.h"
-#include "olap/rowset/segment_v2/segment.h"
 #include "olap/schema.h"
 
 namespace doris {
+namespace vectorized {
+class Block;
+} // namespace vectorized
+
 namespace segment_v2 {
 
 class EmptySegmentIterator : public RowwiseIterator {
diff --git a/be/src/olap/rowset/segment_v2/encoding_info.cpp b/be/src/olap/rowset/segment_v2/encoding_info.cpp
index 0949e9c1b9..c500d7e902 100644
--- a/be/src/olap/rowset/segment_v2/encoding_info.cpp
+++ b/be/src/olap/rowset/segment_v2/encoding_info.cpp
@@ -17,9 +17,13 @@
 
 #include "olap/rowset/segment_v2/encoding_info.h"
 
-#include <type_traits>
+#include <gen_cpp/segment_v2.pb.h>
+
+#include <iterator>
+#include <type_traits>
+#include <unordered_map>
+#include <utility>
 
-#include "gutil/strings/substitute.h"
 #include "olap/olap_common.h"
 #include "olap/rowset/segment_v2/binary_dict_page.h"
 #include "olap/rowset/segment_v2/binary_plain_page.h"
@@ -29,6 +33,7 @@
 #include "olap/rowset/segment_v2/frame_of_reference_page.h"
 #include "olap/rowset/segment_v2/plain_page.h"
 #include "olap/rowset/segment_v2/rle_page.h"
+#include "olap/types.h"
 
 namespace doris {
 namespace segment_v2 {
diff --git a/be/src/olap/rowset/segment_v2/encoding_info.h b/be/src/olap/rowset/segment_v2/encoding_info.h
index 7783862462..4086f78294 100644
--- a/be/src/olap/rowset/segment_v2/encoding_info.h
+++ b/be/src/olap/rowset/segment_v2/encoding_info.h
@@ -17,15 +17,18 @@
 
 #pragma once
 
+#include <stddef.h>
+
 #include <functional>
+#include <memory>
 
 #include "common/status.h"
-#include "gen_cpp/segment_v2.pb.h"
-#include "olap/types.h"
+#include "util/slice.h"
 
 namespace doris {
 
 class TypeInfo;
+enum class FieldType;
 
 namespace segment_v2 {
 
@@ -33,6 +36,7 @@ class PageBuilder;
 class PageDecoder;
 struct PageBuilderOptions;
 struct PageDecoderOptions;
+enum EncodingTypePB : int;
 
 // For better performance, some encodings (like BitShuffle) need to be decoded before being added to the PageCache.
 class DataPagePreDecoder {
diff --git a/be/src/olap/rowset/segment_v2/index_page.cpp b/be/src/olap/rowset/segment_v2/index_page.cpp
index 604cab1f3d..3b8bc03977 100644
--- a/be/src/olap/rowset/segment_v2/index_page.cpp
+++ b/be/src/olap/rowset/segment_v2/index_page.cpp
@@ -17,9 +17,11 @@
 
 #include "olap/rowset/segment_v2/index_page.h"
 
-#include <string>
+#include <gen_cpp/segment_v2.pb.h>
+
+#include <algorithm>
+#include <ostream>
 
-#include "common/logging.h"
 #include "util/coding.h"
 
 namespace doris {
diff --git a/be/src/olap/rowset/segment_v2/index_page.h b/be/src/olap/rowset/segment_v2/index_page.h
index dcfcedafb1..106fc0ffc5 100644
--- a/be/src/olap/rowset/segment_v2/index_page.h
+++ b/be/src/olap/rowset/segment_v2/index_page.h
@@ -17,13 +17,15 @@
 
 #pragma once
 
+#include <butil/macros.h>
+#include <gen_cpp/segment_v2.pb.h>
+#include <glog/logging.h>
+#include <stdint.h>
+
 #include <cstddef>
-#include <memory>
 #include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/segment_v2.pb.h"
-#include "gutil/macros.h"
 #include "olap/rowset/segment_v2/page_pointer.h"
 #include "util/faststring.h"
 #include "util/slice.h"
@@ -77,7 +79,6 @@ private:
     uint32_t _count = 0;
 };
 
-class IndexPageIterator;
 class IndexPageReader {
 public:
     IndexPageReader() : _parsed(false) {}
diff --git a/be/src/olap/rowset/segment_v2/indexed_column_reader.cpp b/be/src/olap/rowset/segment_v2/indexed_column_reader.cpp
index 0dc0bffca6..8c5997f99f 100644
--- a/be/src/olap/rowset/segment_v2/indexed_column_reader.cpp
+++ b/be/src/olap/rowset/segment_v2/indexed_column_reader.cpp
@@ -17,11 +17,19 @@
 
 #include "olap/rowset/segment_v2/indexed_column_reader.h"
 
+#include <gen_cpp/segment_v2.pb.h>
+
+#include <algorithm>
+
 #include "gutil/strings/substitute.h" // for Substitute
-#include "io/fs/local_file_system.h"
 #include "olap/key_coder.h"
+#include "olap/olap_common.h"
 #include "olap/rowset/segment_v2/encoding_info.h" // for EncodingInfo
+#include "olap/rowset/segment_v2/options.h"
+#include "olap/rowset/segment_v2/page_decoder.h"
 #include "olap/rowset/segment_v2/page_io.h"
+#include "olap/types.h"
+#include "util/block_compression.h"
 
 namespace doris {
 using namespace ErrorCode;
diff --git a/be/src/olap/rowset/segment_v2/indexed_column_reader.h b/be/src/olap/rowset/segment_v2/indexed_column_reader.h
index 10b87050ae..3ee39e04c0 100644
--- a/be/src/olap/rowset/segment_v2/indexed_column_reader.h
+++ b/be/src/olap/rowset/segment_v2/indexed_column_reader.h
@@ -17,30 +17,33 @@
 
 #pragma once
 
-#include <memory>
+#include <gen_cpp/segment_v2.pb.h>
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
+
 #include <string>
+#include <utility>
 
 #include "common/status.h"
-#include "gen_cpp/segment_v2.pb.h"
-#include "io/fs/file_reader.h"
-#include "io/fs/file_system.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "olap/rowset/segment_v2/common.h"
 #include "olap/rowset/segment_v2/index_page.h"
 #include "olap/rowset/segment_v2/page_handle.h"
 #include "olap/rowset/segment_v2/page_pointer.h"
 #include "olap/rowset/segment_v2/parsed_page.h"
-#include "util/block_compression.h"
 #include "util/slice.h"
+#include "vec/data_types/data_type.h"
 
 namespace doris {
 
 class KeyCoder;
 class TypeInfo;
+class BlockCompressionCodec;
 
 namespace segment_v2 {
 
 class EncodingInfo;
-class IndexedColumnIterator;
 
 // thread-safe reader for IndexedColumn (see comments of `IndexedColumnWriter` to understand what IndexedColumn is)
 class IndexedColumnReader {
diff --git a/be/src/olap/rowset/segment_v2/indexed_column_writer.cpp b/be/src/olap/rowset/segment_v2/indexed_column_writer.cpp
index bd912c6eb8..0afbcc349c 100644
--- a/be/src/olap/rowset/segment_v2/indexed_column_writer.cpp
+++ b/be/src/olap/rowset/segment_v2/indexed_column_writer.cpp
@@ -17,10 +17,14 @@
 
 #include "olap/rowset/segment_v2/indexed_column_writer.h"
 
+#include <gen_cpp/segment_v2.pb.h>
+
+#include <ostream>
 #include <string>
 
 #include "common/logging.h"
 #include "olap/key_coder.h"
+#include "olap/olap_common.h"
 #include "olap/rowset/segment_v2/encoding_info.h"
 #include "olap/rowset/segment_v2/index_page.h"
 #include "olap/rowset/segment_v2/options.h"
@@ -29,7 +33,7 @@
 #include "olap/rowset/segment_v2/page_pointer.h"
 #include "olap/types.h"
 #include "util/block_compression.h"
-#include "util/coding.h"
+#include "util/slice.h"
 
 namespace doris {
 namespace segment_v2 {
diff --git a/be/src/olap/rowset/segment_v2/indexed_column_writer.h b/be/src/olap/rowset/segment_v2/indexed_column_writer.h
index e4aadb7c3f..ab6812b5e5 100644
--- a/be/src/olap/rowset/segment_v2/indexed_column_writer.h
+++ b/be/src/olap/rowset/segment_v2/indexed_column_writer.h
@@ -17,16 +17,17 @@
 
 #pragma once
 
+#include <butil/macros.h>
+#include <gen_cpp/segment_v2.pb.h>
+#include <stdint.h>
+
 #include <cstddef>
 #include <memory>
-#include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/segment_v2.pb.h"
-#include "gutil/macros.h"
 #include "olap/rowset/segment_v2/common.h"
 #include "olap/rowset/segment_v2/page_pointer.h"
-#include "util/slice.h"
+#include "util/faststring.h"
 #include "vec/common/arena.h"
 
 namespace doris {
diff --git a/be/src/olap/rowset/segment_v2/inverted_index_cache.cpp b/be/src/olap/rowset/segment_v2/inverted_index_cache.cpp
index 00ae62c378..1c6a61003b 100644
--- a/be/src/olap/rowset/segment_v2/inverted_index_cache.cpp
+++ b/be/src/olap/rowset/segment_v2/inverted_index_cache.cpp
@@ -17,9 +17,23 @@
 
 #include "olap/rowset/segment_v2/inverted_index_cache.h"
 
+#include <CLucene/debug/mem.h>
+#include <CLucene/search/IndexSearcher.h>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
+#include <string.h>
+#include <sys/resource.h>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <iostream>
+
+#include "common/logging.h"
+#include "olap/olap_common.h"
 #include "olap/rowset/segment_v2/inverted_index_compound_directory.h"
 #include "olap/rowset/segment_v2/inverted_index_compound_reader.h"
+#include "runtime/thread_context.h"
 #include "util/defer_op.h"
+#include "util/runtime_profile.h"
 
 namespace doris {
 namespace segment_v2 {
diff --git a/be/src/olap/rowset/segment_v2/inverted_index_cache.h b/be/src/olap/rowset/segment_v2/inverted_index_cache.h
index 77d428aa42..3c225bb9d9 100644
--- a/be/src/olap/rowset/segment_v2/inverted_index_cache.h
+++ b/be/src/olap/rowset/segment_v2/inverted_index_cache.h
@@ -17,21 +17,36 @@
 
 #pragma once
 
-#include <CLucene.h>
+#include <CLucene.h> // IWYU pragma: keep
+#include <CLucene/util/Misc.h>
+#include <butil/macros.h>
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
 
-#include <iostream>
-#include <map>
+#include <atomic>
 #include <memory>
-#include <mutex>
 #include <roaring/roaring.hh>
-#include <vector>
+#include <string>
+#include <utility>
 
+#include "common/config.h"
+#include "common/status.h"
 #include "io/fs/file_system.h"
+#include "io/fs/path.h"
 #include "olap/lru_cache.h"
 #include "runtime/memory/mem_tracker.h"
+#include "util/slice.h"
 #include "util/time.h"
 
+namespace lucene {
+namespace search {
+class IndexSearcher;
+} // namespace search
+} // namespace lucene
+
 namespace doris {
+struct OlapReaderStatistics;
 
 namespace segment_v2 {
 using IndexSearcherPtr = std::shared_ptr<lucene::search::IndexSearcher>;
@@ -165,7 +180,6 @@ private:
 };
 
 enum class InvertedIndexQueryType;
-
 class InvertedIndexQueryCacheHandle;
 
 class InvertedIndexQueryCache {
diff --git a/be/src/olap/rowset/segment_v2/inverted_index_compound_directory.cpp b/be/src/olap/rowset/segment_v2/inverted_index_compound_directory.cpp
index c3b474dfe1..969623c780 100644
--- a/be/src/olap/rowset/segment_v2/inverted_index_compound_directory.cpp
+++ b/be/src/olap/rowset/segment_v2/inverted_index_compound_directory.cpp
@@ -18,10 +18,11 @@
 #include "olap/rowset/segment_v2/inverted_index_compound_directory.h"
 
 #include "CLucene/SharedHeader.h"
-#include "CLucene/StdHeader.h"
+#include "common/status.h"
+#include "io/fs/file_reader.h"
 #include "io/fs/file_writer.h"
-#include "olap/iterators.h"
-#include "util/md5.h"
+#include "io/fs/path.h"
+#include "util/slice.h"
 
 #ifdef _CL_HAVE_IO_H
 #include <io.h>
@@ -35,12 +36,29 @@
 #ifdef _CL_HAVE_DIRECT_H
 #include <direct.h>
 #endif
+#include <CLucene/LuceneThreads.h>
+#include <CLucene/clucene-config.h>
+#include <CLucene/debug/error.h>
+#include <CLucene/debug/mem.h>
 #include <CLucene/index/IndexReader.h>
 #include <CLucene/index/IndexWriter.h>
 #include <CLucene/store/LockFactory.h>
+#include <CLucene/store/RAMDirectory.h>
 #include <CLucene/util/Misc.h>
 #include <assert.h>
-#include <errno.h>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
+#include <glog/logging.h>
+#include <stdio.h>
+#include <string.h>
+#include <wchar.h>
+
+#include <algorithm>
+#include <filesystem>
+#include <iostream>
+#include <memory>
+#include <mutex>
+#include <utility>
 
 #define CL_MAX_PATH 4096
 #define CL_MAX_DIR CL_MAX_PATH
diff --git a/be/src/olap/rowset/segment_v2/inverted_index_compound_directory.h b/be/src/olap/rowset/segment_v2/inverted_index_compound_directory.h
index b87498a34e..b06f980f6b 100644
--- a/be/src/olap/rowset/segment_v2/inverted_index_compound_directory.h
+++ b/be/src/olap/rowset/segment_v2/inverted_index_compound_directory.h
@@ -17,17 +17,28 @@
 
 #pragma once
 
-#include <CLucene.h>
+#include <CLucene.h> // IWYU pragma: keep
+#include <CLucene/store/Directory.h>
+#include <CLucene/store/IndexInput.h>
+#include <CLucene/store/IndexOutput.h>
+#include <stdint.h>
 
-#include <iostream>
-#include <map>
-#include <memory>
-#include <mutex>
+#include <string>
 #include <vector>
 
+#include "CLucene/SharedHeader.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "io/fs/file_system.h"
 #include "util/lock.h"
 
+class CLuceneError;
+
+namespace lucene {
+namespace store {
+class LockFactory;
+} // namespace store
+} // namespace lucene
+
 namespace doris {
 
 namespace segment_v2 {
@@ -76,6 +87,7 @@ protected:
 public:
     class FSIndexOutput;
     class FSIndexInput;
+
     friend class DorisCompoundDirectory::FSIndexOutput;
     friend class DorisCompoundDirectory::FSIndexInput;
 
diff --git a/be/src/olap/rowset/segment_v2/inverted_index_compound_reader.cpp b/be/src/olap/rowset/segment_v2/inverted_index_compound_reader.cpp
index 9ce20703b4..15646679cf 100644
--- a/be/src/olap/rowset/segment_v2/inverted_index_compound_reader.cpp
+++ b/be/src/olap/rowset/segment_v2/inverted_index_compound_reader.cpp
@@ -17,8 +17,28 @@
 
 #include "olap/rowset/segment_v2/inverted_index_compound_reader.h"
 
+#include <CLucene/clucene-config.h>
+#include <CLucene/debug/error.h>
+#include <CLucene/debug/mem.h>
+#include <CLucene/store/RAMDirectory.h>
+#include <CLucene/util/Misc.h>
+#include <stdio.h>
+#include <string.h>
+#include <wchar.h>
+
+#include <algorithm>
+#include <memory>
+#include <utility>
+
+#include "CLucene/SharedHeader.h"
 #include "olap/rowset/segment_v2/inverted_index_compound_directory.h"
 
+namespace doris {
+namespace io {
+class FileWriter;
+} // namespace io
+} // namespace doris
+
 #define BUFFER_LENGTH 16384
 #define CL_MAX_PATH 4096
 
diff --git a/be/src/olap/rowset/segment_v2/inverted_index_compound_reader.h b/be/src/olap/rowset/segment_v2/inverted_index_compound_reader.h
index d6e95d9d50..5c0d963247 100644
--- a/be/src/olap/rowset/segment_v2/inverted_index_compound_reader.h
+++ b/be/src/olap/rowset/segment_v2/inverted_index_compound_reader.h
@@ -17,14 +17,28 @@
 
 #pragma once
 
-#include <CLucene.h>
+#include <CLucene.h> // IWYU pragma: keep
+#include <CLucene/SharedHeader.h>
+#include <CLucene/store/Directory.h>
+#include <CLucene/store/IndexInput.h>
+#include <CLucene/store/IndexOutput.h>
+#include <CLucene/util/Equators.h>
+#include <CLucene/util/VoidMap.h>
+#include <stdint.h>
 
-#include <iostream>
 #include <map>
-#include <memory>
 #include <mutex>
+#include <string>
 #include <vector>
 
+class CLuceneError;
+
+namespace lucene {
+namespace store {
+class RAMDirectory;
+} // namespace store
+} // namespace lucene
+
 namespace doris {
 
 namespace segment_v2 {
@@ -32,6 +46,7 @@ namespace segment_v2 {
 class CLUCENE_EXPORT DorisCompoundReader : public lucene::store::Directory {
 private:
     class ReaderFileEntry;
+
     friend class DorisCompoundReader::ReaderFileEntry;
 
     int32_t readBufferSize;
diff --git a/be/src/olap/rowset/segment_v2/inverted_index_desc.cpp b/be/src/olap/rowset/segment_v2/inverted_index_desc.cpp
index 170409cac4..4bdf44ead0 100644
--- a/be/src/olap/rowset/segment_v2/inverted_index_desc.cpp
+++ b/be/src/olap/rowset/segment_v2/inverted_index_desc.cpp
@@ -17,7 +17,10 @@
 
 #include "olap/rowset/segment_v2/inverted_index_desc.h"
 
+#include <fmt/format.h>
+
 #include "gutil/strings/strip.h"
+#include "olap/olap_common.h"
 
 namespace doris::segment_v2 {
 const std::string segment_suffix = ".dat";
diff --git a/be/src/olap/rowset/segment_v2/inverted_index_desc.h b/be/src/olap/rowset/segment_v2/inverted_index_desc.h
index 0da5dfddbb..1cf4636d80 100644
--- a/be/src/olap/rowset/segment_v2/inverted_index_desc.h
+++ b/be/src/olap/rowset/segment_v2/inverted_index_desc.h
@@ -17,12 +17,13 @@
 
 #pragma once
 
-#include <memory>
+#include <stdint.h>
+
 #include <string>
 
-#include "olap/olap_common.h"
-
 namespace doris {
+struct RowsetId;
+
 namespace segment_v2 {
 
 class InvertedIndexDescriptor {
diff --git a/be/src/olap/rowset/segment_v2/inverted_index_reader.cpp b/be/src/olap/rowset/segment_v2/inverted_index_reader.cpp
index 9a7520dc57..6e94640255 100644
--- a/be/src/olap/rowset/segment_v2/inverted_index_reader.cpp
+++ b/be/src/olap/rowset/segment_v2/inverted_index_reader.cpp
@@ -17,22 +17,44 @@
 
 #include "olap/rowset/segment_v2/inverted_index_reader.h"
 
+#include <CLucene/analysis/AnalysisHeader.h>
+#include <CLucene/analysis/Analyzers.h>
 #include <CLucene/analysis/LanguageBasedAnalyzer.h>
-#include <CLucene/search/BooleanQuery.h>
-#include <CLucene/search/PhraseQuery.h>
+#include <CLucene/analysis/standard/StandardAnalyzer.h>
+#include <CLucene/clucene-config.h>
+#include <CLucene/debug/error.h>
+#include <CLucene/debug/mem.h>
+#include <CLucene/index/IndexReader.h>
+#include <CLucene/index/Term.h>
+#include <CLucene/search/IndexSearcher.h>
+#include <CLucene/search/Query.h>
+#include <CLucene/search/RangeQuery.h>
+#include <CLucene/search/TermQuery.h>
+#include <CLucene/store/Directory.h>
+#include <CLucene/store/IndexInput.h>
+#include <CLucene/util/CLStreams.h>
 #include <CLucene/util/FutureArrays.h>
-#include <CLucene/util/NumericUtils.h>
+#include <CLucene/util/bkd/bkd_docid_iterator.h>
+#include <math.h>
+#include <string.h>
+
+#include <CLucene/util/croaring/roaring.hh>
+#include <algorithm>
+#include <filesystem>
+#include <ostream>
+#include <set>
 
 #include "common/config.h"
-#include "gutil/strings/strip.h"
+#include "common/logging.h"
 #include "io/fs/file_system.h"
 #include "olap/key_coder.h"
+#include "olap/olap_common.h"
 #include "olap/rowset/segment_v2/inverted_index_cache.h"
 #include "olap/rowset/segment_v2/inverted_index_compound_directory.h"
 #include "olap/rowset/segment_v2/inverted_index_desc.h"
-#include "olap/tablet_schema.h"
-#include "olap/utils.h"
+#include "olap/types.h"
 #include "util/faststring.h"
+#include "util/runtime_profile.h"
 #include "util/time.h"
 #include "vec/common/string_ref.h"
 
diff --git a/be/src/olap/rowset/segment_v2/inverted_index_reader.h b/be/src/olap/rowset/segment_v2/inverted_index_reader.h
index 731b6ca6b2..9438d5a07d 100644
--- a/be/src/olap/rowset/segment_v2/inverted_index_reader.h
+++ b/be/src/olap/rowset/segment_v2/inverted_index_reader.h
@@ -17,30 +17,44 @@
 
 #pragma once
 
-#include <CLucene.h>
-#include <CLucene/util/BitSet.h>
 #include <CLucene/util/bkd/bkd_reader.h>
+#include <stdint.h>
 
-#include <roaring/roaring.hh>
+#include <memory>
+#include <string>
+#include <utility>
+#include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/segment_v2.pb.h"
-#include "gutil/macros.h"
 #include "io/fs/file_system.h"
+#include "io/fs/path.h"
 #include "olap/inverted_index_parser.h"
-#include "olap/olap_common.h"
-#include "olap/rowset/segment_v2/common.h"
-#include "olap/rowset/segment_v2/inverted_index_cache.h"
 #include "olap/rowset/segment_v2/inverted_index_compound_reader.h"
 #include "olap/tablet_schema.h"
 
+namespace lucene {
+namespace store {
+class Directory;
+} // namespace store
+namespace util {
+namespace bkd {
+class bkd_docid_set_iterator;
+} // namespace bkd
+} // namespace util
+} // namespace lucene
+namespace roaring {
+class Roaring;
+} // namespace roaring
+
 namespace doris {
 class KeyCoder;
 class TypeInfo;
+struct OlapReaderStatistics;
 
 namespace segment_v2 {
 
 class InvertedIndexIterator;
+class InvertedIndexQueryCacheHandle;
 
 enum class InvertedIndexReaderType {
     UNKNOWN = -1,
diff --git a/be/src/olap/rowset/segment_v2/inverted_index_writer.cpp b/be/src/olap/rowset/segment_v2/inverted_index_writer.cpp
index ae95408648..7cfc1807c0 100644
--- a/be/src/olap/rowset/segment_v2/inverted_index_writer.cpp
+++ b/be/src/olap/rowset/segment_v2/inverted_index_writer.cpp
@@ -17,20 +17,33 @@
 
 #include "olap/rowset/segment_v2/inverted_index_writer.h"
 
-#include <CLucene.h>
+#include <CLucene.h> // IWYU pragma: keep
 #include <CLucene/analysis/LanguageBasedAnalyzer.h>
 #include <CLucene/util/bkd/bkd_writer.h>
+#include <glog/logging.h>
 
+#include <algorithm>
+#include <cstdint>
+#include <limits>
 #include <memory>
+#include <ostream>
+#include <roaring/roaring.hh>
+#include <vector>
 
+#include "common/config.h"
 #include "olap/field.h"
+#include "olap/inverted_index_parser.h"
+#include "olap/key_coder.h"
 #include "olap/olap_common.h"
 #include "olap/rowset/segment_v2/common.h"
 #include "olap/rowset/segment_v2/inverted_index_cache.h"
 #include "olap/rowset/segment_v2/inverted_index_compound_directory.h"
 #include "olap/rowset/segment_v2/inverted_index_desc.h"
 #include "olap/tablet_schema.h"
+#include "olap/types.h"
+#include "runtime/collection_value.h"
 #include "util/faststring.h"
+#include "util/slice.h"
 #include "util/string_util.h"
 
 #define FINALIZE_OUTPUT(x) \
@@ -531,4 +544,4 @@ Status InvertedIndexColumnWriter::create(const Field* field,
     }
     return Status::OK();
 }
-} // namespace doris::segment_v2
\ No newline at end of file
+} // namespace doris::segment_v2
diff --git a/be/src/olap/rowset/segment_v2/inverted_index_writer.h b/be/src/olap/rowset/segment_v2/inverted_index_writer.h
index 16678e6c93..a92ec6220b 100644
--- a/be/src/olap/rowset/segment_v2/inverted_index_writer.h
+++ b/be/src/olap/rowset/segment_v2/inverted_index_writer.h
@@ -17,14 +17,20 @@
 
 #pragma once
 
+#include <butil/macros.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <memory>
+#include <string>
+
 #include "common/status.h"
 #include "io/fs/file_system.h"
-#include "olap/inverted_index_parser.h"
-#include "olap/olap_common.h"
-#include "olap/tablet_schema.h"
 
 namespace doris {
 class CollectionValue;
+class Field;
+class TabletIndex;
 
 namespace segment_v2 {
 
diff --git a/be/src/olap/rowset/segment_v2/ngram_bloom_filter.cpp b/be/src/olap/rowset/segment_v2/ngram_bloom_filter.cpp
index a761c5a6e4..54f49b8bb6 100644
--- a/be/src/olap/rowset/segment_v2/ngram_bloom_filter.cpp
+++ b/be/src/olap/rowset/segment_v2/ngram_bloom_filter.cpp
@@ -17,8 +17,11 @@
 
 #include "olap/rowset/segment_v2/ngram_bloom_filter.h"
 
+#include <gen_cpp/segment_v2.pb.h>
+#include <glog/logging.h>
+
+#include "gutil/strings/substitute.h"
 #include "util/cityhash102/city.h"
-#include "util/debug_util.h"
 
 namespace doris {
 namespace segment_v2 {
diff --git a/be/src/olap/rowset/segment_v2/ngram_bloom_filter.h b/be/src/olap/rowset/segment_v2/ngram_bloom_filter.h
index c7684a796f..b69e467831 100644
--- a/be/src/olap/rowset/segment_v2/ngram_bloom_filter.h
+++ b/be/src/olap/rowset/segment_v2/ngram_bloom_filter.h
@@ -17,10 +17,17 @@
 
 #pragma once
 
+#include <stddef.h>
+#include <stdint.h>
+
+#include <vector>
+
+#include "common/status.h"
 #include "olap/rowset/segment_v2/bloom_filter.h"
 
 namespace doris {
 namespace segment_v2 {
+enum HashStrategyPB : int;
 
 class NGramBloomFilter : public BloomFilter {
 public:
diff --git a/be/src/olap/rowset/segment_v2/ordinal_page_index.cpp b/be/src/olap/rowset/segment_v2/ordinal_page_index.cpp
index 5a077ae80a..d977e29f11 100644
--- a/be/src/olap/rowset/segment_v2/ordinal_page_index.cpp
+++ b/be/src/olap/rowset/segment_v2/ordinal_page_index.cpp
@@ -17,12 +17,19 @@
 
 #include "olap/rowset/segment_v2/ordinal_page_index.h"
 
-#include "common/logging.h"
+#include <gen_cpp/segment_v2.pb.h>
+
+#include <algorithm>
+#include <filesystem>
+#include <ostream>
+#include <string>
+
 #include "io/fs/file_writer.h"
-#include "io/fs/local_file_system.h"
 #include "olap/key_coder.h"
+#include "olap/olap_common.h"
 #include "olap/rowset/segment_v2/page_handle.h"
 #include "olap/rowset/segment_v2/page_io.h"
+#include "util/slice.h"
 
 namespace doris {
 namespace segment_v2 {
diff --git a/be/src/olap/rowset/segment_v2/ordinal_page_index.h b/be/src/olap/rowset/segment_v2/ordinal_page_index.h
index d399b8db23..17dba9e675 100644
--- a/be/src/olap/rowset/segment_v2/ordinal_page_index.h
+++ b/be/src/olap/rowset/segment_v2/ordinal_page_index.h
@@ -17,18 +17,19 @@
 
 #pragma once
 
+#include <butil/macros.h>
+#include <glog/logging.h>
+
 #include <cstdint>
 #include <memory>
-#include <string>
+#include <utility>
+#include <vector>
 
 #include "common/status.h"
-#include "gutil/macros.h"
-#include "io/fs/file_reader.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "olap/rowset/segment_v2/common.h"
 #include "olap/rowset/segment_v2/index_page.h"
 #include "olap/rowset/segment_v2/page_pointer.h"
-#include "util/coding.h"
-#include "util/slice.h"
 
 namespace doris {
 
@@ -37,6 +38,8 @@ class FileWriter;
 }
 
 namespace segment_v2 {
+class ColumnIndexMetaPB;
+class OrdinalIndexPB;
 
 // Ordinal index is implemented by one IndexPage that stores the first value ordinal
 // and file pointer for each data page.
diff --git a/be/src/olap/rowset/segment_v2/page_io.cpp b/be/src/olap/rowset/segment_v2/page_io.cpp
index 7eb3dd0549..ee8b761053 100644
--- a/be/src/olap/rowset/segment_v2/page_io.cpp
+++ b/be/src/olap/rowset/segment_v2/page_io.cpp
@@ -17,13 +17,24 @@
 
 #include "olap/rowset/segment_v2/page_io.h"
 
+#include <gen_cpp/segment_v2.pb.h>
+#include <stdint.h>
+
+#include <algorithm>
 #include <cstring>
+#include <memory>
+#include <ostream>
 #include <string>
+#include <utility>
 
 #include "common/logging.h"
 #include "gutil/strings/substitute.h"
+#include "io/fs/file_reader.h"
 #include "io/fs/file_writer.h"
+#include "olap/olap_common.h"
 #include "olap/page_cache.h"
+#include "olap/rowset/segment_v2/encoding_info.h"
+#include "olap/rowset/segment_v2/page_handle.h"
 #include "util/block_compression.h"
 #include "util/coding.h"
 #include "util/crc32c.h"
diff --git a/be/src/olap/rowset/segment_v2/page_io.h b/be/src/olap/rowset/segment_v2/page_io.h
index 06e345c27a..5ccccc2a3d 100644
--- a/be/src/olap/rowset/segment_v2/page_io.h
+++ b/be/src/olap/rowset/segment_v2/page_io.h
@@ -17,15 +17,13 @@
 
 #pragma once
 
+#include <gen_cpp/segment_v2.pb.h>
+
 #include <vector>
 
 #include "common/logging.h"
 #include "common/status.h"
-#include "gen_cpp/segment_v2.pb.h"
-#include "io/fs/file_reader.h"
-#include "olap/iterators.h"
-#include "olap/rowset/segment_v2/encoding_info.h"
-#include "olap/rowset/segment_v2/page_handle.h"
+#include "io/io_common.h"
 #include "olap/rowset/segment_v2/page_pointer.h"
 #include "util/slice.h"
 
@@ -34,16 +32,14 @@ namespace doris {
 class BlockCompressionCodec;
 struct OlapReaderStatistics;
 
-namespace fs {
-class ReadableBlock;
-class WritableBlock;
-} // namespace fs
-
 namespace io {
 class FileWriter;
+class FileReader;
 } // namespace io
 
 namespace segment_v2 {
+class EncodingInfo;
+class PageHandle;
 
 struct PageReadOptions {
     // block to read page
diff --git a/be/src/olap/rowset/segment_v2/page_pointer.h b/be/src/olap/rowset/segment_v2/page_pointer.h
index adddbbebaf..4693523986 100644
--- a/be/src/olap/rowset/segment_v2/page_pointer.h
+++ b/be/src/olap/rowset/segment_v2/page_pointer.h
@@ -17,10 +17,11 @@
 
 #pragma once
 
+#include <gen_cpp/segment_v2.pb.h>
+
 #include <cstdint>
 #include <string>
 
-#include "gen_cpp/segment_v2.pb.h"
 #include "util/coding.h"
 #include "util/faststring.h"
 
diff --git a/be/src/olap/rowset/segment_v2/parsed_page.h b/be/src/olap/rowset/segment_v2/parsed_page.h
index fe21f11dc3..159b0cca72 100644
--- a/be/src/olap/rowset/segment_v2/parsed_page.h
+++ b/be/src/olap/rowset/segment_v2/parsed_page.h
@@ -17,10 +17,11 @@
 
 #pragma once
 
+#include <gen_cpp/segment_v2.pb.h>
+
 #include <memory>
 
 #include "common/status.h"
-#include "gen_cpp/segment_v2.pb.h"
 #include "olap/rowset/segment_v2/binary_dict_page.h"
 #include "olap/rowset/segment_v2/common.h"
 #include "olap/rowset/segment_v2/encoding_info.h"
diff --git a/be/src/olap/rowset/segment_v2/segment.cpp b/be/src/olap/rowset/segment_v2/segment.cpp
index 4a86e04a0f..cf952c2856 100644
--- a/be/src/olap/rowset/segment_v2/segment.cpp
+++ b/be/src/olap/rowset/segment_v2/segment.cpp
@@ -17,30 +17,53 @@
 
 #include "olap/rowset/segment_v2/segment.h"
 
+#include <gen_cpp/PlanNodes_types.h>
 #include <gen_cpp/olap_file.pb.h>
+#include <gen_cpp/segment_v2.pb.h>
+#include <string.h>
 
 #include <memory>
 #include <utility>
 
-#include "common/config.h"
-#include "common/logging.h" // LOG
-#include "io/cache/file_cache_manager.h"
-#include "io/fs/file_reader_options.h"
+#include "io/fs/file_reader.h"
 #include "io/fs/file_system.h"
+#include "io/io_common.h"
+#include "olap/block_column_predicate.h"
+#include "olap/column_predicate.h"
 #include "olap/iterators.h"
+#include "olap/primary_key_index.h"
 #include "olap/rowset/segment_v2/empty_segment_iterator.h"
+#include "olap/rowset/segment_v2/indexed_column_reader.h"
 #include "olap/rowset/segment_v2/page_io.h"
+#include "olap/rowset/segment_v2/page_pointer.h"
 #include "olap/rowset/segment_v2/segment_iterator.h"
 #include "olap/rowset/segment_v2/segment_writer.h" // k_segment_magic_length
+#include "olap/short_key_index.h"
 #include "olap/storage_engine.h"
 #include "olap/tablet_schema.h"
+#include "olap/types.h"
+#include "olap/utils.h"
+#include "runtime/memory/mem_tracker.h"
+#include "runtime/query_fragments_ctx.h"
+#include "runtime/runtime_predicate.h"
+#include "runtime/runtime_state.h"
+#include "util/coding.h"
 #include "util/crc32c.h"
 #include "util/slice.h" // Slice
+#include "vec/columns/column.h"
+#include "vec/common/string_ref.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_factory.hpp"
 #include "vec/olap/vgeneric_iterators.h"
 
 namespace doris {
+namespace io {
+class FileCacheManager;
+class FileReaderOptions;
+} // namespace io
+
 namespace segment_v2 {
+class InvertedIndexIterator;
 
 using io::FileCacheManager;
 
diff --git a/be/src/olap/rowset/segment_v2/segment.h b/be/src/olap/rowset/segment_v2/segment.h
index c8414566fc..bfb09eb1fd 100644
--- a/be/src/olap/rowset/segment_v2/segment.h
+++ b/be/src/olap/rowset/segment_v2/segment.h
@@ -17,38 +17,47 @@
 
 #pragma once
 
+#include <butil/macros.h>
+#include <gen_cpp/olap_file.pb.h>
+#include <gen_cpp/segment_v2.pb.h>
+#include <glog/logging.h>
+
 #include <cstdint>
+#include <map>
 #include <memory> // for unique_ptr
 #include <string>
-#include <vector>
+#include <unordered_map>
 
 #include "common/status.h" // Status
-#include "gen_cpp/segment_v2.pb.h"
-#include "gutil/macros.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "io/fs/file_system.h"
-#include "olap/iterators.h"
-#include "olap/primary_key_index.h"
+#include "olap/olap_common.h"
 #include "olap/rowset/segment_v2/column_reader.h" // ColumnReader
 #include "olap/rowset/segment_v2/page_handle.h"
-#include "olap/short_key_index.h"
 #include "olap/tablet_schema.h"
-#include "util/faststring.h"
 #include "util/once.h"
+#include "util/slice.h"
 
 namespace doris {
 
-class TabletSchema;
 class ShortKeyIndexDecoder;
 class Schema;
 class StorageReadOptions;
+class MemTracker;
+class PrimaryKeyIndexReader;
+class RowwiseIterator;
+
+namespace io {
+class FileReaderOptions;
+} // namespace io
+struct RowLocation;
 
 namespace segment_v2 {
 
 class BitmapIndexIterator;
-class ColumnReader;
-class ColumnIterator;
 class Segment;
-class SegmentIterator;
+class InvertedIndexIterator;
+
 using SegmentSharedPtr = std::shared_ptr<Segment>;
 // A Segment is used to represent a segment in memory format. When segment is
 // generated, it won't be modified, so this struct aimed to help read operation.
diff --git a/be/src/olap/rowset/segment_v2/segment_iterator.cpp b/be/src/olap/rowset/segment_v2/segment_iterator.cpp
index 5d3c375f54..a1c55855bf 100644
--- a/be/src/olap/rowset/segment_v2/segment_iterator.cpp
+++ b/be/src/olap/rowset/segment_v2/segment_iterator.cpp
@@ -17,25 +17,63 @@
 
 #include "olap/rowset/segment_v2/segment_iterator.h"
 
+#include <assert.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/olap_file.pb.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
 #include <memory>
+#include <numeric>
 #include <set>
 #include <utility>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/config.h"
 #include "common/consts.h"
+#include "common/logging.h"
+#include "common/object_pool.h"
 #include "common/status.h"
+#include "io/io_common.h"
 #include "olap/column_predicate.h"
 #include "olap/like_column_predicate.h"
 #include "olap/olap_common.h"
+#include "olap/primary_key_index.h"
+#include "olap/rowset/segment_v2/bitmap_index_reader.h"
 #include "olap/rowset/segment_v2/column_reader.h"
+#include "olap/rowset/segment_v2/indexed_column_reader.h"
+#include "olap/rowset/segment_v2/inverted_index_reader.h"
+#include "olap/rowset/segment_v2/row_ranges.h"
 #include "olap/rowset/segment_v2/segment.h"
 #include "olap/short_key_index.h"
+#include "olap/tablet_schema.h"
+#include "olap/types.h"
+#include "olap/utils.h"
+#include "runtime/query_fragments_ctx.h"
+#include "runtime/runtime_predicate.h"
+#include "runtime/runtime_state.h"
+#include "runtime/thread_context.h"
+#include "util/defer_op.h"
 #include "util/doris_metrics.h"
 #include "util/key_util.h"
 #include "util/simd/bits.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/string_ref.h"
+#include "vec/common/typeid_cast.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_factory.hpp"
 #include "vec/data_types/data_type_number.h"
+#include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
 #include "vec/exprs/vliteral.h"
 #include "vec/exprs/vslot_ref.h"
 
diff --git a/be/src/olap/rowset/segment_v2/segment_iterator.h b/be/src/olap/rowset/segment_v2/segment_iterator.h
index 9650ad9e27..5e3485eee0 100644
--- a/be/src/olap/rowset/segment_v2/segment_iterator.h
+++ b/be/src/olap/rowset/segment_v2/segment_iterator.h
@@ -17,36 +17,54 @@
 
 #pragma once
 
+#include <gen_cpp/Exprs_types.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <map>
 #include <memory>
+#include <ostream>
 #include <roaring/roaring.hh>
+#include <set>
+#include <string>
+#include <unordered_map>
+#include <utility>
 #include <vector>
 
 #include "common/status.h"
-#include "io/fs/file_reader.h"
-#include "io/fs/file_system.h"
+#include "io/fs/file_reader_writer_fwd.h"
+#include "olap/block_column_predicate.h"
+#include "olap/column_predicate.h"
+#include "olap/field.h"
+#include "olap/iterators.h"
 #include "olap/olap_common.h"
 #include "olap/row_cursor.h"
+#include "olap/row_cursor_cell.h"
 #include "olap/rowset/segment_v2/common.h"
-#include "olap/rowset/segment_v2/inverted_index_reader.h"
-#include "olap/rowset/segment_v2/row_ranges.h"
 #include "olap/rowset/segment_v2/segment.h"
 #include "olap/schema.h"
-#include "vec/exprs/vexpr.h"
+#include "util/runtime_profile.h"
+#include "util/slice.h"
+#include "vec/columns/column.h"
+#include "vec/core/block.h"
+#include "vec/data_types/data_type.h"
 
 namespace doris {
 
-class RowCursor;
-class ShortKeyIndexIterator;
+class ObjectPool;
 
-namespace fs {
-class ReadableBlock;
-}
+namespace vectorized {
+class VExpr;
+class VExprContext;
+} // namespace vectorized
+struct RowLocation;
 
 namespace segment_v2 {
 
 class BitmapIndexIterator;
-class BitmapIndexReader;
 class ColumnIterator;
+class InvertedIndexIterator;
+class RowRanges;
 
 struct ColumnPredicateInfo {
     ColumnPredicateInfo() = default;
diff --git a/be/src/olap/rowset/segment_v2/segment_writer.cpp b/be/src/olap/rowset/segment_v2/segment_writer.cpp
index 3363f71cde..9f772e63e3 100644
--- a/be/src/olap/rowset/segment_v2/segment_writer.cpp
+++ b/be/src/olap/rowset/segment_v2/segment_writer.cpp
@@ -17,24 +17,40 @@
 
 #include "olap/rowset/segment_v2/segment_writer.h"
 
+#include <assert.h>
+#include <gen_cpp/segment_v2.pb.h>
+#include <parallel_hashmap/phmap.h>
+
+#include <algorithm>
+#include <ostream>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/config.h"
-#include "common/consts.h"
 #include "common/logging.h" // LOG
+#include "gutil/port.h"
 #include "io/fs/file_writer.h"
 #include "olap/data_dir.h"
+#include "olap/key_coder.h"
+#include "olap/olap_common.h"
 #include "olap/primary_key_index.h"
-#include "olap/row_cursor.h"                      // RowCursor
+#include "olap/row_cursor.h"                      // IWYU pragma: keep
 #include "olap/rowset/rowset_writer_context.h"    // RowsetWriterContext
 #include "olap/rowset/segment_v2/column_writer.h" // ColumnWriter
 #include "olap/rowset/segment_v2/page_io.h"
-#include "olap/schema.h"
+#include "olap/rowset/segment_v2/page_pointer.h"
 #include "olap/short_key_index.h"
 #include "runtime/memory/mem_tracker.h"
 #include "service/point_query_executor.h"
+#include "util/coding.h"
 #include "util/crc32c.h"
 #include "util/faststring.h"
 #include "util/key_util.h"
 #include "vec/common/schema_util.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/olap/olap_data_convertor.h"
 
 namespace doris {
 namespace segment_v2 {
diff --git a/be/src/olap/rowset/segment_v2/segment_writer.h b/be/src/olap/rowset/segment_v2/segment_writer.h
index 8ffd4eaa5b..c907a357df 100644
--- a/be/src/olap/rowset/segment_v2/segment_writer.h
+++ b/be/src/olap/rowset/segment_v2/segment_writer.h
@@ -17,29 +17,38 @@
 
 #pragma once
 
+#include <butil/macros.h>
+#include <gen_cpp/olap_file.pb.h>
+#include <gen_cpp/segment_v2.pb.h>
+#include <stddef.h>
+
 #include <cstdint>
+#include <functional>
 #include <memory> // unique_ptr
 #include <string>
+#include <unordered_set>
 #include <vector>
 
 #include "common/status.h" // Status
-#include "gen_cpp/segment_v2.pb.h"
-#include "gutil/macros.h"
+#include "gutil/strings/substitute.h"
+#include "olap/olap_define.h"
+#include "olap/rowset/segment_v2/column_writer.h"
 #include "olap/tablet_schema.h"
 #include "util/faststring.h"
-#include "vec/core/block.h"
-#include "vec/olap/olap_data_convertor.h"
+#include "util/slice.h"
 
 namespace doris {
+namespace vectorized {
+class Block;
+class IOlapColumnDataAccessor;
+class OlapBlockDataConvertor;
+} // namespace vectorized
 
 // TODO(lingbin): Should be a conf that can be dynamically adjusted, or a member in the context
 const uint32_t MAX_SEGMENT_SIZE = static_cast<uint32_t>(OLAP_MAX_COLUMN_SEGMENT_FILE_SIZE *
                                                         OLAP_COLUMN_FILE_SEGMENT_SIZE_SCALE);
 class DataDir;
 class MemTracker;
-class RowCursor;
-class TabletSchema;
-class TabletColumn;
 class ShortKeyIndexBuilder;
 class PrimaryKeyIndexBuilder;
 class KeyCoder;
@@ -51,8 +60,6 @@ class FileWriter;
 
 namespace segment_v2 {
 
-class ColumnWriter;
-
 extern const char* k_segment_magic;
 extern const uint32_t k_segment_magic_length;
 
diff --git a/be/src/olap/rowset/segment_v2/zone_map_index.cpp b/be/src/olap/rowset/segment_v2/zone_map_index.cpp
index 5b63bcda03..9a4ec628c2 100644
--- a/be/src/olap/rowset/segment_v2/zone_map_index.cpp
+++ b/be/src/olap/rowset/segment_v2/zone_map_index.cpp
@@ -17,12 +17,27 @@
 
 #include "olap/rowset/segment_v2/zone_map_index.h"
 
+#include <gen_cpp/segment_v2.pb.h>
+#include <glog/logging.h>
+
+#include <algorithm>
+#include <type_traits>
+
+#include "olap/olap_common.h"
 #include "olap/rowset/segment_v2/encoding_info.h"
 #include "olap/rowset/segment_v2/indexed_column_reader.h"
 #include "olap/rowset/segment_v2/indexed_column_writer.h"
 #include "olap/types.h"
+#include "runtime/primitive_type.h"
+#include "util/slice.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_string.h"
+#include "vec/common/string_ref.h"
+#include "vec/common/unaligned.h"
+#include "vec/data_types/data_type.h"
 
 namespace doris {
+struct uint24_t;
 
 namespace segment_v2 {
 
diff --git a/be/src/olap/rowset/segment_v2/zone_map_index.h b/be/src/olap/rowset/segment_v2/zone_map_index.h
index 073ea2da31..ed3418dc48 100644
--- a/be/src/olap/rowset/segment_v2/zone_map_index.h
+++ b/be/src/olap/rowset/segment_v2/zone_map_index.h
@@ -17,16 +17,19 @@
 
 #pragma once
 
+#include <gen_cpp/segment_v2.pb.h>
+#include <stddef.h>
+#include <stdint.h>
+
 #include <memory>
 #include <string>
+#include <utility>
 #include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/segment_v2.pb.h"
-#include "io/fs/file_reader.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "olap/field.h"
-#include "olap/rowset/segment_v2/binary_plain_page.h"
-#include "util/slice.h"
+#include "runtime/define_primitive_type.h"
 #include "vec/common/arena.h"
 
 namespace doris {
diff --git a/be/src/olap/rowset/unique_rowset_id_generator.cpp b/be/src/olap/rowset/unique_rowset_id_generator.cpp
index 9db7faf35e..0d7d77915e 100644
--- a/be/src/olap/rowset/unique_rowset_id_generator.cpp
+++ b/be/src/olap/rowset/unique_rowset_id_generator.cpp
@@ -17,9 +17,12 @@
 
 #include "olap/rowset/unique_rowset_id_generator.h"
 
+#include <memory>
+#include <mutex>
+
 #include "util/doris_metrics.h"
+#include "util/metrics.h"
 #include "util/spinlock.h"
-#include "util/stack_util.h"
 #include "util/uid_util.h"
 
 namespace doris {
diff --git a/be/src/olap/rowset/unique_rowset_id_generator.h b/be/src/olap/rowset/unique_rowset_id_generator.h
index ef8d4cfde2..051eefd245 100644
--- a/be/src/olap/rowset/unique_rowset_id_generator.h
+++ b/be/src/olap/rowset/unique_rowset_id_generator.h
@@ -17,8 +17,12 @@
 
 #pragma once
 
-#include <atomic>
+#include <stdint.h>
 
+#include <atomic>
+#include <unordered_set>
+
+#include "olap/olap_common.h"
 #include "olap/rowset/rowset_id_generator.h"
 #include "util/spinlock.h"
 #include "util/uid_util.h"
diff --git a/be/src/olap/rowset/vertical_beta_rowset_writer.cpp b/be/src/olap/rowset/vertical_beta_rowset_writer.cpp
index e1e3ab46de..6525407a91 100644
--- a/be/src/olap/rowset/vertical_beta_rowset_writer.cpp
+++ b/be/src/olap/rowset/vertical_beta_rowset_writer.cpp
@@ -17,7 +17,28 @@
 
 #include "olap/rowset/vertical_beta_rowset_writer.h"
 
+#include <gen_cpp/olap_file.pb.h>
+
+#include <algorithm>
+#include <atomic>
+#include <mutex>
+#include <ostream>
+#include <string>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/logging.h"
+#include "gutil/strings/substitute.h"
+#include "io/fs/file_reader_writer_fwd.h"
+#include "io/fs/file_system.h"
+#include "io/fs/file_writer.h"
 #include "olap/rowset/beta_rowset.h"
+#include "olap/rowset/rowset_meta.h"
+#include "olap/rowset/rowset_writer_context.h"
+#include "util/slice.h"
+#include "util/spinlock.h"
+#include "vec/core/block.h"
 
 namespace doris {
 using namespace ErrorCode;
diff --git a/be/src/olap/rowset/vertical_beta_rowset_writer.h b/be/src/olap/rowset/vertical_beta_rowset_writer.h
index 56e2fb4d6a..71b132b9b4 100644
--- a/be/src/olap/rowset/vertical_beta_rowset_writer.h
+++ b/be/src/olap/rowset/vertical_beta_rowset_writer.h
@@ -17,10 +17,20 @@
 
 #pragma once
 
+#include <stddef.h>
+#include <stdint.h>
+
+#include <memory>
+#include <vector>
+
+#include "common/status.h"
 #include "olap/rowset/beta_rowset_writer.h"
 #include "olap/rowset/segment_v2/segment_writer.h"
 
 namespace doris {
+namespace vectorized {
+class Block;
+} // namespace vectorized
 
 // for vertical compaction
 class VerticalBetaRowsetWriter : public BetaRowsetWriter {
diff --git a/be/src/olap/schema.cpp b/be/src/olap/schema.cpp
index ab75f1f93a..373773fffe 100644
--- a/be/src/olap/schema.cpp
+++ b/be/src/olap/schema.cpp
@@ -17,12 +17,24 @@
 
 #include "olap/schema.h"
 
-#include "olap/uint24.h"
-#include "vec/columns/column_complex.h"
+#include <glog/logging.h>
+
+#include <boost/iterator/iterator_facade.hpp>
+#include <ostream>
+#include <unordered_set>
+#include <utility>
+
+#include "common/config.h"
+#include "runtime/define_primitive_type.h"
+#include "vec/columns/column_array.h"
 #include "vec/columns/column_dictionary.h"
+#include "vec/columns/column_map.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/columns/column_struct.h"
+#include "vec/columns/columns_number.h"
 #include "vec/columns/predicate_column.h"
 #include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_factory.hpp"
 
 namespace doris {
diff --git a/be/src/olap/schema.h b/be/src/olap/schema.h
index 4d9594b344..b592563bd0 100644
--- a/be/src/olap/schema.h
+++ b/be/src/olap/schema.h
@@ -17,14 +17,22 @@
 
 #pragma once
 
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <memory>
+#include <string>
 #include <vector>
 
 #include "common/consts.h"
+#include "io/io_common.h"
 #include "olap/field.h"
-#include "olap/row_cursor_cell.h"
+#include "olap/olap_common.h"
 #include "olap/tablet_schema.h"
-#include "olap/types.h"
-#include "runtime/descriptors.h"
+#include "olap/utils.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 
 namespace doris {
 
diff --git a/be/src/olap/schema_change.cpp b/be/src/olap/schema_change.cpp
index 2089b7c62d..9742b4d095 100644
--- a/be/src/olap/schema_change.cpp
+++ b/be/src/olap/schema_change.cpp
@@ -17,37 +17,66 @@
 
 #include "olap/schema_change.h"
 
+#include <gen_cpp/AgentService_types.h>
+#include <gen_cpp/Exprs_types.h>
 #include <gen_cpp/olap_file.pb.h>
 
-#include "common/config.h"
+#include <algorithm>
+#include <exception>
+#include <map>
+#include <mutex>
+#include <roaring/roaring.hh>
+#include <tuple>
+
+#include "common/logging.h"
 #include "common/status.h"
+#include "gutil/hash/hash.h"
 #include "gutil/integral_types.h"
+#include "gutil/strings/numbers.h"
+#include "io/fs/file_system.h"
+#include "io/io_common.h"
+#include "olap/data_dir.h"
+#include "olap/delete_handler.h"
+#include "olap/field.h"
+#include "olap/iterators.h"
 #include "olap/merger.h"
 #include "olap/olap_common.h"
-#include "olap/row_cursor.h"
+#include "olap/olap_define.h"
 #include "olap/rowset/beta_rowset.h"
-#include "olap/rowset/rowset_id_generator.h"
+#include "olap/rowset/rowset_meta.h"
+#include "olap/rowset/rowset_reader_context.h"
+#include "olap/rowset/rowset_writer_context.h"
 #include "olap/rowset/segment_v2/column_reader.h"
 #include "olap/rowset/segment_v2/inverted_index_desc.h"
 #include "olap/rowset/segment_v2/inverted_index_writer.h"
+#include "olap/rowset/segment_v2/segment.h"
+#include "olap/schema.h"
 #include "olap/segment_loader.h"
 #include "olap/storage_engine.h"
 #include "olap/tablet.h"
+#include "olap/tablet_manager.h"
+#include "olap/tablet_meta.h"
 #include "olap/tablet_schema.h"
 #include "olap/types.h"
 #include "olap/utils.h"
 #include "olap/wrapper_field.h"
 #include "runtime/memory/mem_tracker.h"
+#include "runtime/runtime_state.h"
 #include "util/defer_op.h"
 #include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/aggregate_functions/aggregate_function_reader.h"
-#include "vec/aggregate_functions/aggregate_function_simple_factory.h"
 #include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/common/assert_cast.h"
 #include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/exprs/vexpr.h"
 #include "vec/exprs/vexpr_context.h"
+#include "vec/olap/olap_data_convertor.h"
 
 namespace doris {
+class CollectionValue;
+
 using namespace ErrorCode;
 
 constexpr int ALTER_TABLE_BATCH_SIZE = 4096;
diff --git a/be/src/olap/schema_change.h b/be/src/olap/schema_change.h
index 16226dfead..b74113c8c7 100644
--- a/be/src/olap/schema_change.h
+++ b/be/src/olap/schema_change.h
@@ -17,24 +17,52 @@
 
 #pragma once
 
-#include <cstdint>
+#include <butil/macros.h>
+#include <fmt/format.h>
+#include <gen_cpp/Descriptors_types.h>
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <memory>
+#include <ostream>
+#include <set>
+#include <shared_mutex>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
+#include <vector>
+
+#include "common/config.h"
+#include "common/object_pool.h"
 #include "common/status.h"
-#include "gen_cpp/AgentService_types.h"
 #include "olap/column_mapping.h"
-#include "olap/delete_handler.h"
+#include "olap/olap_common.h"
 #include "olap/rowset/rowset.h"
+#include "olap/rowset/rowset_reader.h"
 #include "olap/rowset/rowset_writer.h"
+#include "olap/rowset/segment_v2/inverted_index_writer.h"
 #include "olap/tablet.h"
-#include "vec/columns/column.h"
-#include "vec/core/block.h"
-#include "vec/olap/olap_data_convertor.h"
+#include "olap/tablet_schema.h"
+#include "runtime/descriptors.h"
+#include "runtime/memory/mem_tracker.h"
+#include "vec/data_types/data_type.h"
 
 namespace doris {
+class DeleteHandler;
+class Field;
+class TAlterInvertedIndexReq;
+class TAlterTabletReqV2;
+class TExpr;
+enum AlterTabletType : int;
+enum RowsetTypePB : int;
+enum SegmentsOverlapPB : int;
 
-namespace segment_v2 {
-class InvertedIndexColumnWriter;
-}
+namespace vectorized {
+class Block;
+class OlapBlockDataConvertor;
+} // namespace vectorized
 
 class BlockChanger {
 public:
diff --git a/be/src/olap/segment_loader.cpp b/be/src/olap/segment_loader.cpp
index fa1e6c8751..eaf6651fc5 100644
--- a/be/src/olap/segment_loader.cpp
+++ b/be/src/olap/segment_loader.cpp
@@ -17,8 +17,9 @@
 
 #include "olap/segment_loader.h"
 
-#include "olap/rowset/rowset.h"
-#include "olap/tablet_schema.h"
+#include "common/config.h"
+#include "olap/olap_define.h"
+#include "olap/rowset/beta_rowset.h"
 #include "util/stopwatch.hpp"
 
 namespace doris {
diff --git a/be/src/olap/segment_loader.h b/be/src/olap/segment_loader.h
index 03a54acf55..13da58a2e5 100644
--- a/be/src/olap/segment_loader.h
+++ b/be/src/olap/segment_loader.h
@@ -17,20 +17,28 @@
 
 #pragma once
 
+#include <butil/macros.h>
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <atomic>
 #include <memory>
+#include <ostream>
 #include <string>
 #include <utility>
+#include <vector>
 
-#include "gutil/macros.h" // for DISALLOW_COPY_AND_ASSIGN
+#include "common/status.h"
 #include "olap/lru_cache.h"
 #include "olap/olap_common.h" // for rowset id
-#include "olap/rowset/beta_rowset.h"
-#include "olap/tablet_schema.h"
+#include "olap/rowset/segment_v2/segment.h"
 #include "util/time.h"
 
 namespace doris {
 
 class SegmentCacheHandle;
+class BetaRowset;
 
 // SegmentLoader is used to load the Segment of BetaRowset.
 // An LRUCache is encapsulated inside it, which is used to cache the opened segments.
diff --git a/be/src/olap/short_key_index.cpp b/be/src/olap/short_key_index.cpp
index 19055b3b12..69622cee45 100644
--- a/be/src/olap/short_key_index.cpp
+++ b/be/src/olap/short_key_index.cpp
@@ -17,7 +17,9 @@
 
 #include "olap/short_key_index.h"
 
-#include <string>
+#include <gen_cpp/segment_v2.pb.h>
+
+#include <ostream>
 
 #include "gutil/strings/substitute.h"
 #include "util/coding.h"
diff --git a/be/src/olap/short_key_index.h b/be/src/olap/short_key_index.h
index a217a928aa..a5ed018f72 100644
--- a/be/src/olap/short_key_index.h
+++ b/be/src/olap/short_key_index.h
@@ -17,14 +17,16 @@
 
 #pragma once
 
+#include <gen_cpp/segment_v2.pb.h>
+#include <glog/logging.h>
+#include <sys/types.h>
+
+#include <algorithm>
 #include <cstdint>
 #include <iterator>
-#include <string>
 #include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/segment_v2.pb.h"
-#include "util/debug_util.h"
 #include "util/faststring.h"
 #include "util/slice.h"
 
diff --git a/be/src/olap/skiplist.h b/be/src/olap/skiplist.h
index 5c4a2b40d6..477a3e1e12 100644
--- a/be/src/olap/skiplist.h
+++ b/be/src/olap/skiplist.h
@@ -26,10 +26,11 @@
 //
 // ... prev vs. next pointer ordering ...
 
+#include <gen_cpp/olap_file.pb.h>
+
 #include <atomic>
 
 #include "common/logging.h"
-#include "gen_cpp/olap_file.pb.h"
 #include "util/random.h"
 #include "vec/common/arena.h"
 
diff --git a/be/src/olap/snapshot_manager.cpp b/be/src/olap/snapshot_manager.cpp
index 687f76a0ec..cf498c0b7b 100644
--- a/be/src/olap/snapshot_manager.cpp
+++ b/be/src/olap/snapshot_manager.cpp
@@ -17,27 +17,44 @@
 
 #include "olap/snapshot_manager.h"
 
-#include <ctype.h>
-#include <errno.h>
-#include <stdio.h>
+#include <fmt/format.h>
+#include <gen_cpp/AgentService_types.h>
+#include <gen_cpp/Types_constants.h>
+#include <gen_cpp/olap_file.pb.h>
 #include <thrift/protocol/TDebugProtocol.h>
 
 #include <algorithm>
+#include <ctime>
 #include <filesystem>
-#include <iterator>
+#include <list>
 #include <map>
+#include <new>
+#include <ostream>
 #include <set>
+#include <shared_mutex>
+#include <unordered_map>
+#include <utility>
 
+#include "common/config.h"
+#include "common/logging.h"
 #include "common/status.h"
-#include "gen_cpp/Types_constants.h"
 #include "io/fs/local_file_system.h"
+#include "olap/data_dir.h"
+#include "olap/olap_common.h"
+#include "olap/olap_define.h"
 #include "olap/rowset/rowset.h"
 #include "olap/rowset/rowset_factory.h"
+#include "olap/rowset/rowset_meta.h"
 #include "olap/rowset/rowset_writer.h"
+#include "olap/rowset/rowset_writer_context.h"
 #include "olap/storage_engine.h"
+#include "olap/tablet_manager.h"
 #include "olap/tablet_meta.h"
 #include "olap/tablet_schema.h"
+#include "olap/tablet_schema_cache.h"
+#include "olap/utils.h"
 #include "runtime/thread_context.h"
+#include "util/uid_util.h"
 
 using std::filesystem::path;
 using std::map;
diff --git a/be/src/olap/snapshot_manager.h b/be/src/olap/snapshot_manager.h
index 982bf36375..3536cebd1a 100644
--- a/be/src/olap/snapshot_manager.h
+++ b/be/src/olap/snapshot_manager.h
@@ -17,27 +17,23 @@
 
 #pragma once
 
-#include <condition_variable>
-#include <ctime>
-#include <list>
-#include <map>
+#include <stdint.h>
+
+#include <memory>
 #include <mutex>
-#include <set>
 #include <string>
-#include <thread>
 #include <vector>
 
 #include "common/status.h"
-#include "olap/data_dir.h"
-#include "olap/field.h"
-#include "olap/olap_common.h"
-#include "olap/olap_define.h"
-#include "olap/push_handler.h"
+#include "olap/rowset/rowset.h"
 #include "olap/tablet.h"
-#include "olap/tablet_meta_manager.h"
-#include "util/doris_metrics.h"
+#include "olap/tablet_schema.h"
+#include "runtime/memory/mem_tracker.h"
 
 namespace doris {
+class RowsetMetaPB;
+class TSnapshotRequest;
+struct RowsetId;
 
 class SnapshotManager {
 public:
diff --git a/be/src/olap/storage_engine.cpp b/be/src/olap/storage_engine.cpp
index 154145d823..722220d301 100644
--- a/be/src/olap/storage_engine.cpp
+++ b/be/src/olap/storage_engine.cpp
@@ -17,44 +17,61 @@
 
 #include "olap/storage_engine.h"
 
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
+#include <fmt/format.h>
+#include <gen_cpp/AgentService_types.h>
 #include <rapidjson/document.h>
-#include <signal.h>
-#include <sys/syscall.h>
+#include <rapidjson/encodings.h>
+#include <rapidjson/prettywriter.h>
+#include <rapidjson/stringbuffer.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/resource.h>
 #include <thrift/protocol/TDebugProtocol.h>
 
 #include <algorithm>
-#include <boost/algorithm/string.hpp>
-#include <boost/algorithm/string/classification.hpp>
-#include <boost/algorithm/string/split.hpp>
-#include <boost/algorithm/string/trim.hpp>
-#include <cstdio>
+#include <boost/algorithm/string/case_conv.hpp>
+#include <boost/container/detail/std_fwd.hpp>
 #include <filesystem>
+#include <iterator>
+#include <list>
 #include <new>
-#include <queue>
+#include <ostream>
 #include <random>
 #include <set>
+#include <thread>
+#include <utility>
 
 #include "agent/task_worker_pool.h"
+#include "common/config.h"
+#include "common/logging.h"
+#include "gutil/strings/substitute.h"
 #include "io/fs/local_file_system.h"
 #include "olap/base_compaction.h"
 #include "olap/cumulative_compaction.h"
 #include "olap/data_dir.h"
-#include "olap/lru_cache.h"
 #include "olap/memtable_flush_executor.h"
-#include "olap/push_handler.h"
-#include "olap/reader.h"
+#include "olap/olap_define.h"
+#include "olap/rowset/rowset_meta.h"
 #include "olap/rowset/rowset_meta_manager.h"
 #include "olap/rowset/unique_rowset_id_generator.h"
-#include "olap/schema_change.h"
 #include "olap/segment_loader.h"
+#include "olap/tablet_manager.h"
 #include "olap/tablet_meta.h"
-#include "olap/tablet_meta_manager.h"
-#include "olap/utils.h"
+#include "olap/task/engine_task.h"
+#include "olap/txn_manager.h"
+#include "runtime/memory/mem_tracker.h"
+#include "runtime/stream_load/stream_load_recorder.h"
 #include "util/doris_metrics.h"
-#include "util/pretty_printer.h"
-#include "util/scoped_cleanup.h"
-#include "util/time.h"
+#include "util/metrics.h"
+#include "util/priority_thread_pool.hpp"
+#include "util/spinlock.h"
+#include "util/stopwatch.hpp"
+#include "util/thread.h"
+#include "util/threadpool.h"
 #include "util/trace.h"
+#include "util/uid_util.h"
 
 using apache::thrift::ThriftDebugString;
 using std::filesystem::canonical;
@@ -69,7 +86,6 @@ using std::map;
 using std::nothrow;
 using std::pair;
 using std::set;
-using std::set_difference;
 using std::string;
 using std::stringstream;
 using std::vector;
diff --git a/be/src/olap/storage_engine.h b/be/src/olap/storage_engine.h
index db113f1d93..481353ec0d 100644
--- a/be/src/olap/storage_engine.h
+++ b/be/src/olap/storage_engine.h
@@ -17,48 +17,54 @@
 
 #pragma once
 
-#include <pthread.h>
-#include <rapidjson/document.h>
+#include <butil/macros.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/olap_file.pb.h>
+#include <stdint.h>
 
+#include <atomic>
 #include <condition_variable>
 #include <ctime>
-#include <list>
 #include <map>
+#include <memory>
 #include <mutex>
 #include <set>
 #include <string>
-#include <thread>
+#include <unordered_map>
+#include <unordered_set>
 #include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/AgentService_types.h"
-#include "gen_cpp/BackendService_types.h"
-#include "gen_cpp/MasterService_types.h"
 #include "gutil/ref_counted.h"
 #include "olap/compaction_permit_limiter.h"
 #include "olap/olap_common.h"
-#include "olap/olap_define.h"
-#include "olap/olap_meta.h"
 #include "olap/options.h"
+#include "olap/rowset/rowset.h"
 #include "olap/rowset/rowset_id_generator.h"
-#include "olap/tablet_manager.h"
-#include "olap/task/engine_task.h"
-#include "olap/txn_manager.h"
+#include "olap/rowset/segment_v2/segment.h"
+#include "olap/tablet.h"
 #include "runtime/heartbeat_flags.h"
-#include "runtime/stream_load/stream_load_recorder.h"
 #include "util/countdown_latch.h"
-#include "util/thread.h"
-#include "util/threadpool.h"
 
 namespace doris {
 
 class DataDir;
 class EngineTask;
-class BlockManager;
 class MemTableFlushExecutor;
-class Tablet;
 class TaskWorkerPool;
 class BetaRowsetWriter;
+class BaseCompaction;
+class CumulativeCompaction;
+class CumulativeCompactionPolicy;
+class MemTracker;
+class PriorityThreadPool;
+class StreamLoadRecorder;
+class TCloneReq;
+class TCreateTabletReq;
+class TabletManager;
+class Thread;
+class ThreadPool;
+class TxnManager;
 
 using SegCompactionCandidates = std::vector<segment_v2::SegmentSharedPtr>;
 using SegCompactionCandidatesSharedPtr = std::shared_ptr<SegCompactionCandidates>;
diff --git a/be/src/olap/storage_policy.cpp b/be/src/olap/storage_policy.cpp
index 68f561397d..494c93cec1 100644
--- a/be/src/olap/storage_policy.cpp
+++ b/be/src/olap/storage_policy.cpp
@@ -17,6 +17,10 @@
 
 #include "olap/storage_policy.h"
 
+#include <glog/logging.h>
+
+#include <algorithm>
+#include <cstdlib>
 #include <mutex>
 #include <unordered_map>
 
diff --git a/be/src/olap/storage_policy.h b/be/src/olap/storage_policy.h
index 04f05540ae..0a566d1002 100644
--- a/be/src/olap/storage_policy.h
+++ b/be/src/olap/storage_policy.h
@@ -17,6 +17,15 @@
 
 #pragma once
 
+#include <fmt/format.h>
+#include <stdint.h>
+
+#include <memory>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
 #include "io/fs/file_system.h"
 #include "io/fs/remote_file_system.h"
 
diff --git a/be/src/olap/tablet.cpp b/be/src/olap/tablet.cpp
index 7603d6fbe7..96311c4c77 100644
--- a/be/src/olap/tablet.cpp
+++ b/be/src/olap/tablet.cpp
@@ -17,68 +17,113 @@
 
 #include "olap/tablet.h"
 
+#include <assert.h>
+#include <butil/logging.h>
 #include <bvar/reducer.h>
 #include <bvar/window.h>
-#include <ctype.h>
-#include <fmt/core.h>
-#include <glog/logging.h>
-#include <opentelemetry/common/threadlocal.h>
-#include <pthread.h>
+#include <fmt/format.h>
+#include <gen_cpp/FrontendService_types.h>
+#include <gen_cpp/MasterService_types.h>
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/olap_file.pb.h>
+#include <gen_cpp/types.pb.h>
+#include <rapidjson/document.h>
+#include <rapidjson/encodings.h>
 #include <rapidjson/prettywriter.h>
+#include <rapidjson/rapidjson.h>
 #include <rapidjson/stringbuffer.h>
-#include <stdio.h>
-#include <sys/stat.h>
+#include <stdlib.h>
+#include <time.h>
 
 #include <algorithm>
 #include <atomic>
-#include <cstdint>
+#include <boost/container/detail/std_fwd.hpp>
+#include <roaring/roaring.hh>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <filesystem>
+#include <iterator>
+#include <limits>
 #include <map>
 #include <memory>
 #include <mutex>
 #include <set>
 #include <shared_mutex>
 #include <string>
+#include <tuple>
+#include <type_traits>
 #include <unordered_set>
 
 #include "agent/utils.h"
 #include "common/config.h"
+#include "common/consts.h"
 #include "common/logging.h"
 #include "common/status.h"
+#include "gutil/ref_counted.h"
 #include "gutil/strings/stringpiece.h"
+#include "gutil/strings/substitute.h"
+#include "io/fs/file_reader.h"
+#include "io/fs/file_reader_writer_fwd.h"
+#include "io/fs/file_system.h"
+#include "io/fs/file_writer.h"
 #include "io/fs/path.h"
 #include "io/fs/remote_file_system.h"
 #include "olap/base_compaction.h"
 #include "olap/base_tablet.h"
 #include "olap/cumulative_compaction.h"
+#include "olap/cumulative_compaction_policy.h"
 #include "olap/olap_common.h"
 #include "olap/olap_define.h"
-#include "olap/reader.h"
-#include "olap/row_cursor.h"
+#include "olap/olap_meta.h"
+#include "olap/primary_key_index.h"
+#include "olap/rowid_conversion.h"
+#include "olap/rowset/beta_rowset.h"
 #include "olap/rowset/rowset.h"
 #include "olap/rowset/rowset_factory.h"
 #include "olap/rowset/rowset_meta.h"
 #include "olap/rowset/rowset_meta_manager.h"
+#include "olap/rowset/rowset_writer.h"
+#include "olap/rowset/rowset_writer_context.h"
 #include "olap/rowset/segment_v2/column_reader.h"
+#include "olap/rowset/segment_v2/common.h"
+#include "olap/rowset/segment_v2/indexed_column_reader.h"
 #include "olap/schema_change.h"
 #include "olap/storage_engine.h"
 #include "olap/storage_policy.h"
+#include "olap/tablet_manager.h"
 #include "olap/tablet_meta.h"
-#include "olap/tablet_meta_manager.h"
 #include "olap/tablet_schema.h"
 #include "olap/txn_manager.h"
+#include "olap/types.h"
+#include "olap/utils.h"
 #include "segment_loader.h"
 #include "service/point_query_executor.h"
 #include "util/defer_op.h"
-#include "util/path_util.h"
+#include "util/doris_metrics.h"
 #include "util/pretty_printer.h"
 #include "util/scoped_cleanup.h"
+#include "util/stopwatch.hpp"
+#include "util/threadpool.h"
 #include "util/time.h"
 #include "util/trace.h"
 #include "util/uid_util.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_string.h"
+#include "vec/common/string_ref.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_factory.hpp"
 #include "vec/jsonb/serialize.h"
 
 namespace doris {
+class TupleDescriptor;
+
+namespace vectorized {
+class Block;
+} // namespace vectorized
+
 using namespace ErrorCode;
 
 using std::pair;
diff --git a/be/src/olap/tablet.h b/be/src/olap/tablet.h
index faa5a5c8e0..1fa6d1bc7f 100644
--- a/be/src/olap/tablet.h
+++ b/be/src/olap/tablet.h
@@ -17,48 +17,66 @@
 
 #pragma once
 
+#include <butil/macros.h>
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
+
 #include <atomic>
 #include <functional>
+#include <list>
+#include <map>
 #include <memory>
 #include <mutex>
+#include <ostream>
 #include <set>
 #include <shared_mutex>
 #include <string>
 #include <unordered_map>
-#include <unordered_set>
+#include <utility>
 #include <vector>
 
-#include "gen_cpp/AgentService_types.h"
-#include "gen_cpp/MasterService_types.h"
-#include "gen_cpp/olap_file.pb.h"
-#include "io/fs/file_system.h"
+#include "common/config.h"
+#include "common/status.h"
 #include "olap/base_tablet.h"
-#include "olap/cumulative_compaction_policy.h"
 #include "olap/data_dir.h"
 #include "olap/olap_common.h"
-#include "olap/olap_define.h"
-#include "olap/olap_tuple.h"
 #include "olap/rowset/rowset.h"
+#include "olap/rowset/rowset_meta.h"
 #include "olap/rowset/rowset_reader.h"
 #include "olap/rowset/rowset_tree.h"
 #include "olap/rowset/segment_v2/segment.h"
 #include "olap/tablet_meta.h"
-#include "olap/utils.h"
+#include "olap/tablet_schema.h"
 #include "olap/version_graph.h"
+#include "util/metrics.h"
 #include "util/once.h"
+#include "util/slice.h"
 
 namespace doris {
 
-class DataDir;
 class Tablet;
-class TabletMeta;
 class CumulativeCompactionPolicy;
 class CumulativeCompaction;
 class BaseCompaction;
 class RowsetWriter;
-
 struct TabletTxnInfo;
 struct RowsetWriterContext;
+class RowIdConversion;
+class TTabletInfo;
+class TabletMetaPB;
+class TupleDescriptor;
+enum CompressKind : int;
+
+namespace io {
+class RemoteFileSystem;
+} // namespace io
+namespace vectorized {
+class Block;
+} // namespace vectorized
+struct RowLocation;
+enum KeysType : int;
+enum SortType : int;
 
 using TabletSharedPtr = std::shared_ptr<Tablet>;
 
diff --git a/be/src/olap/tablet_manager.cpp b/be/src/olap/tablet_manager.cpp
index 8ad1fe9d94..94986cdcfb 100644
--- a/be/src/olap/tablet_manager.cpp
+++ b/be/src/olap/tablet_manager.cpp
@@ -17,33 +17,55 @@
 
 #include "olap/tablet_manager.h"
 
+#include <fmt/format.h>
+#include <gen_cpp/AgentService_types.h>
+#include <gen_cpp/BackendService_types.h>
+#include <gen_cpp/Descriptors_types.h>
+#include <gen_cpp/MasterService_types.h>
 #include <gen_cpp/Types_types.h>
+#include <gen_cpp/olap_file.pb.h>
 #include <re2/re2.h>
+#include <unistd.h>
 
 #include <algorithm>
-#include <cstdint>
-#include <cstdio>
-#include <filesystem>
+#include <list>
+#include <ostream>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/config.h"
+#include "common/logging.h"
+#include "gutil/integral_types.h"
 #include "gutil/strings/strcat.h"
-#include "olap/base_compaction.h"
-#include "olap/cumulative_compaction.h"
+#include "gutil/strings/substitute.h"
+#include "io/fs/local_file_system.h"
 #include "olap/data_dir.h"
 #include "olap/olap_common.h"
-#include "olap/push_handler.h"
+#include "olap/olap_define.h"
+#include "olap/rowset/rowset.h"
+#include "olap/storage_engine.h"
 #include "olap/tablet.h"
 #include "olap/tablet_meta.h"
 #include "olap/tablet_meta_manager.h"
-#include "rapidjson/prettywriter.h"
-#include "rapidjson/stringbuffer.h"
+#include "olap/tablet_schema.h"
+#include "olap/txn_manager.h"
+#include "runtime/exec_env.h"
+#include "runtime/memory/mem_tracker.h"
 #include "runtime/thread_context.h"
 #include "service/backend_options.h"
 #include "util/doris_metrics.h"
 #include "util/histogram.h"
+#include "util/metrics.h"
 #include "util/path_util.h"
 #include "util/scoped_cleanup.h"
+#include "util/stopwatch.hpp"
 #include "util/time.h"
 #include "util/trace.h"
+#include "util/uid_util.h"
+
+namespace doris {
+class CumulativeCompactionPolicy;
+} // namespace doris
 
 using std::list;
 using std::map;
diff --git a/be/src/olap/tablet_manager.h b/be/src/olap/tablet_manager.h
index dca1b635ad..0ddeb4aaac 100644
--- a/be/src/olap/tablet_manager.h
+++ b/be/src/olap/tablet_manager.h
@@ -17,30 +17,37 @@
 
 #pragma once
 
-#include <atomic>
-#include <list>
+#include <butil/macros.h>
+#include <gen_cpp/BackendService_types.h>
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <functional>
 #include <map>
+#include <memory>
 #include <mutex>
 #include <set>
 #include <shared_mutex>
 #include <string>
 #include <unordered_map>
 #include <unordered_set>
+#include <utility>
 #include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/AgentService_types.h"
-#include "gen_cpp/BackendService_types.h"
-#include "gen_cpp/MasterService_types.h"
-#include "olap/olap_define.h"
-#include "olap/olap_meta.h"
-#include "olap/options.h"
+#include "olap/olap_common.h"
 #include "olap/tablet.h"
+#include "olap/tablet_meta.h"
 
 namespace doris {
 
-class Tablet;
 class DataDir;
+class CumulativeCompactionPolicy;
+class MemTracker;
+class TCreateTabletReq;
+class TTablet;
+class TTabletInfo;
 
 // TabletManager provides get, add, delete tablet method for storage engine
 // NOTE: If you want to add a method that needs to hold meta-lock before you can call it,
diff --git a/be/src/olap/tablet_meta.cpp b/be/src/olap/tablet_meta.cpp
index a318dc508b..f5548e8061 100644
--- a/be/src/olap/tablet_meta.cpp
+++ b/be/src/olap/tablet_meta.cpp
@@ -17,17 +17,29 @@
 
 #include "olap/tablet_meta.h"
 
-#include <sstream>
+#include <gen_cpp/Descriptors_types.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/olap_common.pb.h>
+#include <gen_cpp/olap_file.pb.h>
+#include <gen_cpp/segment_v2.pb.h>
+#include <gen_cpp/types.pb.h>
+#include <json2pb/pb_to_json.h>
+#include <time.h>
 
-#include "common/consts.h"
+#include <set>
+#include <utility>
+
+#include "common/config.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "io/fs/file_writer.h"
+#include "olap/data_dir.h"
 #include "olap/file_header.h"
 #include "olap/olap_common.h"
 #include "olap/olap_define.h"
 #include "olap/tablet_meta_manager.h"
+#include "olap/utils.h"
 #include "util/string_util.h"
 #include "util/uid_util.h"
-#include "util/url_coding.h"
 
 using std::string;
 using std::unordered_map;
@@ -108,25 +120,25 @@ TabletMeta::TabletMeta(int64_t table_id, int64_t partition_id, int64_t tablet_id
     // compression_type used to compress segment page
     switch (compression_type) {
     case TCompressionType::NO_COMPRESSION:
-        schema->set_compression_type(NO_COMPRESSION);
+        schema->set_compression_type(segment_v2::NO_COMPRESSION);
         break;
     case TCompressionType::SNAPPY:
-        schema->set_compression_type(SNAPPY);
+        schema->set_compression_type(segment_v2::SNAPPY);
         break;
     case TCompressionType::LZ4:
-        schema->set_compression_type(LZ4);
+        schema->set_compression_type(segment_v2::LZ4);
         break;
     case TCompressionType::LZ4F:
-        schema->set_compression_type(LZ4F);
+        schema->set_compression_type(segment_v2::LZ4F);
         break;
     case TCompressionType::ZLIB:
-        schema->set_compression_type(ZLIB);
+        schema->set_compression_type(segment_v2::ZLIB);
         break;
     case TCompressionType::ZSTD:
-        schema->set_compression_type(ZSTD);
+        schema->set_compression_type(segment_v2::ZSTD);
         break;
     default:
-        schema->set_compression_type(LZ4F);
+        schema->set_compression_type(segment_v2::LZ4F);
         break;
     }
 
diff --git a/be/src/olap/tablet_meta.h b/be/src/olap/tablet_meta.h
index 6410563bd3..d2870a8dfe 100644
--- a/be/src/olap/tablet_meta.h
+++ b/be/src/olap/tablet_meta.h
@@ -17,25 +17,39 @@
 
 #pragma once
 
+#include <gen_cpp/AgentService_types.h>
+#include <gen_cpp/olap_file.pb.h>
+#include <stdint.h>
+
 #include <atomic>
 #include <cstddef>
+#include <map>
+#include <memory>
 #include <mutex>
+#include <ostream>
+#include <roaring/roaring.hh>
 #include <shared_mutex>
 #include <string>
+#include <tuple>
+#include <unordered_map>
 #include <vector>
 
 #include "common/logging.h"
-#include "gen_cpp/olap_file.pb.h"
-#include "olap/delete_handler.h"
+#include "common/status.h"
+#include "gutil/stringprintf.h"
+#include "io/fs/file_system.h"
+#include "olap/lru_cache.h"
 #include "olap/olap_common.h"
-#include "olap/olap_define.h"
-#include "olap/rowid_conversion.h"
-#include "olap/rowset/rowset.h"
 #include "olap/rowset/rowset_meta.h"
 #include "olap/tablet_schema.h"
 #include "util/uid_util.h"
 
+namespace json2pb {
+struct Pb2JsonOptions;
+} // namespace json2pb
+
 namespace doris {
+class TColumn;
 
 // Lifecycle states that a Tablet can be in. Legal state transitions for a
 // Tablet object:
@@ -64,11 +78,10 @@ enum TabletState {
     TABLET_SHUTDOWN
 };
 
-class RowsetMeta;
-class Rowset;
 class DataDir;
 class TabletMeta;
 class DeleteBitmap;
+
 using TabletMetaSharedPtr = std::shared_ptr<TabletMeta>;
 using DeleteBitmapPtr = std::shared_ptr<DeleteBitmap>;
 
diff --git a/be/src/olap/tablet_meta_manager.cpp b/be/src/olap/tablet_meta_manager.cpp
index 336cd1d592..ac4deacd14 100644
--- a/be/src/olap/tablet_meta_manager.cpp
+++ b/be/src/olap/tablet_meta_manager.cpp
@@ -18,19 +18,32 @@
 #include "olap/tablet_meta_manager.h"
 
 #include <fmt/format.h>
+#include <gen_cpp/olap_file.pb.h>
+#include <rocksdb/db.h>
 
 #include <boost/algorithm/string/trim.hpp>
 #include <fstream>
-#include <sstream>
+#include <memory>
 #include <string>
 #include <vector>
 
 #include "common/logging.h"
 #include "json2pb/json_to_pb.h"
 #include "json2pb/pb_to_json.h"
+#include "olap/data_dir.h"
 #include "olap/olap_define.h"
 #include "olap/olap_meta.h"
-#include "olap/storage_engine.h"
+#include "olap/utils.h"
+
+namespace rocksdb {
+class Iterator;
+class Slice;
+class Status;
+struct ColumnFamilyOptions;
+struct DBOptions;
+struct ReadOptions;
+struct WriteOptions;
+} // namespace rocksdb
 
 using rocksdb::DB;
 using rocksdb::DBOptions;
diff --git a/be/src/olap/tablet_meta_manager.h b/be/src/olap/tablet_meta_manager.h
index 2eba2bfd6c..15d6763477 100644
--- a/be/src/olap/tablet_meta_manager.h
+++ b/be/src/olap/tablet_meta_manager.h
@@ -17,13 +17,18 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+
+#include <functional>
 #include <string>
 
-#include "olap/data_dir.h"
-#include "olap/olap_define.h"
+#include "common/status.h"
+#include "gutil/stringprintf.h"
 #include "olap/tablet_meta.h"
 
 namespace doris {
+class DataDir;
+class OlapMeta;
 
 const std::string OLD_HEADER_PREFIX = "hdr_";
 
diff --git a/be/src/olap/tablet_schema.cpp b/be/src/olap/tablet_schema.cpp
index 2f6a54529d..174a1e13fb 100644
--- a/be/src/olap/tablet_schema.cpp
+++ b/be/src/olap/tablet_schema.cpp
@@ -17,14 +17,26 @@
 
 #include "olap/tablet_schema.h"
 
+#include <gen_cpp/Descriptors_types.h>
 #include <gen_cpp/olap_file.pb.h>
+#include <glog/logging.h>
 
+#include <algorithm>
+#include <cctype>
+// IWYU pragma: no_include <bits/std_abs.h>
+#include <cmath> // IWYU pragma: keep
+#include <ostream>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/consts.h"
 #include "common/status.h"
 #include "exec/tablet_info.h"
-#include "gen_cpp/descriptors.pb.h"
+#include "olap/olap_define.h"
+#include "olap/types.h"
 #include "olap/utils.h"
+#include "runtime/thread_context.h"
 #include "tablet_meta.h"
-#include "vec/aggregate_functions/aggregate_function_reader.h"
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
 #include "vec/core/block.h"
 #include "vec/data_types/data_type.h"
diff --git a/be/src/olap/tablet_schema.h b/be/src/olap/tablet_schema.h
index 74d8073031..a77bc03ebd 100644
--- a/be/src/olap/tablet_schema.h
+++ b/be/src/olap/tablet_schema.h
@@ -17,14 +17,25 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/olap_common.pb.h>
+#include <gen_cpp/olap_file.pb.h>
+#include <gen_cpp/segment_v2.pb.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <map>
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
 #include <vector>
 
-#include "gen_cpp/olap_file.pb.h"
-#include "gen_cpp/segment_v2.pb.h"
-#include "olap/olap_define.h"
-#include "olap/types.h"
+#include "common/status.h"
+#include "gutil/stringprintf.h"
+#include "olap/olap_common.h"
 #include "vec/aggregate_functions/aggregate_function.h"
-#include "vec/data_types/data_type.h"
 
 namespace doris {
 namespace vectorized {
@@ -32,6 +43,8 @@ class Block;
 }
 
 struct OlapTableIndexSchema;
+class TColumn;
+class TOlapTableIndex;
 
 class TabletColumn {
 public:
diff --git a/be/src/olap/task/engine_alter_tablet_task.cpp b/be/src/olap/task/engine_alter_tablet_task.cpp
index a214ad1a1b..df801a6d1b 100644
--- a/be/src/olap/task/engine_alter_tablet_task.cpp
+++ b/be/src/olap/task/engine_alter_tablet_task.cpp
@@ -17,9 +17,18 @@
 
 #include "olap/task/engine_alter_tablet_task.h"
 
+#include <fmt/format.h>
+#include <gen_cpp/AgentService_types.h>
+#include <glog/logging.h>
+
+#include <ostream>
+#include <string>
+
+#include "common/config.h"
 #include "olap/schema_change.h"
-#include "runtime/memory/mem_tracker.h"
+#include "runtime/memory/mem_tracker_limiter.h"
 #include "runtime/thread_context.h"
+#include "util/doris_metrics.h"
 
 namespace doris {
 
diff --git a/be/src/olap/task/engine_alter_tablet_task.h b/be/src/olap/task/engine_alter_tablet_task.h
index 8f7ba64b55..0c7851c98c 100644
--- a/be/src/olap/task/engine_alter_tablet_task.h
+++ b/be/src/olap/task/engine_alter_tablet_task.h
@@ -17,11 +17,15 @@
 
 #pragma once
 
-#include "gen_cpp/AgentService_types.h"
-#include "olap/olap_define.h"
+#include <memory>
+
+#include "common/status.h"
 #include "olap/task/engine_task.h"
 
 namespace doris {
+class MemTrackerLimiter;
+class TAlterInvertedIndexReq;
+class TAlterTabletReqV2;
 
 // base class for storage engine
 // add "Engine" as task prefix to prevent duplicate name with agent task
diff --git a/be/src/olap/task/engine_batch_load_task.cpp b/be/src/olap/task/engine_batch_load_task.cpp
index cd8b73717d..afe3d0b366 100644
--- a/be/src/olap/task/engine_batch_load_task.cpp
+++ b/be/src/olap/task/engine_batch_load_task.cpp
@@ -17,27 +17,42 @@
 
 #include "olap/task/engine_batch_load_task.h"
 
+#include <fmt/format.h>
+#include <gen_cpp/AgentService_types.h>
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/Types_types.h>
 #include <pthread.h>
 #include <thrift/protocol/TDebugProtocol.h>
 
 #include <cstdio>
 #include <ctime>
+#include <filesystem>
 #include <fstream>
-#include <iostream>
-#include <sstream>
+#include <list>
 #include <string>
+#include <system_error>
 
 #include "boost/lexical_cast.hpp"
-#include "gen_cpp/AgentService_types.h"
+#include "common/config.h"
+#include "common/logging.h"
 #include "http/http_client.h"
+#include "olap/data_dir.h"
 #include "olap/olap_common.h"
 #include "olap/olap_define.h"
 #include "olap/push_handler.h"
 #include "olap/storage_engine.h"
 #include "olap/tablet.h"
+#include "olap/tablet_manager.h"
+#include "runtime/memory/mem_tracker_limiter.h"
 #include "runtime/thread_context.h"
 #include "util/doris_metrics.h"
 #include "util/pretty_printer.h"
+#include "util/runtime_profile.h"
+#include "util/stopwatch.hpp"
+
+namespace doris {
+class TTabletInfo;
+} // namespace doris
 
 using apache::thrift::ThriftDebugString;
 using std::list;
diff --git a/be/src/olap/task/engine_batch_load_task.h b/be/src/olap/task/engine_batch_load_task.h
index e314b6faf7..036b49e26d 100644
--- a/be/src/olap/task/engine_batch_load_task.h
+++ b/be/src/olap/task/engine_batch_load_task.h
@@ -18,22 +18,23 @@
 #ifndef DORIS_BE_SRC_OLAP_TASK_ENGINE_BATCH_LOAD_TASK_H
 #define DORIS_BE_SRC_OLAP_TASK_ENGINE_BATCH_LOAD_TASK_H
 
-#include <utility>
+#include <stdint.h>
+
+#include <memory>
+#include <string>
 #include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/AgentService_types.h"
-#include "gen_cpp/MasterService_types.h"
-#include "olap/olap_common.h"
-#include "olap/olap_define.h"
 #include "olap/task/engine_task.h"
 
 namespace doris {
+class MemTrackerLimiter;
+class TPushReq;
+class TTabletInfo;
 
 const uint32_t PUSH_MAX_RETRY = 1;
 const uint32_t MAX_RETRY = 3;
 const uint32_t DEFAULT_DOWNLOAD_TIMEOUT = 3600;
-class StorageEngine;
 
 class EngineBatchLoadTask : public EngineTask {
 public:
diff --git a/be/src/olap/task/engine_checksum_task.cpp b/be/src/olap/task/engine_checksum_task.cpp
index 1d6c3e52a3..fd98691415 100644
--- a/be/src/olap/task/engine_checksum_task.cpp
+++ b/be/src/olap/task/engine_checksum_task.cpp
@@ -17,7 +17,26 @@
 
 #include "olap/task/engine_checksum_task.h"
 
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <ostream>
+#include <string>
+#include <vector>
+
+#include "io/io_common.h"
+#include "olap/olap_common.h"
+#include "olap/olap_define.h"
+#include "olap/reader.h"
+#include "olap/rowset/rowset.h"
+#include "olap/storage_engine.h"
+#include "olap/tablet.h"
+#include "olap/tablet_manager.h"
+#include "olap/utils.h"
+#include "runtime/memory/mem_tracker_limiter.h"
 #include "runtime/thread_context.h"
+#include "vec/common/sip_hash.h"
+#include "vec/core/block.h"
 #include "vec/olap/block_reader.h"
 
 namespace doris {
diff --git a/be/src/olap/task/engine_checksum_task.h b/be/src/olap/task/engine_checksum_task.h
index 04afa1a5cd..58daba515c 100644
--- a/be/src/olap/task/engine_checksum_task.h
+++ b/be/src/olap/task/engine_checksum_task.h
@@ -18,11 +18,16 @@
 #ifndef DORIS_BE_SRC_OLAP_TASK_ENGINE_CHECKSUM_TASK_H
 #define DORIS_BE_SRC_OLAP_TASK_ENGINE_CHECKSUM_TASK_H
 
-#include "gen_cpp/AgentService_types.h"
-#include "olap/olap_define.h"
+#include <gen_cpp/Types_types.h>
+#include <stdint.h>
+
+#include <memory>
+
+#include "common/status.h"
 #include "olap/task/engine_task.h"
 
 namespace doris {
+class MemTrackerLimiter;
 
 // base class for storage engine
 // add "Engine" as task prefix to prevent duplicate name with agent task
diff --git a/be/src/olap/task/engine_clone_task.cpp b/be/src/olap/task/engine_clone_task.cpp
index 1689d54512..dfb18ee376 100644
--- a/be/src/olap/task/engine_clone_task.cpp
+++ b/be/src/olap/task/engine_clone_task.cpp
@@ -17,25 +17,51 @@
 
 #include "olap/task/engine_clone_task.h"
 
-#include <memory>
-#include <set>
-#include <system_error>
+#include <fcntl.h>
+#include <fmt/format.h>
+#include <gen_cpp/AgentService_types.h>
+#include <gen_cpp/BackendService.h>
+#include <gen_cpp/HeartbeatService_types.h>
+#include <gen_cpp/MasterService_types.h>
+#include <gen_cpp/Status_types.h>
+#include <gen_cpp/Types_constants.h>
+#include <sys/stat.h>
 
-#include "gen_cpp/BackendService.h"
-#include "gen_cpp/Types_constants.h"
+#include <filesystem>
+#include <memory>
+#include <mutex>
+#include <ostream>
+#include <set>
+#include <shared_mutex>
+#include <system_error>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
+
+#include "common/config.h"
+#include "common/logging.h"
 #include "gutil/strings/split.h"
 #include "gutil/strings/stringpiece.h"
+#include "gutil/strings/strip.h"
 #include "http/http_client.h"
+#include "io/fs/file_system.h"
 #include "io/fs/local_file_system.h"
+#include "io/fs/path.h"
+#include "olap/data_dir.h"
+#include "olap/olap_common.h"
+#include "olap/olap_define.h"
 #include "olap/rowset/rowset.h"
-#include "olap/rowset/rowset_factory.h"
 #include "olap/snapshot_manager.h"
 #include "olap/storage_engine.h"
+#include "olap/tablet.h"
+#include "olap/tablet_manager.h"
 #include "olap/tablet_meta.h"
 #include "runtime/client_cache.h"
+#include "runtime/memory/mem_tracker_limiter.h"
 #include "runtime/thread_context.h"
 #include "util/defer_op.h"
 #include "util/network_util.h"
+#include "util/stopwatch.hpp"
 #include "util/thrift_rpc_helper.h"
 
 using std::set;
diff --git a/be/src/olap/task/engine_clone_task.h b/be/src/olap/task/engine_clone_task.h
index 67a164b6c2..b3a055a500 100644
--- a/be/src/olap/task/engine_clone_task.h
+++ b/be/src/olap/task/engine_clone_task.h
@@ -18,15 +18,26 @@
 #ifndef DORIS_BE_SRC_OLAP_TASK_ENGINE_CLONE_TASK_H
 #define DORIS_BE_SRC_OLAP_TASK_ENGINE_CLONE_TASK_H
 
-#include "agent/utils.h"
+#include <gen_cpp/Types_types.h>
+#include <stdint.h>
+
+#include <memory>
+#include <string>
+#include <vector>
+
 #include "common/status.h"
-#include "gen_cpp/AgentService_types.h"
-#include "gen_cpp/HeartbeatService.h"
-#include "gen_cpp/MasterService_types.h"
-#include "olap/olap_define.h"
+#include "gutil/strings/substitute.h"
+#include "olap/tablet_meta.h"
 #include "olap/task/engine_task.h"
 
 namespace doris {
+class DataDir;
+class MemTrackerLimiter;
+class TCloneReq;
+class TMasterInfo;
+class TTabletInfo;
+class Tablet;
+struct Version;
 
 // base class for storage engine
 // add "Engine" as task prefix to prevent duplicate name with agent task
diff --git a/be/src/olap/task/engine_publish_version_task.cpp b/be/src/olap/task/engine_publish_version_task.cpp
index 8fd438354a..91384ea89c 100644
--- a/be/src/olap/task/engine_publish_version_task.cpp
+++ b/be/src/olap/task/engine_publish_version_task.cpp
@@ -17,13 +17,26 @@
 
 #include "olap/task/engine_publish_version_task.h"
 
+#include <gen_cpp/AgentService_types.h>
+#include <gen_cpp/olap_file.pb.h>
 #include <util/defer_op.h>
-
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
 #include <map>
+#include <memory>
+#include <ostream>
+#include <set>
+#include <shared_mutex>
+#include <string>
+#include <utility>
 
-#include "olap/data_dir.h"
-#include "olap/rowset/rowset_meta_manager.h"
+#include "common/logging.h"
+#include "olap/storage_engine.h"
 #include "olap/tablet_manager.h"
+#include "olap/tablet_meta.h"
+#include "olap/txn_manager.h"
+#include "olap/utils.h"
+#include "util/threadpool.h"
 
 namespace doris {
 
diff --git a/be/src/olap/task/engine_publish_version_task.h b/be/src/olap/task/engine_publish_version_task.h
index 959584d116..daed805518 100644
--- a/be/src/olap/task/engine_publish_version_task.h
+++ b/be/src/olap/task/engine_publish_version_task.h
@@ -18,13 +18,25 @@
 #ifndef DORIS_BE_SRC_OLAP_TASK_ENGINE_PUBLISH_VERSION_TASK_H
 #define DORIS_BE_SRC_OLAP_TASK_ENGINE_PUBLISH_VERSION_TASK_H
 
-#include "gen_cpp/AgentService_types.h"
-#include "olap/olap_define.h"
+#include <gen_cpp/Types_types.h>
+#include <stdint.h>
+
+#include <atomic>
+#include <condition_variable>
+#include <mutex>
+#include <vector>
+
+#include "common/status.h"
+#include "olap/olap_common.h"
+#include "olap/rowset/rowset.h"
+#include "olap/tablet.h"
 #include "olap/task/engine_task.h"
 
 namespace doris {
 
 class EnginePublishVersionTask;
+class TPublishVersionRequest;
+
 class TabletPublishTxnTask {
 public:
     TabletPublishTxnTask(EnginePublishVersionTask* engine_task, TabletSharedPtr tablet,
diff --git a/be/src/olap/task/engine_storage_migration_task.cpp b/be/src/olap/task/engine_storage_migration_task.cpp
index 43c593b509..7bf512d407 100644
--- a/be/src/olap/task/engine_storage_migration_task.cpp
+++ b/be/src/olap/task/engine_storage_migration_task.cpp
@@ -17,10 +17,32 @@
 
 #include "olap/task/engine_storage_migration_task.h"
 
-#include <ctime>
+#include <fmt/format.h>
+#include <gen_cpp/olap_file.pb.h>
+#include <unistd.h>
 
+#include <algorithm>
+#include <ctime>
+#include <memory>
+#include <new>
+#include <ostream>
+#include <set>
+#include <utility>
+
+#include "common/config.h"
+#include "common/logging.h"
+#include "gutil/strings/numbers.h"
+#include "io/fs/local_file_system.h"
+#include "olap/data_dir.h"
+#include "olap/olap_common.h"
+#include "olap/olap_define.h"
+#include "olap/rowset/rowset_meta.h"
 #include "olap/snapshot_manager.h"
-#include "olap/tablet_meta_manager.h"
+#include "olap/storage_engine.h"
+#include "olap/tablet_manager.h"
+#include "olap/txn_manager.h"
+#include "util/doris_metrics.h"
+#include "util/uid_util.h"
 
 namespace doris {
 
diff --git a/be/src/olap/task/engine_storage_migration_task.h b/be/src/olap/task/engine_storage_migration_task.h
index c2c9b7a774..d392de9cf2 100644
--- a/be/src/olap/task/engine_storage_migration_task.h
+++ b/be/src/olap/task/engine_storage_migration_task.h
@@ -18,11 +18,21 @@
 #ifndef DORIS_BE_SRC_OLAP_TASK_ENGINE_STORAGE_MIGRATION_TASK_H
 #define DORIS_BE_SRC_OLAP_TASK_ENGINE_STORAGE_MIGRATION_TASK_H
 
-#include "gen_cpp/AgentService_types.h"
-#include "olap/olap_define.h"
+#include <stdint.h>
+
+#include <mutex>
+#include <shared_mutex>
+#include <string>
+#include <vector>
+
+#include "common/status.h"
+#include "olap/rowset/rowset.h"
+#include "olap/tablet.h"
+#include "olap/tablet_meta.h"
 #include "olap/task/engine_task.h"
 
 namespace doris {
+class DataDir;
 
 /// This task is used to migrate the specified tablet to the specified data directory.
 // Usually used for storage medium migration, or migration of tablets between disks.
diff --git a/be/src/olap/txn_manager.cpp b/be/src/olap/txn_manager.cpp
index 474b4c8936..cfb1b676f0 100644
--- a/be/src/olap/txn_manager.cpp
+++ b/be/src/olap/txn_manager.cpp
@@ -17,39 +17,33 @@
 
 #include "txn_manager.h"
 
-#include <rapidjson/document.h>
-#include <signal.h>
 #include <thrift/protocol/TDebugProtocol.h>
+#include <time.h>
 
-#include <algorithm>
-#include <boost/algorithm/string.hpp>
-#include <boost/algorithm/string/classification.hpp>
-#include <boost/algorithm/string/split.hpp>
-#include <cstdio>
 #include <filesystem>
+#include <iterator>
+#include <list>
 #include <new>
+#include <ostream>
 #include <queue>
-#include <random>
 #include <set>
+#include <string>
 
-#include "olap/base_compaction.h"
-#include "olap/cumulative_compaction.h"
+#include "common/config.h"
+#include "common/logging.h"
 #include "olap/data_dir.h"
 #include "olap/delta_writer.h"
-#include "olap/lru_cache.h"
-#include "olap/push_handler.h"
-#include "olap/reader.h"
+#include "olap/rowset/rowset_meta.h"
 #include "olap/rowset/rowset_meta_manager.h"
-#include "olap/schema_change.h"
 #include "olap/storage_engine.h"
+#include "olap/tablet_manager.h"
 #include "olap/tablet_meta.h"
-#include "olap/tablet_meta_manager.h"
-#include "olap/utils.h"
-#include "rowset/beta_rowset.h"
-#include "util/doris_metrics.h"
-#include "util/pretty_printer.h"
 #include "util/time.h"
 
+namespace doris {
+class OlapMeta;
+} // namespace doris
+
 using apache::thrift::ThriftDebugString;
 using std::filesystem::canonical;
 using std::filesystem::directory_iterator;
@@ -64,7 +58,6 @@ using std::nothrow;
 using std::pair;
 using std::priority_queue;
 using std::set;
-using std::set_difference;
 using std::string;
 using std::stringstream;
 using std::vector;
diff --git a/be/src/olap/txn_manager.h b/be/src/olap/txn_manager.h
index 7319234b27..ad986fbe9e 100644
--- a/be/src/olap/txn_manager.h
+++ b/be/src/olap/txn_manager.h
@@ -17,38 +17,33 @@
 
 #pragma once
 
-#include <pthread.h>
-#include <rapidjson/document.h>
+#include <butil/macros.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/types.pb.h>
+#include <stddef.h>
+#include <stdint.h>
 
-#include <condition_variable>
-#include <list>
+#include <boost/container/detail/std_fwd.hpp>
 #include <map>
+#include <memory>
 #include <mutex>
 #include <set>
-#include <string>
-#include <thread>
+#include <shared_mutex>
 #include <unordered_map>
 #include <unordered_set>
+#include <utility>
 #include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/AgentService_types.h"
-#include "gen_cpp/BackendService_types.h"
-#include "gen_cpp/MasterService_types.h"
-#include "olap/lru_cache.h"
 #include "olap/olap_common.h"
-#include "olap/olap_define.h"
-#include "olap/olap_meta.h"
-#include "olap/options.h"
 #include "olap/rowset/rowset.h"
-#include "olap/rowset/rowset_meta.h"
-#include "olap/rowset/segment_v2/segment.h"
 #include "olap/tablet.h"
 #include "olap/tablet_meta.h"
 #include "util/time.h"
 
 namespace doris {
 class DeltaWriter;
+class OlapMeta;
 
 struct TabletTxnInfo {
     PUniqueId load_id;
diff --git a/be/src/olap/types.cpp b/be/src/olap/types.cpp
index da3fd4a255..e35725f07d 100644
--- a/be/src/olap/types.cpp
+++ b/be/src/olap/types.cpp
@@ -17,9 +17,12 @@
 
 #include "olap/types.h"
 
+#include <gen_cpp/segment_v2.pb.h>
+
 #include <memory>
 
-#include "gen_cpp/segment_v2.pb.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "olap/tablet_schema.h"
 
 namespace doris {
diff --git a/be/src/olap/types.h b/be/src/olap/types.h
index 6b35729551..698b6548f7 100644
--- a/be/src/olap/types.h
+++ b/be/src/olap/types.h
@@ -17,29 +17,43 @@
 
 #pragma once
 
-#include <math.h>
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <limits.h>
 #include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <time.h>
 
+#include <algorithm>
 #include <cinttypes>
 #include <limits>
 #include <memory>
+#include <new>
 #include <sstream>
 #include <string>
+#include <type_traits>
+#include <utility>
+#include <vector>
 
+#include "common/config.h"
+#include "common/status.h"
+#include "gutil/stringprintf.h"
 #include "gutil/strings/numbers.h"
+#include "olap/decimal12.h"
 #include "olap/olap_common.h"
 #include "olap/olap_define.h"
+#include "olap/uint24.h"
 #include "runtime/collection_value.h"
-#include "runtime/jsonb_value.h"
 #include "runtime/map_value.h"
 #include "runtime/struct_value.h"
-#include "util/jsonb_document.h"
-#include "util/jsonb_utils.h"
+#include "util/binary_cast.hpp"
 #include "util/mysql_global.h"
 #include "util/slice.h"
 #include "util/string_parser.hpp"
 #include "util/types.h"
 #include "vec/common/arena.h"
+#include "vec/runtime/vdatetime_value.h"
 
 namespace doris {
 
@@ -47,13 +61,12 @@ namespace segment_v2 {
 class ColumnMetaPB;
 }
 
-struct uint24_t;
-struct decimal12_t;
 class TabletColumn;
 
 extern bool is_olap_string_type(FieldType field_type);
 
 class TypeInfo;
+
 using TypeInfoPtr = std::unique_ptr<const TypeInfo, void (*)(const TypeInfo*)>;
 
 TypeInfoPtr create_static_type_info_ptr(const TypeInfo* type_info);
diff --git a/be/src/olap/utils.cpp b/be/src/olap/utils.cpp
index 1f19881aee..188543b346 100644
--- a/be/src/olap/utils.cpp
+++ b/be/src/olap/utils.cpp
@@ -17,22 +17,24 @@
 
 #include "olap/utils.h"
 
-#include <dirent.h>
-#include <errno.h>
-#include <lz4/lz4.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <sys/stat.h>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
 #include <time.h>
 #include <unistd.h>
+#include <zconf.h>
+#include <zlib.h>
 
-#include <cstdint>
+#include <cmath>
 #include <cstring>
-#include <filesystem>
+#include <memory>
 #include <regex>
+#include <set>
+#include <sstream>
 #include <string>
 #include <vector>
 
+#include "util/sse_util.hpp"
+
 #ifdef DORIS_WITH_LZO
 #include <lzo/lzo1c.h>
 #include <lzo/lzo1x.h>
@@ -42,13 +44,11 @@
 
 #include "common/logging.h"
 #include "common/status.h"
-#include "gutil/strings/substitute.h"
 #include "io/fs/file_reader.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "io/fs/file_writer.h"
 #include "io/fs/local_file_system.h"
 #include "olap/olap_common.h"
-#include "olap/olap_define.h"
-#include "util/errno.h"
 #include "util/string_parser.hpp"
 
 using std::string;
diff --git a/be/src/olap/utils.h b/be/src/olap/utils.h
index 71be36cb71..f91b625db3 100644
--- a/be/src/olap/utils.h
+++ b/be/src/olap/utils.h
@@ -17,26 +17,21 @@
 
 #pragma once
 
-#include <fcntl.h>
-#include <pthread.h>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
+#include <limits.h>
+#include <stdint.h>
 #include <sys/time.h>
-#include <zlib.h>
 
 #include <cstdio>
 #include <cstdlib>
-#include <exception>
-#include <filesystem>
 #include <iterator>
 #include <limits>
-#include <list>
-#include <set>
-#include <sstream>
 #include <string>
 #include <vector>
 
-#include "common/logging.h"
+#include "common/status.h"
 #include "olap/olap_common.h"
-#include "olap/olap_define.h"
 
 namespace doris {
 void write_log_info(char* buf, size_t buf_len, const char* fmt, ...);
diff --git a/be/src/olap/version_graph.cpp b/be/src/olap/version_graph.cpp
index bae1d289a2..0204ab783d 100644
--- a/be/src/olap/version_graph.cpp
+++ b/be/src/olap/version_graph.cpp
@@ -18,12 +18,17 @@
 #include "olap/version_graph.h"
 
 #include <cctz/time_zone.h>
+#include <stddef.h>
 
+#include <algorithm>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <list>
 #include <memory>
-#include <queue>
+#include <ostream>
+#include <utility>
 
 #include "common/logging.h"
-#include "util/time.h"
 
 namespace doris {
 using namespace ErrorCode;
diff --git a/be/src/olap/version_graph.h b/be/src/olap/version_graph.h
index 287d946fcf..25ca25cc5a 100644
--- a/be/src/olap/version_graph.h
+++ b/be/src/olap/version_graph.h
@@ -18,11 +18,16 @@
 #pragma once
 
 #include <rapidjson/document.h>
-#include <rapidjson/prettywriter.h>
-#include <rapidjson/stringbuffer.h>
+#include <stdint.h>
 
+#include <map>
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "common/status.h"
 #include "olap/olap_common.h"
-#include "olap/olap_define.h"
 #include "olap/rowset/rowset_meta.h"
 
 namespace doris {
diff --git a/be/src/olap/wrapper_field.cpp b/be/src/olap/wrapper_field.cpp
index 6f227e1775..cc5e756a31 100644
--- a/be/src/olap/wrapper_field.cpp
+++ b/be/src/olap/wrapper_field.cpp
@@ -17,6 +17,15 @@
 
 #include "olap/wrapper_field.h"
 
+#include <glog/logging.h>
+#include <string.h>
+
+#include <algorithm>
+#include <ostream>
+
+#include "common/config.h"
+#include "olap/olap_common.h"
+#include "olap/olap_define.h"
 #include "olap/row_cursor.h"
 
 namespace doris {
diff --git a/be/src/olap/wrapper_field.h b/be/src/olap/wrapper_field.h
index b8ed3badfa..71afc2f2a4 100644
--- a/be/src/olap/wrapper_field.h
+++ b/be/src/olap/wrapper_field.h
@@ -17,13 +17,20 @@
 
 #pragma once
 
+#include <stdint.h>
+#include <stdlib.h>
+
+#include <memory>
+#include <string>
+
+#include "common/status.h"
 #include "olap/field.h"
-#include "olap/olap_define.h"
 #include "olap/row_cursor_cell.h"
 #include "olap/tablet_schema.h"
-#include "util/hash_util.hpp"
+#include "util/slice.h"
 
 namespace doris {
+enum class FieldType;
 
 class WrapperField {
 public:
diff --git a/be/src/pipeline/exec/aggregation_sink_operator.cpp b/be/src/pipeline/exec/aggregation_sink_operator.cpp
index ce53e4d6a4..70ca1219d4 100644
--- a/be/src/pipeline/exec/aggregation_sink_operator.cpp
+++ b/be/src/pipeline/exec/aggregation_sink_operator.cpp
@@ -17,7 +17,9 @@
 
 #include "aggregation_sink_operator.h"
 
-#include "vec/exec/vaggregation_node.h"
+#include <string>
+
+#include "pipeline/exec/operator.h"
 
 namespace doris::pipeline {
 
diff --git a/be/src/pipeline/exec/aggregation_sink_operator.h b/be/src/pipeline/exec/aggregation_sink_operator.h
index 23115fbe19..3bfc783ccc 100644
--- a/be/src/pipeline/exec/aggregation_sink_operator.h
+++ b/be/src/pipeline/exec/aggregation_sink_operator.h
@@ -17,14 +17,13 @@
 
 #pragma once
 
+#include <stdint.h>
+
 #include "operator.h"
+#include "vec/exec/vaggregation_node.h"
 
 namespace doris {
-namespace vectorized {
-class AggregationNode;
-class VExprContext;
-class Block;
-} // namespace vectorized
+class ExecNode;
 
 namespace pipeline {
 
diff --git a/be/src/pipeline/exec/aggregation_source_operator.cpp b/be/src/pipeline/exec/aggregation_source_operator.cpp
index e3b2aa1f23..80a974c78d 100644
--- a/be/src/pipeline/exec/aggregation_source_operator.cpp
+++ b/be/src/pipeline/exec/aggregation_source_operator.cpp
@@ -17,7 +17,9 @@
 
 #include "aggregation_source_operator.h"
 
-#include "vec/exec/vaggregation_node.h"
+#include <string>
+
+#include "pipeline/exec/operator.h"
 
 namespace doris {
 namespace pipeline {
diff --git a/be/src/pipeline/exec/aggregation_source_operator.h b/be/src/pipeline/exec/aggregation_source_operator.h
index f1c93071da..19c64bc065 100644
--- a/be/src/pipeline/exec/aggregation_source_operator.h
+++ b/be/src/pipeline/exec/aggregation_source_operator.h
@@ -16,12 +16,15 @@
 // under the License.
 #pragma once
 
+#include <stdint.h>
+
+#include "common/status.h"
 #include "operator.h"
+#include "vec/exec/vaggregation_node.h"
 
 namespace doris {
-namespace vectorized {
-class AggregationNode;
-}
+class ExecNode;
+class RuntimeState;
 
 namespace pipeline {
 
diff --git a/be/src/pipeline/exec/analytic_sink_operator.cpp b/be/src/pipeline/exec/analytic_sink_operator.cpp
index 4a17359ea7..a0634c81f6 100644
--- a/be/src/pipeline/exec/analytic_sink_operator.cpp
+++ b/be/src/pipeline/exec/analytic_sink_operator.cpp
@@ -18,6 +18,10 @@
 
 #include "analytic_sink_operator.h"
 
+#include <string>
+
+#include "pipeline/exec/operator.h"
+
 namespace doris::pipeline {
 
 OPERATOR_CODE_GENERATOR(AnalyticSinkOperator, StreamingOperator)
diff --git a/be/src/pipeline/exec/analytic_sink_operator.h b/be/src/pipeline/exec/analytic_sink_operator.h
index 793824ea75..2d16eb1c70 100644
--- a/be/src/pipeline/exec/analytic_sink_operator.h
+++ b/be/src/pipeline/exec/analytic_sink_operator.h
@@ -18,13 +18,13 @@
 
 #pragma once
 
+#include <stdint.h>
+
 #include "operator.h"
 #include "vec/exec/vanalytic_eval_node.h"
 
 namespace doris {
-namespace vectorized {
-class VAnalyticEvalNode;
-} // namespace vectorized
+class ExecNode;
 
 namespace pipeline {
 class AnalyticSinkOperatorBuilder final : public OperatorBuilder<vectorized::VAnalyticEvalNode> {
diff --git a/be/src/pipeline/exec/analytic_source_operator.cpp b/be/src/pipeline/exec/analytic_source_operator.cpp
index 9c9717d24a..20a4bfb9cc 100644
--- a/be/src/pipeline/exec/analytic_source_operator.cpp
+++ b/be/src/pipeline/exec/analytic_source_operator.cpp
@@ -17,6 +17,10 @@
 
 #include "analytic_source_operator.h"
 
+#include <string>
+
+#include "pipeline/exec/operator.h"
+
 namespace doris::pipeline {
 
 OPERATOR_CODE_GENERATOR(AnalyticSourceOperator, SourceOperator)
diff --git a/be/src/pipeline/exec/analytic_source_operator.h b/be/src/pipeline/exec/analytic_source_operator.h
index 34fd9cdfde..60f1a86158 100644
--- a/be/src/pipeline/exec/analytic_source_operator.h
+++ b/be/src/pipeline/exec/analytic_source_operator.h
@@ -17,14 +17,15 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include "common/status.h"
 #include "operator.h"
 #include "vec/exec/vanalytic_eval_node.h"
 
 namespace doris {
-
-namespace vectorized {
-class VAnalyticEvalNode;
-}
+class ExecNode;
+class RuntimeState;
 
 namespace pipeline {
 
diff --git a/be/src/pipeline/exec/data_queue.cpp b/be/src/pipeline/exec/data_queue.cpp
index a6f2c43939..42a1fbf38e 100644
--- a/be/src/pipeline/exec/data_queue.cpp
+++ b/be/src/pipeline/exec/data_queue.cpp
@@ -17,14 +17,16 @@
 
 #include "data_queue.h"
 
-#include <mutex>
+#include <glog/logging.h>
 
+#include <algorithm>
+#include <mutex>
+#include <utility>
+
+#include "gutil/integral_types.h"
 #include "vec/core/block.h"
 
 namespace doris {
-namespace vectorized {
-class Block;
-}
 namespace pipeline {
 
 DataQueue::DataQueue(int child_count) {
diff --git a/be/src/pipeline/exec/data_queue.h b/be/src/pipeline/exec/data_queue.h
index 978fe6e629..299f7e6b4d 100644
--- a/be/src/pipeline/exec/data_queue.h
+++ b/be/src/pipeline/exec/data_queue.h
@@ -16,17 +16,18 @@
 // under the License.
 #pragma once
 
+#include <stdint.h>
+
 #include <atomic>
 #include <deque>
 #include <memory>
 #include <mutex>
 #include <vector>
 
+#include "common/status.h"
 #include "vec/core/block.h"
+
 namespace doris {
-namespace vectorized {
-class Block;
-}
 namespace pipeline {
 
 class DataQueue {
diff --git a/be/src/pipeline/exec/datagen_operator.cpp b/be/src/pipeline/exec/datagen_operator.cpp
index fc4a32342c..7b5cf26032 100644
--- a/be/src/pipeline/exec/datagen_operator.cpp
+++ b/be/src/pipeline/exec/datagen_operator.cpp
@@ -17,8 +17,16 @@
 
 #include "datagen_operator.h"
 
+#include <memory>
+
+#include "pipeline/exec/operator.h"
+#include "util/runtime_profile.h"
 #include "vec/exec/vdata_gen_scan_node.h"
 
+namespace doris {
+class RuntimeState;
+} // namespace doris
+
 namespace doris::pipeline {
 
 OPERATOR_CODE_GENERATOR(DataGenOperator, SourceOperator)
diff --git a/be/src/pipeline/exec/datagen_operator.h b/be/src/pipeline/exec/datagen_operator.h
index 7cea4d54df..48d525e80b 100644
--- a/be/src/pipeline/exec/datagen_operator.h
+++ b/be/src/pipeline/exec/datagen_operator.h
@@ -17,11 +17,16 @@
 
 #pragma once
 
-#include "operator.h"
+#include <stdint.h>
 
-namespace doris::vectorized {
-class VDataGenFunctionScanNode;
-} // namespace doris::vectorized
+#include "common/status.h"
+#include "operator.h"
+#include "vec/exec/vdata_gen_scan_node.h"
+
+namespace doris {
+class ExecNode;
+class RuntimeState;
+} // namespace doris
 
 namespace doris::pipeline {
 
diff --git a/be/src/pipeline/exec/empty_set_operator.cpp b/be/src/pipeline/exec/empty_set_operator.cpp
index be1178e9a6..cc244e1426 100644
--- a/be/src/pipeline/exec/empty_set_operator.cpp
+++ b/be/src/pipeline/exec/empty_set_operator.cpp
@@ -17,7 +17,9 @@
 
 #include "empty_set_operator.h"
 
-#include "vec/exec/vempty_set_node.h"
+#include <string>
+
+#include "pipeline/exec/operator.h"
 
 namespace doris::pipeline {
 
diff --git a/be/src/pipeline/exec/empty_set_operator.h b/be/src/pipeline/exec/empty_set_operator.h
index a6324ca249..67cf23a230 100644
--- a/be/src/pipeline/exec/empty_set_operator.h
+++ b/be/src/pipeline/exec/empty_set_operator.h
@@ -17,10 +17,13 @@
 
 #pragma once
 
+#include <stdint.h>
+
 #include "operator.h"
 #include "vec/exec/vempty_set_node.h"
 
 namespace doris {
+class ExecNode;
 
 namespace pipeline {
 
diff --git a/be/src/pipeline/exec/empty_source_operator.cpp b/be/src/pipeline/exec/empty_source_operator.cpp
index b5b140305b..5142c0c55a 100644
--- a/be/src/pipeline/exec/empty_source_operator.cpp
+++ b/be/src/pipeline/exec/empty_source_operator.cpp
@@ -17,6 +17,8 @@
 
 #include "empty_source_operator.h"
 
+#include "pipeline/exec/operator.h"
+
 namespace doris::pipeline {
 OperatorPtr EmptySourceOperatorBuilder::build_operator() {
     return std::make_shared<EmptySourceOperator>(this);
diff --git a/be/src/pipeline/exec/empty_source_operator.h b/be/src/pipeline/exec/empty_source_operator.h
index 0cfa3bcbf1..fd12b27dab 100644
--- a/be/src/pipeline/exec/empty_source_operator.h
+++ b/be/src/pipeline/exec/empty_source_operator.h
@@ -17,11 +17,21 @@
 
 #pragma once
 
-#include "operator.h"
+#include <stdint.h>
 
-namespace doris::vectorized {
-class VExchangeNode;
-}
+#include <memory>
+
+#include "common/status.h"
+#include "operator.h"
+#include "runtime/descriptors.h"
+
+namespace doris {
+class RuntimeState;
+
+namespace vectorized {
+class Block;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::pipeline {
 
diff --git a/be/src/pipeline/exec/exchange_sink_buffer.cpp b/be/src/pipeline/exec/exchange_sink_buffer.cpp
index da536d66ec..8f6927640b 100644
--- a/be/src/pipeline/exec/exchange_sink_buffer.cpp
+++ b/be/src/pipeline/exec/exchange_sink_buffer.cpp
@@ -17,14 +17,28 @@
 
 #include "exchange_sink_buffer.h"
 
-#include <google/protobuf/stubs/common.h>
+#include <brpc/controller.h>
+#include <butil/errno.h>
+#include <butil/iobuf_inl.h>
+#include <fmt/format.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/internal_service.pb.h>
+#include <glog/logging.h>
+#include <google/protobuf/stubs/callback.h>
+#include <stddef.h>
 
 #include <atomic>
+#include <exception>
+#include <functional>
 #include <memory>
+#include <ostream>
+#include <utility>
 
 #include "common/status.h"
 #include "pipeline/pipeline_fragment_context.h"
-#include "service/brpc.h"
+#include "runtime/exec_env.h"
+#include "runtime/thread_context.h"
+#include "service/backend_options.h"
 #include "util/proto_util.h"
 #include "vec/sink/vdata_stream_sender.h"
 
diff --git a/be/src/pipeline/exec/exchange_sink_buffer.h b/be/src/pipeline/exec/exchange_sink_buffer.h
index 783457c165..17093bca87 100644
--- a/be/src/pipeline/exec/exchange_sink_buffer.h
+++ b/be/src/pipeline/exec/exchange_sink_buffer.h
@@ -17,18 +17,25 @@
 
 #pragma once
 
+#include <gen_cpp/data.pb.h>
+#include <gen_cpp/types.pb.h>
 #include <parallel_hashmap/phmap.h>
+#include <stdint.h>
 
+#include <atomic>
 #include <list>
+#include <memory>
+#include <mutex>
 #include <queue>
-#include <shared_mutex>
+#include <string>
 
 #include "common/global_types.h"
 #include "common/status.h"
-#include "gen_cpp/Types_types.h"
-#include "gen_cpp/internal_service.pb.h"
 
 namespace doris {
+class PTransmitDataParams;
+class TUniqueId;
+
 namespace vectorized {
 class PipChannel;
 class BroadcastPBlockHolder;
diff --git a/be/src/pipeline/exec/exchange_sink_operator.cpp b/be/src/pipeline/exec/exchange_sink_operator.cpp
index 43242c8f3e..ab7a1ac07c 100644
--- a/be/src/pipeline/exec/exchange_sink_operator.cpp
+++ b/be/src/pipeline/exec/exchange_sink_operator.cpp
@@ -17,13 +17,20 @@
 
 #include "exchange_sink_operator.h"
 
+#include <gen_cpp/DataSinks_types.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/types.pb.h>
+
 #include "common/status.h"
 #include "exchange_sink_buffer.h"
-#include "gen_cpp/internal_service.pb.h"
-#include "util/brpc_client_cache.h"
-#include "vec/exprs/vexpr.h"
+#include "pipeline/exec/operator.h"
+#include "runtime/runtime_state.h"
 #include "vec/sink/vdata_stream_sender.h"
 
+namespace doris {
+class DataSink;
+} // namespace doris
+
 namespace doris::pipeline {
 
 ExchangeSinkOperatorBuilder::ExchangeSinkOperatorBuilder(int32_t id, DataSink* sink,
diff --git a/be/src/pipeline/exec/exchange_sink_operator.h b/be/src/pipeline/exec/exchange_sink_operator.h
index c35e30f49b..034ceeffa2 100644
--- a/be/src/pipeline/exec/exchange_sink_operator.h
+++ b/be/src/pipeline/exec/exchange_sink_operator.h
@@ -17,11 +17,19 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include <memory>
+
+#include "common/status.h"
 #include "exchange_sink_buffer.h"
 #include "operator.h"
 #include "vec/sink/vdata_stream_sender.h"
 
 namespace doris {
+class DataSink;
+class RuntimeState;
+class TDataSink;
 
 namespace pipeline {
 class PipelineFragmentContext;
diff --git a/be/src/pipeline/exec/exchange_source_operator.cpp b/be/src/pipeline/exec/exchange_source_operator.cpp
index 88df3c57c7..1dd20f929c 100644
--- a/be/src/pipeline/exec/exchange_source_operator.cpp
+++ b/be/src/pipeline/exec/exchange_source_operator.cpp
@@ -17,7 +17,9 @@
 
 #include "exchange_source_operator.h"
 
-#include "common/status.h"
+#include <memory>
+
+#include "pipeline/exec/operator.h"
 #include "vec/exec/vexchange_node.h"
 #include "vec/runtime/vdata_stream_recvr.h"
 
diff --git a/be/src/pipeline/exec/exchange_source_operator.h b/be/src/pipeline/exec/exchange_source_operator.h
index b6143a86ab..f80685f396 100644
--- a/be/src/pipeline/exec/exchange_source_operator.h
+++ b/be/src/pipeline/exec/exchange_source_operator.h
@@ -17,11 +17,14 @@
 
 #pragma once
 
-#include "operator.h"
+#include <stdint.h>
 
-namespace doris::vectorized {
-class VExchangeNode;
-}
+#include "operator.h"
+#include "vec/exec/vexchange_node.h"
+
+namespace doris {
+class ExecNode;
+} // namespace doris
 
 namespace doris::pipeline {
 
diff --git a/be/src/pipeline/exec/hashjoin_build_sink.cpp b/be/src/pipeline/exec/hashjoin_build_sink.cpp
index 73fdc69252..55aa25397c 100644
--- a/be/src/pipeline/exec/hashjoin_build_sink.cpp
+++ b/be/src/pipeline/exec/hashjoin_build_sink.cpp
@@ -17,7 +17,9 @@
 
 #include "hashjoin_build_sink.h"
 
-#include "vec/exec/join/vhash_join_node.h"
+#include <string>
+
+#include "pipeline/exec/operator.h"
 
 namespace doris::pipeline {
 
diff --git a/be/src/pipeline/exec/hashjoin_build_sink.h b/be/src/pipeline/exec/hashjoin_build_sink.h
index 71579c2488..dfa077b83f 100644
--- a/be/src/pipeline/exec/hashjoin_build_sink.h
+++ b/be/src/pipeline/exec/hashjoin_build_sink.h
@@ -17,14 +17,13 @@
 
 #pragma once
 
+#include <stdint.h>
+
 #include "operator.h"
+#include "vec/exec/join/vhash_join_node.h"
 
 namespace doris {
-namespace vectorized {
-class HashJoinNode;
-class VExprContext;
-class Block;
-} // namespace vectorized
+class ExecNode;
 
 namespace pipeline {
 
diff --git a/be/src/pipeline/exec/hashjoin_probe_operator.cpp b/be/src/pipeline/exec/hashjoin_probe_operator.cpp
index 91cef6d915..e499d391ce 100644
--- a/be/src/pipeline/exec/hashjoin_probe_operator.cpp
+++ b/be/src/pipeline/exec/hashjoin_probe_operator.cpp
@@ -17,7 +17,9 @@
 
 #include "hashjoin_probe_operator.h"
 
-#include "vec/exec/join/vhash_join_node.h"
+#include <string>
+
+#include "pipeline/exec/operator.h"
 
 namespace doris {
 namespace pipeline {
diff --git a/be/src/pipeline/exec/hashjoin_probe_operator.h b/be/src/pipeline/exec/hashjoin_probe_operator.h
index 54105e7f8b..b59a6435bd 100644
--- a/be/src/pipeline/exec/hashjoin_probe_operator.h
+++ b/be/src/pipeline/exec/hashjoin_probe_operator.h
@@ -16,12 +16,15 @@
 // under the License.
 #pragma once
 
+#include <stdint.h>
+
+#include "common/status.h"
 #include "operator.h"
+#include "vec/exec/join/vhash_join_node.h"
 
 namespace doris {
-namespace vectorized {
-class HashJoinNode;
-}
+class ExecNode;
+class RuntimeState;
 
 namespace pipeline {
 
diff --git a/be/src/pipeline/exec/nested_loop_join_build_operator.cpp b/be/src/pipeline/exec/nested_loop_join_build_operator.cpp
index 5e3245b983..c49e070ef1 100644
--- a/be/src/pipeline/exec/nested_loop_join_build_operator.cpp
+++ b/be/src/pipeline/exec/nested_loop_join_build_operator.cpp
@@ -17,7 +17,9 @@
 
 #include "nested_loop_join_build_operator.h"
 
-#include "vec/exec/join/vnested_loop_join_node.h"
+#include <string>
+
+#include "pipeline/exec/operator.h"
 
 namespace doris::pipeline {
 
diff --git a/be/src/pipeline/exec/nested_loop_join_build_operator.h b/be/src/pipeline/exec/nested_loop_join_build_operator.h
index 32bd55d8cf..33918f0a0c 100644
--- a/be/src/pipeline/exec/nested_loop_join_build_operator.h
+++ b/be/src/pipeline/exec/nested_loop_join_build_operator.h
@@ -17,14 +17,13 @@
 
 #pragma once
 
+#include <stdint.h>
+
 #include "operator.h"
+#include "vec/exec/join/vnested_loop_join_node.h"
 
 namespace doris {
-namespace vectorized {
-class VNestedLoopJoinNode;
-class VExprContext;
-class Block;
-} // namespace vectorized
+class ExecNode;
 
 namespace pipeline {
 
diff --git a/be/src/pipeline/exec/nested_loop_join_probe_operator.cpp b/be/src/pipeline/exec/nested_loop_join_probe_operator.cpp
index b839de0a98..cdc3630433 100644
--- a/be/src/pipeline/exec/nested_loop_join_probe_operator.cpp
+++ b/be/src/pipeline/exec/nested_loop_join_probe_operator.cpp
@@ -17,8 +17,16 @@
 
 #include "nested_loop_join_probe_operator.h"
 
+#include <memory>
+
+#include "pipeline/exec/operator.h"
+#include "vec/core/block.h"
 #include "vec/exec/join/vnested_loop_join_node.h"
 
+namespace doris {
+class RuntimeState;
+} // namespace doris
+
 namespace doris::pipeline {
 
 OPERATOR_CODE_GENERATOR(NestLoopJoinProbeOperator, StatefulOperator)
diff --git a/be/src/pipeline/exec/nested_loop_join_probe_operator.h b/be/src/pipeline/exec/nested_loop_join_probe_operator.h
index bcbf08111c..8fe9b1f883 100644
--- a/be/src/pipeline/exec/nested_loop_join_probe_operator.h
+++ b/be/src/pipeline/exec/nested_loop_join_probe_operator.h
@@ -17,14 +17,16 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include "common/status.h"
 #include "operator.h"
+#include "vec/exec/join/vnested_loop_join_node.h"
 
 namespace doris {
-namespace vectorized {
-class VNestedLoopJoinNode;
-class VExprContext;
-class Block;
-} // namespace vectorized
+class ExecNode;
+class RuntimeState;
+
 namespace pipeline {
 
 class NestLoopJoinProbeOperatorBuilder final
diff --git a/be/src/pipeline/exec/operator.cpp b/be/src/pipeline/exec/operator.cpp
index 39cb8b8814..3fd4e34eba 100644
--- a/be/src/pipeline/exec/operator.cpp
+++ b/be/src/pipeline/exec/operator.cpp
@@ -17,6 +17,11 @@
 
 #include "operator.h"
 
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+} // namespace doris
+
 namespace doris::pipeline {
 
 OperatorBase::OperatorBase(OperatorBuilderBase* operator_builder)
diff --git a/be/src/pipeline/exec/operator.h b/be/src/pipeline/exec/operator.h
index de02d4c73e..15afe49ef6 100644
--- a/be/src/pipeline/exec/operator.h
+++ b/be/src/pipeline/exec/operator.h
@@ -17,16 +17,30 @@
 
 #pragma once
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stdint.h>
+
+#include <functional>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <type_traits>
 #include <utility>
+#include <vector>
 
 #include "common/status.h"
-#include "exec/data_sink.h"
 #include "exec/exec_node.h"
-#include "runtime/runtime_state.h"
+#include "runtime/memory/mem_tracker.h"
+#include "util/runtime_profile.h"
 #include "vec/core/block.h"
-#include "vec/exec/vdata_gen_scan_node.h"
-#include "vec/exec/vselect_node.h"
-#include "vec/exec/vunion_node.h"
+
+namespace doris {
+class DataSink;
+class RowDescriptor;
+class RuntimeState;
+class TDataSink;
+} // namespace doris
 
 #define OPERATOR_CODE_GENERATOR(NAME, SUBCLASS)                                                 \
     NAME##Builder::NAME##Builder(int32_t id, ExecNode* exec_node)                               \
diff --git a/be/src/pipeline/exec/repeat_operator.cpp b/be/src/pipeline/exec/repeat_operator.cpp
index d2c9f0a1e2..3c0f819acb 100644
--- a/be/src/pipeline/exec/repeat_operator.cpp
+++ b/be/src/pipeline/exec/repeat_operator.cpp
@@ -17,8 +17,16 @@
 
 #include "repeat_operator.h"
 
+#include <memory>
+
+#include "pipeline/exec/operator.h"
+#include "vec/core/block.h"
 #include "vec/exec/vrepeat_node.h"
 
+namespace doris {
+class RuntimeState;
+} // namespace doris
+
 namespace doris::pipeline {
 
 OPERATOR_CODE_GENERATOR(RepeatOperator, StatefulOperator)
diff --git a/be/src/pipeline/exec/repeat_operator.h b/be/src/pipeline/exec/repeat_operator.h
index b397ea05d3..ed982ad574 100644
--- a/be/src/pipeline/exec/repeat_operator.h
+++ b/be/src/pipeline/exec/repeat_operator.h
@@ -17,14 +17,16 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include "common/status.h"
 #include "operator.h"
+#include "vec/exec/vrepeat_node.h"
 
 namespace doris {
-namespace vectorized {
-class VRepeatNode;
-class VExprContext;
-class Block;
-} // namespace vectorized
+class ExecNode;
+class RuntimeState;
+
 namespace pipeline {
 
 class RepeatOperatorBuilder final : public OperatorBuilder<vectorized::VRepeatNode> {
diff --git a/be/src/pipeline/exec/result_file_sink_operator.cpp b/be/src/pipeline/exec/result_file_sink_operator.cpp
index 72f621ec8b..eeb4449ad6 100644
--- a/be/src/pipeline/exec/result_file_sink_operator.cpp
+++ b/be/src/pipeline/exec/result_file_sink_operator.cpp
@@ -17,8 +17,15 @@
 
 #include "result_file_sink_operator.h"
 
+#include <memory>
+
+#include "pipeline/exec/operator.h"
 #include "vec/sink/vresult_file_sink.h"
 
+namespace doris {
+class DataSink;
+} // namespace doris
+
 namespace doris::pipeline {
 
 ResultFileSinkOperatorBuilder::ResultFileSinkOperatorBuilder(int32_t id, DataSink* sink)
diff --git a/be/src/pipeline/exec/result_file_sink_operator.h b/be/src/pipeline/exec/result_file_sink_operator.h
index 5f21b108a7..beb04ff79c 100644
--- a/be/src/pipeline/exec/result_file_sink_operator.h
+++ b/be/src/pipeline/exec/result_file_sink_operator.h
@@ -17,12 +17,13 @@
 
 #pragma once
 
+#include <stdint.h>
+
 #include "operator.h"
+#include "vec/sink/vresult_file_sink.h"
 
 namespace doris {
-namespace vectorized {
-class VResultFileSink;
-}
+class DataSink;
 
 namespace pipeline {
 
diff --git a/be/src/pipeline/exec/result_sink_operator.cpp b/be/src/pipeline/exec/result_sink_operator.cpp
index 971c459671..c55b3599ab 100644
--- a/be/src/pipeline/exec/result_sink_operator.cpp
+++ b/be/src/pipeline/exec/result_sink_operator.cpp
@@ -17,9 +17,16 @@
 
 #include "result_sink_operator.h"
 
+#include <memory>
+
+#include "pipeline/exec/operator.h"
 #include "runtime/buffer_control_block.h"
 #include "vec/sink/vresult_sink.h"
 
+namespace doris {
+class DataSink;
+} // namespace doris
+
 namespace doris::pipeline {
 
 ResultSinkOperatorBuilder::ResultSinkOperatorBuilder(int32_t id, DataSink* sink)
diff --git a/be/src/pipeline/exec/result_sink_operator.h b/be/src/pipeline/exec/result_sink_operator.h
index 483cd98261..45a324ac81 100644
--- a/be/src/pipeline/exec/result_sink_operator.h
+++ b/be/src/pipeline/exec/result_sink_operator.h
@@ -17,12 +17,13 @@
 
 #pragma once
 
+#include <stdint.h>
+
 #include "operator.h"
+#include "vec/sink/vresult_sink.h"
 
 namespace doris {
-namespace vectorized {
-class VResultSink;
-}
+class DataSink;
 
 namespace pipeline {
 
diff --git a/be/src/pipeline/exec/scan_operator.cpp b/be/src/pipeline/exec/scan_operator.cpp
index 10a550a371..83e41d93ba 100644
--- a/be/src/pipeline/exec/scan_operator.cpp
+++ b/be/src/pipeline/exec/scan_operator.cpp
@@ -17,6 +17,11 @@
 
 #include "scan_operator.h"
 
+#include <fmt/format.h>
+
+#include <memory>
+
+#include "pipeline/exec/operator.h"
 #include "vec/exec/scan/scanner_context.h"
 #include "vec/exec/scan/vscan_node.h"
 
diff --git a/be/src/pipeline/exec/scan_operator.h b/be/src/pipeline/exec/scan_operator.h
index a4810f9730..dbcf8f0ed7 100644
--- a/be/src/pipeline/exec/scan_operator.h
+++ b/be/src/pipeline/exec/scan_operator.h
@@ -17,13 +17,17 @@
 
 #pragma once
 
-#include "operator.h"
+#include <stdint.h>
 
-namespace doris::vectorized {
-class VScanNode;
-class VScanner;
-class ScannerContext;
-} // namespace doris::vectorized
+#include <string>
+
+#include "common/status.h"
+#include "operator.h"
+#include "vec/exec/scan/vscan_node.h"
+
+namespace doris {
+class ExecNode;
+} // namespace doris
 
 namespace doris::pipeline {
 
diff --git a/be/src/pipeline/exec/schema_scan_operator.cpp b/be/src/pipeline/exec/schema_scan_operator.cpp
index 497d425bb8..c58d92392d 100644
--- a/be/src/pipeline/exec/schema_scan_operator.cpp
+++ b/be/src/pipeline/exec/schema_scan_operator.cpp
@@ -17,8 +17,16 @@
 
 #include "schema_scan_operator.h"
 
+#include <memory>
+
+#include "pipeline/exec/operator.h"
+#include "util/runtime_profile.h"
 #include "vec/exec/vschema_scan_node.h"
 
+namespace doris {
+class RuntimeState;
+} // namespace doris
+
 namespace doris::pipeline {
 
 OPERATOR_CODE_GENERATOR(SchemaScanOperator, SourceOperator)
diff --git a/be/src/pipeline/exec/schema_scan_operator.h b/be/src/pipeline/exec/schema_scan_operator.h
index 34b56e3b13..df0eab8391 100644
--- a/be/src/pipeline/exec/schema_scan_operator.h
+++ b/be/src/pipeline/exec/schema_scan_operator.h
@@ -17,11 +17,16 @@
 
 #pragma once
 
-#include "operator.h"
+#include <stdint.h>
 
-namespace doris::vectorized {
-class VSchemaScanNode;
-} // namespace doris::vectorized
+#include "common/status.h"
+#include "operator.h"
+#include "vec/exec/vschema_scan_node.h"
+
+namespace doris {
+class ExecNode;
+class RuntimeState;
+} // namespace doris
 
 namespace doris::pipeline {
 
diff --git a/be/src/pipeline/exec/select_operator.cpp b/be/src/pipeline/exec/select_operator.cpp
index c629c29978..b4eb27eace 100644
--- a/be/src/pipeline/exec/select_operator.cpp
+++ b/be/src/pipeline/exec/select_operator.cpp
@@ -17,7 +17,9 @@
 
 #include "select_operator.h"
 
-#include "vec/exec/vselect_node.h"
+#include <string>
+
+#include "pipeline/exec/operator.h"
 
 namespace doris::pipeline {
 
diff --git a/be/src/pipeline/exec/select_operator.h b/be/src/pipeline/exec/select_operator.h
index d71cb319cf..592d619246 100644
--- a/be/src/pipeline/exec/select_operator.h
+++ b/be/src/pipeline/exec/select_operator.h
@@ -17,14 +17,14 @@
 
 #pragma once
 
+#include <stdint.h>
+
 #include "operator.h"
+#include "vec/exec/vselect_node.h"
 
 namespace doris {
-namespace vectorized {
-class VSelectNode;
-class VExprContext;
-class Block;
-} // namespace vectorized
+class ExecNode;
+
 namespace pipeline {
 
 class SelectOperatorBuilder final : public OperatorBuilder<vectorized::VSelectNode> {
diff --git a/be/src/pipeline/exec/set_probe_sink_operator.cpp b/be/src/pipeline/exec/set_probe_sink_operator.cpp
index 9ac62b08fb..eefea17aad 100644
--- a/be/src/pipeline/exec/set_probe_sink_operator.cpp
+++ b/be/src/pipeline/exec/set_probe_sink_operator.cpp
@@ -17,8 +17,22 @@
 
 #include "set_probe_sink_operator.h"
 
+#include <glog/logging.h>
+
+#include <memory>
+
+#include "pipeline/exec/operator.h"
 #include "vec/exec/vset_operation_node.h"
 
+namespace doris {
+class ExecNode;
+class RuntimeState;
+
+namespace vectorized {
+class Block;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::pipeline {
 
 template <bool is_intersect>
diff --git a/be/src/pipeline/exec/set_probe_sink_operator.h b/be/src/pipeline/exec/set_probe_sink_operator.h
index 899bee2324..91ba62a199 100644
--- a/be/src/pipeline/exec/set_probe_sink_operator.h
+++ b/be/src/pipeline/exec/set_probe_sink_operator.h
@@ -17,14 +17,19 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include "common/status.h"
 #include "operator.h"
+#include "vec/exec/vset_operation_node.h"
 
 namespace doris {
+class ExecNode;
+class RuntimeState;
 
 namespace vectorized {
-template <bool is_intersect>
-class VSetOperationNode;
-}
+class Block;
+} // namespace vectorized
 
 namespace pipeline {
 
diff --git a/be/src/pipeline/exec/set_sink_operator.cpp b/be/src/pipeline/exec/set_sink_operator.cpp
index ea85b4f54e..c44f8d4db8 100644
--- a/be/src/pipeline/exec/set_sink_operator.cpp
+++ b/be/src/pipeline/exec/set_sink_operator.cpp
@@ -17,8 +17,15 @@
 
 #include "set_sink_operator.h"
 
+#include <memory>
+
+#include "pipeline/exec/operator.h"
 #include "vec/exec/vset_operation_node.h"
 
+namespace doris {
+class ExecNode;
+} // namespace doris
+
 namespace doris::pipeline {
 
 template <bool is_intersect>
diff --git a/be/src/pipeline/exec/set_sink_operator.h b/be/src/pipeline/exec/set_sink_operator.h
index bd93592769..f622bef19a 100644
--- a/be/src/pipeline/exec/set_sink_operator.h
+++ b/be/src/pipeline/exec/set_sink_operator.h
@@ -17,14 +17,13 @@
 
 #pragma once
 
+#include <stdint.h>
+
 #include "operator.h"
+#include "vec/exec/vset_operation_node.h"
 
 namespace doris {
-
-namespace vectorized {
-template <bool is_intersect>
-class VSetOperationNode;
-}
+class ExecNode;
 
 namespace pipeline {
 
diff --git a/be/src/pipeline/exec/set_source_operator.cpp b/be/src/pipeline/exec/set_source_operator.cpp
index 078df0476e..79566d4d4a 100644
--- a/be/src/pipeline/exec/set_source_operator.cpp
+++ b/be/src/pipeline/exec/set_source_operator.cpp
@@ -17,8 +17,15 @@
 
 #include "set_source_operator.h"
 
+#include <memory>
+
+#include "pipeline/exec/operator.h"
 #include "vec/exec/vset_operation_node.h"
 
+namespace doris {
+class ExecNode;
+} // namespace doris
+
 namespace doris::pipeline {
 
 template <bool is_intersect>
diff --git a/be/src/pipeline/exec/set_source_operator.h b/be/src/pipeline/exec/set_source_operator.h
index 14de0079fa..52108e6c08 100644
--- a/be/src/pipeline/exec/set_source_operator.h
+++ b/be/src/pipeline/exec/set_source_operator.h
@@ -17,14 +17,15 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include "common/status.h"
 #include "operator.h"
+#include "vec/exec/vset_operation_node.h"
 
 namespace doris {
-
-namespace vectorized {
-template <bool is_intersect>
-class VSetOperationNode;
-}
+class ExecNode;
+class RuntimeState;
 
 namespace pipeline {
 
diff --git a/be/src/pipeline/exec/sort_sink_operator.cpp b/be/src/pipeline/exec/sort_sink_operator.cpp
index cdcd75aaaa..5326b5d757 100644
--- a/be/src/pipeline/exec/sort_sink_operator.cpp
+++ b/be/src/pipeline/exec/sort_sink_operator.cpp
@@ -17,7 +17,9 @@
 
 #include "sort_sink_operator.h"
 
-#include "vec/exec/vsort_node.h"
+#include <string>
+
+#include "pipeline/exec/operator.h"
 
 namespace doris::pipeline {
 
diff --git a/be/src/pipeline/exec/sort_sink_operator.h b/be/src/pipeline/exec/sort_sink_operator.h
index c9fc6b5255..6ee8f4e170 100644
--- a/be/src/pipeline/exec/sort_sink_operator.h
+++ b/be/src/pipeline/exec/sort_sink_operator.h
@@ -17,13 +17,13 @@
 
 #pragma once
 
+#include <stdint.h>
+
 #include "operator.h"
+#include "vec/exec/vsort_node.h"
 
 namespace doris {
-
-namespace vectorized {
-class VSortNode;
-}
+class ExecNode;
 
 namespace pipeline {
 
diff --git a/be/src/pipeline/exec/sort_source_operator.cpp b/be/src/pipeline/exec/sort_source_operator.cpp
index a2b859d949..0c2b5f4b6e 100644
--- a/be/src/pipeline/exec/sort_source_operator.cpp
+++ b/be/src/pipeline/exec/sort_source_operator.cpp
@@ -17,7 +17,9 @@
 
 #include "sort_source_operator.h"
 
-#include "vec/exec/vsort_node.h"
+#include <string>
+
+#include "pipeline/exec/operator.h"
 
 namespace doris::pipeline {
 
diff --git a/be/src/pipeline/exec/sort_source_operator.h b/be/src/pipeline/exec/sort_source_operator.h
index 017e91f60c..389b86236f 100644
--- a/be/src/pipeline/exec/sort_source_operator.h
+++ b/be/src/pipeline/exec/sort_source_operator.h
@@ -17,13 +17,15 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include "common/status.h"
 #include "operator.h"
+#include "vec/exec/vsort_node.h"
 
 namespace doris {
-
-namespace vectorized {
-class VSortNode;
-}
+class ExecNode;
+class RuntimeState;
 
 namespace pipeline {
 
diff --git a/be/src/pipeline/exec/streaming_aggregation_sink_operator.cpp b/be/src/pipeline/exec/streaming_aggregation_sink_operator.cpp
index c896862702..29276f4112 100644
--- a/be/src/pipeline/exec/streaming_aggregation_sink_operator.cpp
+++ b/be/src/pipeline/exec/streaming_aggregation_sink_operator.cpp
@@ -17,8 +17,21 @@
 
 #include "streaming_aggregation_sink_operator.h"
 
+#include <gen_cpp/Metrics_types.h>
+
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "pipeline/exec/data_queue.h"
+#include "pipeline/exec/operator.h"
 #include "vec/exec/vaggregation_node.h"
 
+namespace doris {
+class ExecNode;
+class RuntimeState;
+} // namespace doris
+
 namespace doris::pipeline {
 
 StreamingAggSinkOperator::StreamingAggSinkOperator(OperatorBuilderBase* operator_builder,
diff --git a/be/src/pipeline/exec/streaming_aggregation_sink_operator.h b/be/src/pipeline/exec/streaming_aggregation_sink_operator.h
index 3d47f7fb37..02525c275e 100644
--- a/be/src/pipeline/exec/streaming_aggregation_sink_operator.h
+++ b/be/src/pipeline/exec/streaming_aggregation_sink_operator.h
@@ -17,17 +17,22 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include <memory>
+
+#include "common/status.h"
 #include "operator.h"
-#include "pipeline/exec/data_queue.h"
+#include "util/runtime_profile.h"
+#include "vec/core/block.h"
+#include "vec/exec/vaggregation_node.h"
 
 namespace doris {
-namespace vectorized {
-class AggregationNode;
-class VExprContext;
-class Block;
-} // namespace vectorized
+class ExecNode;
+class RuntimeState;
 
 namespace pipeline {
+class DataQueue;
 
 class StreamingAggSinkOperatorBuilder final : public OperatorBuilder<vectorized::AggregationNode> {
 public:
diff --git a/be/src/pipeline/exec/streaming_aggregation_source_operator.cpp b/be/src/pipeline/exec/streaming_aggregation_source_operator.cpp
index f048cd4b62..a293052232 100644
--- a/be/src/pipeline/exec/streaming_aggregation_source_operator.cpp
+++ b/be/src/pipeline/exec/streaming_aggregation_source_operator.cpp
@@ -17,9 +17,19 @@
 
 #include "streaming_aggregation_source_operator.h"
 
+#include <utility>
+
+#include "pipeline/exec/data_queue.h"
+#include "pipeline/exec/operator.h"
+#include "runtime/descriptors.h"
+#include "util/runtime_profile.h"
+#include "vec/core/block.h"
 #include "vec/exec/vaggregation_node.h"
 
 namespace doris {
+class ExecNode;
+class RuntimeState;
+
 namespace pipeline {
 StreamingAggSourceOperator::StreamingAggSourceOperator(OperatorBuilderBase* templ, ExecNode* node,
                                                        std::shared_ptr<DataQueue> queue)
diff --git a/be/src/pipeline/exec/streaming_aggregation_source_operator.h b/be/src/pipeline/exec/streaming_aggregation_source_operator.h
index 74b4be5420..ef13e03c4b 100644
--- a/be/src/pipeline/exec/streaming_aggregation_source_operator.h
+++ b/be/src/pipeline/exec/streaming_aggregation_source_operator.h
@@ -16,16 +16,23 @@
 // under the License.
 #pragma once
 
-#include <atomic>
+#include <stdint.h>
 
+#include <memory>
+
+#include "common/status.h"
 #include "operator.h"
-#include "pipeline/exec/data_queue.h"
+#include "vec/exec/vaggregation_node.h"
 
 namespace doris {
+class ExecNode;
+class RuntimeState;
+
 namespace vectorized {
-class AggregationNode;
-}
+class Block;
+} // namespace vectorized
 namespace pipeline {
+class DataQueue;
 
 class StreamingAggSourceOperatorBuilder final
         : public OperatorBuilder<vectorized::AggregationNode> {
diff --git a/be/src/pipeline/exec/table_function_operator.cpp b/be/src/pipeline/exec/table_function_operator.cpp
index 6d38726742..3f6133c3fc 100644
--- a/be/src/pipeline/exec/table_function_operator.cpp
+++ b/be/src/pipeline/exec/table_function_operator.cpp
@@ -17,6 +17,15 @@
 
 #include "table_function_operator.h"
 
+#include <memory>
+
+#include "pipeline/exec/operator.h"
+#include "vec/core/block.h"
+
+namespace doris {
+class RuntimeState;
+} // namespace doris
+
 namespace doris::pipeline {
 
 OPERATOR_CODE_GENERATOR(TableFunctionOperator, StatefulOperator)
diff --git a/be/src/pipeline/exec/table_function_operator.h b/be/src/pipeline/exec/table_function_operator.h
index 1d544e8834..a4b4e77141 100644
--- a/be/src/pipeline/exec/table_function_operator.h
+++ b/be/src/pipeline/exec/table_function_operator.h
@@ -17,9 +17,17 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include "common/status.h"
 #include "operator.h"
 #include "vec/exec/vtable_function_node.h"
 
+namespace doris {
+class ExecNode;
+class RuntimeState;
+} // namespace doris
+
 namespace doris::pipeline {
 
 class TableFunctionOperatorBuilder final : public OperatorBuilder<vectorized::VTableFunctionNode> {
diff --git a/be/src/pipeline/exec/union_sink_operator.cpp b/be/src/pipeline/exec/union_sink_operator.cpp
index 488d2dc58a..7ea31cdeaf 100644
--- a/be/src/pipeline/exec/union_sink_operator.cpp
+++ b/be/src/pipeline/exec/union_sink_operator.cpp
@@ -17,7 +17,19 @@
 
 #include "union_sink_operator.h"
 
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/status.h"
+#include "pipeline/exec/data_queue.h"
+#include "pipeline/exec/operator.h"
+#include "runtime/runtime_state.h"
+#include "util/runtime_profile.h"
+
+namespace doris {
+class ExecNode;
+} // namespace doris
 
 namespace doris::pipeline {
 
diff --git a/be/src/pipeline/exec/union_sink_operator.h b/be/src/pipeline/exec/union_sink_operator.h
index 707a980cf4..dfe7085d5e 100644
--- a/be/src/pipeline/exec/union_sink_operator.h
+++ b/be/src/pipeline/exec/union_sink_operator.h
@@ -17,16 +17,21 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include <memory>
+
+#include "common/status.h"
 #include "operator.h"
-#include "pipeline/exec/data_queue.h"
 #include "vec/core/block.h"
+#include "vec/exec/vunion_node.h"
+
 namespace doris {
-namespace vectorized {
-class VUnionNode;
-class Block;
-} // namespace vectorized
+class ExecNode;
+class RuntimeState;
 
 namespace pipeline {
+class DataQueue;
 
 class UnionSinkOperatorBuilder final : public OperatorBuilder<vectorized::VUnionNode> {
 public:
diff --git a/be/src/pipeline/exec/union_source_operator.cpp b/be/src/pipeline/exec/union_source_operator.cpp
index b65e7b3f51..83069f3e90 100644
--- a/be/src/pipeline/exec/union_source_operator.cpp
+++ b/be/src/pipeline/exec/union_source_operator.cpp
@@ -17,15 +17,18 @@
 
 #include "pipeline/exec/union_source_operator.h"
 
-#include <opentelemetry/common/threadlocal.h>
+#include <functional>
+#include <utility>
 
 #include "common/status.h"
 #include "pipeline/exec/data_queue.h"
+#include "pipeline/exec/operator.h"
+#include "runtime/descriptors.h"
+#include "vec/core/block.h"
 
 namespace doris {
-namespace vectorized {
-class Block;
-}
+class ExecNode;
+class RuntimeState;
 
 namespace pipeline {
 
diff --git a/be/src/pipeline/exec/union_source_operator.h b/be/src/pipeline/exec/union_source_operator.h
index d330b45a82..33a6bef587 100644
--- a/be/src/pipeline/exec/union_source_operator.h
+++ b/be/src/pipeline/exec/union_source_operator.h
@@ -16,15 +16,24 @@
 // under the License.
 #pragma once
 
+#include <stdint.h>
+
+#include <memory>
+
+#include "common/status.h"
 #include "operator.h"
-#include "pipeline/exec/data_queue.h"
+#include "vec/exec/vunion_node.h"
 
 namespace doris {
+class ExecNode;
+class RuntimeState;
+
 namespace vectorized {
-class VUnionNode;
-}
+class Block;
+} // namespace vectorized
 
 namespace pipeline {
+class DataQueue;
 
 class UnionSourceOperatorBuilder final : public OperatorBuilder<vectorized::VUnionNode> {
 public:
diff --git a/be/src/pipeline/pipeline.cpp b/be/src/pipeline/pipeline.cpp
index c5a94ba640..ccc7786442 100644
--- a/be/src/pipeline/pipeline.cpp
+++ b/be/src/pipeline/pipeline.cpp
@@ -17,7 +17,11 @@
 
 #include "pipeline.h"
 
-#include "pipeline_fragment_context.h"
+#include <ostream>
+#include <typeinfo>
+#include <utility>
+
+#include "pipeline/exec/operator.h"
 
 namespace doris::pipeline {
 
diff --git a/be/src/pipeline/pipeline.h b/be/src/pipeline/pipeline.h
index dc1277f721..4cf9b53086 100644
--- a/be/src/pipeline/pipeline.h
+++ b/be/src/pipeline/pipeline.h
@@ -17,16 +17,28 @@
 
 #pragma once
 
+#include <glog/logging.h>
+#include <stdint.h>
+
+#include <algorithm>
 #include <atomic>
 #include <memory>
 #include <vector>
 
 #include "common/status.h"
 #include "exec/operator.h"
+#include "util/runtime_profile.h"
+
+namespace doris {
+namespace pipeline {
+class PipelineFragmentContext;
+} // namespace pipeline
+} // namespace doris
 
 namespace doris::pipeline {
 
 class Pipeline;
+
 using PipelinePtr = std::shared_ptr<Pipeline>;
 using Pipelines = std::vector<PipelinePtr>;
 using PipelineId = uint32_t;
diff --git a/be/src/pipeline/pipeline_fragment_context.cpp b/be/src/pipeline/pipeline_fragment_context.cpp
index a2b1afbe48..9a5c23fb6c 100644
--- a/be/src/pipeline/pipeline_fragment_context.cpp
+++ b/be/src/pipeline/pipeline_fragment_context.cpp
@@ -18,12 +18,28 @@
 #include "pipeline_fragment_context.h"
 
 #include <gen_cpp/DataSinks_types.h>
-#include <thrift/protocol/TDebugProtocol.h>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Planner_types.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+#include <opentelemetry/trace/span.h>
+#include <opentelemetry/trace/span_context.h>
+#include <opentelemetry/trace/tracer.h>
+#include <pthread.h>
+#include <stdlib.h>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <map>
+#include <ostream>
+#include <typeinfo>
+#include <utility>
 
+#include "common/config.h"
+#include "common/logging.h"
 #include "exec/data_sink.h"
+#include "exec/exec_node.h"
 #include "exec/scan_node.h"
-#include "gen_cpp/FrontendService.h"
-#include "gen_cpp/HeartbeatService_types.h"
+#include "io/fs/stream_load_pipe.h"
 #include "pipeline/exec/aggregation_sink_operator.h"
 #include "pipeline/exec/aggregation_source_operator.h"
 #include "pipeline/exec/analytic_sink_operator.h"
@@ -38,7 +54,7 @@
 #include "pipeline/exec/exchange_source_operator.h"
 #include "pipeline/exec/hashjoin_build_sink.h"
 #include "pipeline/exec/hashjoin_probe_operator.h"
-#include "pipeline/exec/mysql_scan_operator.h"
+#include "pipeline/exec/mysql_scan_operator.h" // IWYU pragma: keep
 #include "pipeline/exec/nested_loop_join_build_operator.h"
 #include "pipeline/exec/nested_loop_join_probe_operator.h"
 #include "pipeline/exec/olap_table_sink_operator.h"
@@ -49,9 +65,9 @@
 #include "pipeline/exec/scan_operator.h"
 #include "pipeline/exec/schema_scan_operator.h"
 #include "pipeline/exec/select_operator.h"
-#include "pipeline/exec/set_probe_sink_operator.h"
-#include "pipeline/exec/set_sink_operator.h"
-#include "pipeline/exec/set_source_operator.h"
+#include "pipeline/exec/set_probe_sink_operator.h" // IWYU pragma: keep
+#include "pipeline/exec/set_sink_operator.h"       // IWYU pragma: keep
+#include "pipeline/exec/set_source_operator.h"     // IWYU pragma: keep
 #include "pipeline/exec/sort_sink_operator.h"
 #include "pipeline/exec/sort_source_operator.h"
 #include "pipeline/exec/streaming_aggregation_sink_operator.h"
@@ -61,32 +77,41 @@
 #include "pipeline/exec/union_sink_operator.h"
 #include "pipeline/exec/union_source_operator.h"
 #include "pipeline_task.h"
-#include "runtime/client_cache.h"
+#include "runtime/exec_env.h"
 #include "runtime/fragment_mgr.h"
+#include "runtime/runtime_filter_mgr.h"
 #include "runtime/runtime_state.h"
 #include "runtime/stream_load/new_load_stream_mgr.h"
 #include "runtime/stream_load/stream_load_context.h"
+#include "runtime/thread_context.h"
+#include "service/backend_options.h"
 #include "task_scheduler.h"
 #include "util/container_util.hpp"
+#include "util/debug_util.h"
+#include "util/telemetry/telemetry.h"
+#include "util/uid_util.h"
+#include "vec/common/assert_cast.h"
 #include "vec/exec/join/vhash_join_node.h"
-#include "vec/exec/join/vnested_loop_join_node.h"
 #include "vec/exec/scan/new_es_scan_node.h"
 #include "vec/exec/scan/new_file_scan_node.h"
-#include "vec/exec/scan/new_jdbc_scan_node.h"
 #include "vec/exec/scan/new_odbc_scan_node.h"
 #include "vec/exec/scan/new_olap_scan_node.h"
 #include "vec/exec/scan/vmeta_scan_node.h"
 #include "vec/exec/scan/vscan_node.h"
 #include "vec/exec/vaggregation_node.h"
 #include "vec/exec/vexchange_node.h"
-#include "vec/exec/vrepeat_node.h"
-#include "vec/exec/vschema_scan_node.h"
-#include "vec/exec/vset_operation_node.h"
-#include "vec/exec/vsort_node.h"
 #include "vec/exec/vunion_node.h"
 #include "vec/runtime/vdata_stream_mgr.h"
-#include "vec/sink/vresult_file_sink.h"
-#include "vec/sink/vresult_sink.h"
+
+namespace apache {
+namespace thrift {
+class TException;
+
+namespace transport {
+class TTransportException;
+} // namespace transport
+} // namespace thrift
+} // namespace apache
 
 using apache::thrift::transport::TTransportException;
 using apache::thrift::TException;
diff --git a/be/src/pipeline/pipeline_fragment_context.h b/be/src/pipeline/pipeline_fragment_context.h
index b8db2728d2..5e23ab0e2b 100644
--- a/be/src/pipeline/pipeline_fragment_context.h
+++ b/be/src/pipeline/pipeline_fragment_context.h
@@ -17,27 +17,43 @@
 
 #pragma once
 
-#include <future>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/types.pb.h>
+#include <stddef.h>
+#include <stdint.h>
 
-#include "io/fs/stream_load_pipe.h"
+#include <atomic>
+#include <condition_variable>
+#include <functional>
+#include <future>
+#include <memory>
+#include <mutex>
+#include <string>
+#include <vector>
+
+#include "common/status.h"
 #include "pipeline/pipeline.h"
 #include "pipeline/pipeline_task.h"
+#include "runtime/query_fragments_ctx.h"
 #include "runtime/runtime_state.h"
+#include "util/runtime_profile.h"
+#include "util/stopwatch.hpp"
 
 namespace doris {
 class ExecNode;
 class DataSink;
 struct ReportStatusRequest;
+class ExecEnv;
+class RuntimeFilterMergeControllerEntity;
+class TDataSink;
+class TPipelineFragmentParams;
 
-namespace vectorized {
-template <bool is_intersect>
-class VSetOperationNode;
-}
+namespace taskgroup {
+class TaskGroup;
+} // namespace taskgroup
 
 namespace pipeline {
 
-class PipelineTask;
-
 class PipelineFragmentContext : public std::enable_shared_from_this<PipelineFragmentContext> {
 public:
     // Callback to report execution status of plan fragment.
diff --git a/be/src/pipeline/pipeline_task.cpp b/be/src/pipeline/pipeline_task.cpp
index 8121414a6b..6d69ebfd89 100644
--- a/be/src/pipeline/pipeline_task.cpp
+++ b/be/src/pipeline/pipeline_task.cpp
@@ -17,8 +17,28 @@
 
 #include "pipeline_task.h"
 
+#include <fmt/format.h>
+#include <gen_cpp/Metrics_types.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <ostream>
+
+#include "pipeline/exec/operator.h"
+#include "pipeline/pipeline.h"
 #include "pipeline_fragment_context.h"
+#include "runtime/descriptors.h"
+#include "runtime/query_fragments_ctx.h"
+#include "runtime/thread_context.h"
 #include "task_queue.h"
+#include "util/defer_op.h"
+
+namespace doris {
+class RuntimeState;
+namespace taskgroup {
+class TaskGroup;
+} // namespace taskgroup
+} // namespace doris
 
 namespace doris::pipeline {
 
diff --git a/be/src/pipeline/pipeline_task.h b/be/src/pipeline/pipeline_task.h
index 043da50d80..02eb8be992 100644
--- a/be/src/pipeline/pipeline_task.h
+++ b/be/src/pipeline/pipeline_task.h
@@ -17,9 +17,29 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "common/status.h"
 #include "exec/operator.h"
 #include "pipeline.h"
+#include "util/runtime_profile.h"
 #include "util/stopwatch.hpp"
+#include "vec/core/block.h"
+
+namespace doris {
+class QueryFragmentsCtx;
+class RuntimeState;
+namespace pipeline {
+class PipelineFragmentContext;
+} // namespace pipeline
+namespace taskgroup {
+class TaskGroup;
+} // namespace taskgroup
+} // namespace doris
 
 namespace doris::pipeline {
 
diff --git a/be/src/pipeline/task_queue.cpp b/be/src/pipeline/task_queue.cpp
index dac6ca69ab..7ad69bfa21 100644
--- a/be/src/pipeline/task_queue.cpp
+++ b/be/src/pipeline/task_queue.cpp
@@ -17,6 +17,12 @@
 
 #include "task_queue.h"
 
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <string>
+
+#include "common/logging.h"
+#include "pipeline/pipeline_task.h"
 #include "runtime/task_group/task_group.h"
 
 namespace doris {
diff --git a/be/src/pipeline/task_queue.h b/be/src/pipeline/task_queue.h
index 2aa4881028..6b16b70f99 100644
--- a/be/src/pipeline/task_queue.h
+++ b/be/src/pipeline/task_queue.h
@@ -16,14 +16,23 @@
 // under the License.
 #pragma once
 
-#include <queue>
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <atomic>
+#include <condition_variable>
+#include <memory>
+#include <mutex>
+#include <ostream>
+#include <queue>
+#include <set>
+
+#include "common/status.h"
 #include "pipeline_task.h"
+#include "runtime/task_group/task_group.h"
 
 namespace doris {
-namespace taskgroup {
-class TaskGroup;
-}
 
 namespace pipeline {
 
diff --git a/be/src/pipeline/task_scheduler.cpp b/be/src/pipeline/task_scheduler.cpp
index 70d1bd518f..7362dd38b3 100644
--- a/be/src/pipeline/task_scheduler.cpp
+++ b/be/src/pipeline/task_scheduler.cpp
@@ -17,9 +17,30 @@
 
 #include "task_scheduler.h"
 
+#include <fmt/format.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/types.pb.h>
+#include <glog/logging.h>
+#include <sched.h>
+
+#include <algorithm>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <functional>
+#include <ostream>
+#include <string>
+#include <thread>
+
 #include "common/signal_handler.h"
+#include "pipeline/pipeline_task.h"
+#include "pipeline/task_queue.h"
 #include "pipeline_fragment_context.h"
+#include "runtime/query_fragments_ctx.h"
+#include "util/sse_util.hpp"
 #include "util/thread.h"
+#include "util/threadpool.h"
+#include "util/uid_util.h"
+#include "vec/runtime/vdatetime_value.h"
 
 namespace doris::pipeline {
 
@@ -342,4 +363,4 @@ void TaskScheduler::try_update_task_group(const taskgroup::TaskGroupInfo& task_g
     _task_queue->update_task_group(task_group_info, task_group);
 }
 
-} // namespace doris::pipeline
\ No newline at end of file
+} // namespace doris::pipeline
diff --git a/be/src/pipeline/task_scheduler.h b/be/src/pipeline/task_scheduler.h
index 777923dbd3..b2a665f034 100644
--- a/be/src/pipeline/task_scheduler.h
+++ b/be/src/pipeline/task_scheduler.h
@@ -17,11 +17,30 @@
 
 #pragma once
 
+#include <stddef.h>
+
+#include <atomic>
+#include <condition_variable>
+#include <list>
+#include <memory>
+#include <mutex>
+#include <utility>
+#include <vector>
+
 #include "common/status.h"
-#include "pipeline.h"
+#include "gutil/ref_counted.h"
 #include "pipeline_task.h"
-#include "task_queue.h"
-#include "util/threadpool.h"
+#include "runtime/task_group/task_group.h"
+#include "util/thread.h"
+
+namespace doris {
+class ExecEnv;
+class ThreadPool;
+
+namespace pipeline {
+class TaskQueue;
+} // namespace pipeline
+} // namespace doris
 
 namespace doris::pipeline {
 
diff --git a/be/src/runtime/block_spill_manager.cpp b/be/src/runtime/block_spill_manager.cpp
index b627b5543f..b45e4c3ad6 100644
--- a/be/src/runtime/block_spill_manager.cpp
+++ b/be/src/runtime/block_spill_manager.cpp
@@ -17,10 +17,16 @@
 
 #include "runtime/block_spill_manager.h"
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+
+#include <algorithm>
 #include <boost/uuid/random_generator.hpp>
 #include <boost/uuid/uuid_io.hpp>
+#include <numeric>
 #include <random>
 
+#include "io/fs/file_system.h"
 #include "io/fs/local_file_system.h"
 #include "util/time.h"
 #include "vec/core/block_spill_reader.h"
diff --git a/be/src/runtime/block_spill_manager.h b/be/src/runtime/block_spill_manager.h
index f406b8c9ac..5601d58aed 100644
--- a/be/src/runtime/block_spill_manager.h
+++ b/be/src/runtime/block_spill_manager.h
@@ -17,24 +17,28 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include <memory>
 #include <mutex>
 #include <string>
 #include <unordered_map>
 #include <vector>
 
+#include "common/status.h"
 #include "olap/options.h"
+#include "vec/core/block_spill_reader.h"
+#include "vec/core/block_spill_writer.h"
 
 namespace doris {
 class RuntimeProfile;
 
 namespace vectorized {
-class BlockSpillWriter;
-class BlockSpillReader;
+
 using BlockSpillWriterUPtr = std::unique_ptr<BlockSpillWriter>;
 using BlockSpillReaderUPtr = std::unique_ptr<BlockSpillReader>;
 } // namespace vectorized
 
-class ExecEnv;
 class BlockSpillManager {
 public:
     BlockSpillManager(const std::vector<StorePath>& paths);
diff --git a/be/src/runtime/broker_mgr.cpp b/be/src/runtime/broker_mgr.cpp
index 5e57922399..6102da5e40 100644
--- a/be/src/runtime/broker_mgr.cpp
+++ b/be/src/runtime/broker_mgr.cpp
@@ -17,16 +17,26 @@
 
 #include "runtime/broker_mgr.h"
 
+#include <gen_cpp/PaloBrokerService_types.h>
+#include <gen_cpp/TPaloBrokerService.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <thrift/Thrift.h>
+#include <thrift/transport/TTransportException.h>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
 #include <sstream>
+#include <vector>
 
 #include "common/config.h"
-#include "gen_cpp/PaloBrokerService_types.h"
-#include "gen_cpp/TPaloBrokerService.h"
+#include "common/status.h"
 #include "runtime/client_cache.h"
 #include "runtime/exec_env.h"
 #include "service/backend_options.h"
 #include "util/doris_metrics.h"
-#include "util/thrift_util.h"
+#include "util/hash_util.hpp"
+#include "util/metrics.h"
+#include "util/thread.h"
 
 namespace doris {
 
diff --git a/be/src/runtime/broker_mgr.h b/be/src/runtime/broker_mgr.h
index 7f55d78fff..fd7a67e43c 100644
--- a/be/src/runtime/broker_mgr.h
+++ b/be/src/runtime/broker_mgr.h
@@ -17,20 +17,20 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+
 #include <mutex>
 #include <string>
-#include <thread>
 #include <unordered_set>
 
-#include "gen_cpp/Types_types.h"
 #include "gutil/ref_counted.h"
 #include "util/countdown_latch.h"
-#include "util/hash_util.hpp"
-#include "util/thread.h"
+#include "util/hash_util.hpp" // IWYU pragma: keep
 
 namespace doris {
 
 class ExecEnv;
+class Thread;
 
 class BrokerMgr {
 public:
diff --git a/be/src/runtime/buffer_control_block.cpp b/be/src/runtime/buffer_control_block.cpp
index 73de8a0852..7565e3edac 100644
--- a/be/src/runtime/buffer_control_block.cpp
+++ b/be/src/runtime/buffer_control_block.cpp
@@ -17,11 +17,20 @@
 
 #include "runtime/buffer_control_block.h"
 
-#include "gen_cpp/PaloInternalService_types.h"
-#include "gen_cpp/internal_service.pb.h"
+#include <gen_cpp/Data_types.h>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/internal_service.pb.h>
+#include <glog/logging.h>
+#include <google/protobuf/stubs/callback.h>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <ostream>
+#include <string>
+#include <utility>
+#include <vector>
+
 #include "runtime/exec_env.h"
 #include "runtime/thread_context.h"
-#include "service/brpc.h"
 #include "util/thrift_util.h"
 
 namespace doris {
diff --git a/be/src/runtime/buffer_control_block.h b/be/src/runtime/buffer_control_block.h
index 46e6431f73..88fa8d2be4 100644
--- a/be/src/runtime/buffer_control_block.h
+++ b/be/src/runtime/buffer_control_block.h
@@ -17,14 +17,18 @@
 
 #pragma once
 
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/Types_types.h>
+#include <stdint.h>
+
 #include <atomic>
 #include <condition_variable>
 #include <deque>
 #include <list>
+#include <memory>
 #include <mutex>
 
 #include "common/status.h"
-#include "gen_cpp/Types_types.h"
 #include "runtime/query_statistics.h"
 
 namespace google {
@@ -39,7 +43,6 @@ class Controller;
 
 namespace doris {
 
-class TFetchDataResult;
 class PFetchDataResult;
 
 struct GetResultBatchCtx {
diff --git a/be/src/runtime/cache/result_cache.cpp b/be/src/runtime/cache/result_cache.cpp
index da4c1fc283..fec1d71f36 100644
--- a/be/src/runtime/cache/result_cache.cpp
+++ b/be/src/runtime/cache/result_cache.cpp
@@ -16,7 +16,15 @@
 // under the License.
 #include "runtime/cache/result_cache.h"
 
-#include "gen_cpp/internal_service.pb.h"
+#include <gen_cpp/internal_service.pb.h>
+#include <glog/logging.h>
+
+#include <iostream>
+#include <list>
+#include <utility>
+
+#include "olap/olap_define.h"
+#include "runtime/cache/cache_utils.h"
 #include "util/doris_metrics.h"
 
 namespace doris {
diff --git a/be/src/runtime/cache/result_cache.h b/be/src/runtime/cache/result_cache.h
index fc36f61b59..a629e7ed7d 100644
--- a/be/src/runtime/cache/result_cache.h
+++ b/be/src/runtime/cache/result_cache.h
@@ -17,22 +17,20 @@
 
 #pragma once
 
-#include <cassert>
 #include <cstdio>
-#include <cstdlib>
-#include <exception>
-#include <iostream>
-#include <list>
-#include <map>
-#include <mutex>
 #include <shared_mutex>
-#include <thread>
+#include <unordered_map>
 
-#include "common/config.h"
-#include "runtime/cache/cache_utils.h"
+#include "gutil/integral_types.h"
 #include "runtime/cache/result_node.h"
+#include "util/uid_util.h"
 
 namespace doris {
+class PCacheResponse;
+class PClearCacheRequest;
+class PFetchCacheRequest;
+class PFetchCacheResult;
+class PUpdateCacheRequest;
 
 typedef std::unordered_map<UniqueId, ResultNode*> ResultNodeMap;
 
diff --git a/be/src/runtime/cache/result_node.cpp b/be/src/runtime/cache/result_node.cpp
index 273b9cc5e6..dd00e75390 100644
--- a/be/src/runtime/cache/result_node.cpp
+++ b/be/src/runtime/cache/result_node.cpp
@@ -16,7 +16,16 @@
 // under the License.
 #include "runtime/cache/result_node.h"
 
-#include "gen_cpp/internal_service.pb.h"
+#include <gen_cpp/internal_service.pb.h>
+#include <gen_cpp/types.pb.h>
+#include <glog/logging.h>
+
+#include <iostream>
+#include <limits>
+#include <utility>
+
+#include "common/config.h"
+#include "olap/olap_define.h"
 #include "runtime/cache/cache_utils.h"
 
 namespace doris {
diff --git a/be/src/runtime/cache/result_node.h b/be/src/runtime/cache/result_node.h
index 439ee64529..3edf48261b 100644
--- a/be/src/runtime/cache/result_node.h
+++ b/be/src/runtime/cache/result_node.h
@@ -17,34 +17,19 @@
 
 #pragma once
 
-#include <sys/time.h>
+#include <gen_cpp/internal_service.pb.h>
 
-#include <algorithm>
-#include <cassert>
 #include <cstdio>
-#include <cstdlib>
-#include <exception>
-#include <iostream>
 #include <list>
-#include <map>
-#include <string>
+#include <shared_mutex>
+#include <unordered_map>
 
-#include "common/config.h"
-#include "gen_cpp/internal_service.pb.h"
-#include "olap/olap_define.h"
+#include "gutil/integral_types.h"
 #include "runtime/cache/cache_utils.h"
 #include "util/uid_util.h"
 
 namespace doris {
 
-class PCacheParam;
-class PCacheValue;
-class PCacheResponse;
-class PFetchCacheRequest;
-class PFetchCacheResult;
-class PUpdateCacheRequest;
-class PClearCacheRequest;
-
 /**
 * Cache one partition data, request param must match version and time of cache
 */
diff --git a/be/src/runtime/client_cache.cpp b/be/src/runtime/client_cache.cpp
index d4edbca6c0..3da31caf5c 100644
--- a/be/src/runtime/client_cache.cpp
+++ b/be/src/runtime/client_cache.cpp
@@ -17,13 +17,10 @@
 
 #include "runtime/client_cache.h"
 
-#include <thrift/protocol/TBinaryProtocol.h>
-
 #include <memory>
-#include <sstream>
+#include <utility>
 
 #include "common/logging.h"
-#include "gen_cpp/FrontendService.h"
 #include "util/doris_metrics.h"
 #include "util/network_util.h"
 
diff --git a/be/src/runtime/client_cache.h b/be/src/runtime/client_cache.h
index c461a45082..7ba5cafabe 100644
--- a/be/src/runtime/client_cache.h
+++ b/be/src/runtime/client_cache.h
@@ -17,19 +17,31 @@
 
 #pragma once
 
-#include <list>
-#include <mutex>
-#include <string>
-#include <unordered_map>
-#include <vector>
+#include <gen_cpp/BackendService.h>     // IWYU pragma: keep
+#include <gen_cpp/FrontendService.h>    // IWYU pragma: keep
+#include <gen_cpp/TPaloBrokerService.h> // IWYU pragma: keep
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <string.h>
+#include <unistd.h>
 
+#include <functional>
+#include <list>
+#include <memory>
+#include <mutex>
+#include <ostream>
+#include <string>
+#include <typeinfo>
+#include <unordered_map>
+
+#include "common/config.h"
 #include "common/status.h"
 #include "util/hash_util.hpp"
 #include "util/metrics.h"
 #include "util/thrift_client.h"
+#include "util/thrift_server.h"
 
 namespace doris {
-
 // Helper class which implements the majority of the caching
 // functionality without using templates (i.e. pointers to the
 // superclass of all ThriftClients and a void* for the key).
@@ -268,13 +280,12 @@ private:
 
 // Doris backend client cache, used by a backend to send requests
 // to any other backend.
-class BackendServiceClient;
 using BackendServiceClientCache = ClientCache<BackendServiceClient>;
 using BackendServiceConnection = ClientConnection<BackendServiceClient>;
-class FrontendServiceClient;
+
 using FrontendServiceClientCache = ClientCache<FrontendServiceClient>;
 using FrontendServiceConnection = ClientConnection<FrontendServiceClient>;
-class TPaloBrokerServiceClient;
+
 using BrokerServiceClientCache = ClientCache<TPaloBrokerServiceClient>;
 using BrokerServiceConnection = ClientConnection<TPaloBrokerServiceClient>;
 
diff --git a/be/src/runtime/collection_value.cpp b/be/src/runtime/collection_value.cpp
index e052130bb0..1d50172069 100644
--- a/be/src/runtime/collection_value.cpp
+++ b/be/src/runtime/collection_value.cpp
@@ -17,12 +17,7 @@
 
 #include "runtime/collection_value.h"
 
-#include <functional>
-
-#include "common/object_pool.h"
-#include "common/utils.h"
-#include "runtime/types.h"
-#include "vec/common/string_ref.h"
+#include <string.h>
 
 namespace doris {
 
diff --git a/be/src/runtime/collection_value.h b/be/src/runtime/collection_value.h
index 02a51e4bd1..82dfc78fe5 100644
--- a/be/src/runtime/collection_value.h
+++ b/be/src/runtime/collection_value.h
@@ -17,41 +17,16 @@
 
 #pragma once
 
-#include <type_traits>
+#include <stdint.h>
 
-#include "runtime/primitive_type.h"
+#include <functional>
+#include <utility>
 
 namespace doris {
-class FunctionContext;
-} // namespace doris
-
-namespace doris {
-
-using doris::FunctionContext;
 
 using MemFootprint = std::pair<int64_t, uint8_t*>;
 using GenMemFootprintFunc = std::function<MemFootprint(int64_t size)>;
 
-struct ArrayIteratorFunctionsBase;
-class ArrayIterator;
-class Status;
-class ObjectPool;
-
-template <PrimitiveType type>
-struct ArrayIteratorFunctions;
-template <typename T>
-inline constexpr std::enable_if_t<std::is_base_of_v<ArrayIteratorFunctionsBase, T>, bool>
-        IsTypeFixedWidth = true;
-
-template <>
-inline constexpr bool IsTypeFixedWidth<ArrayIteratorFunctions<TYPE_CHAR>> = false;
-template <>
-inline constexpr bool IsTypeFixedWidth<ArrayIteratorFunctions<TYPE_VARCHAR>> = false;
-template <>
-inline constexpr bool IsTypeFixedWidth<ArrayIteratorFunctions<TYPE_STRING>> = false;
-template <>
-inline constexpr bool IsTypeFixedWidth<ArrayIteratorFunctions<TYPE_ARRAY>> = false;
-
 /**
  * The format of array-typed slot.
  * A new array needs to be initialized before using it.
diff --git a/be/src/runtime/datetime_value.h b/be/src/runtime/datetime_value.h
index d56f0e77cd..88279af1b0 100644
--- a/be/src/runtime/datetime_value.h
+++ b/be/src/runtime/datetime_value.h
@@ -20,6 +20,7 @@
 #include <re2/re2.h>
 #include <stdint.h>
 
+// IWYU pragma: no_include <bits/chrono.h>
 #include <chrono>
 #include <cstddef>
 #include <iostream>
diff --git a/be/src/runtime/decimalv2_value.cpp b/be/src/runtime/decimalv2_value.cpp
index 0256f6210f..d8552f3fa1 100644
--- a/be/src/runtime/decimalv2_value.cpp
+++ b/be/src/runtime/decimalv2_value.cpp
@@ -17,7 +17,10 @@
 
 #include "runtime/decimalv2_value.h"
 
-#include <algorithm>
+#include <fmt/format.h>
+
+#include <cmath>
+#include <cstring>
 #include <iostream>
 #include <utility>
 
diff --git a/be/src/runtime/decimalv2_value.h b/be/src/runtime/decimalv2_value.h
index ddce791e45..4a6e2cb6cf 100644
--- a/be/src/runtime/decimalv2_value.h
+++ b/be/src/runtime/decimalv2_value.h
@@ -17,16 +17,16 @@
 
 #pragma once
 
-#include <cctype>
-#include <climits>
+#include <glog/logging.h>
+#include <stdint.h>
+
+// IWYU pragma: no_include <bits/std_abs.h>
+#include <cmath> // IWYU pragma: keep
 #include <cstdlib>
-#include <cstring>
 #include <iostream>
-#include <sstream>
 #include <string>
 #include <string_view>
 
-#include "udf/udf.h"
 #include "util/hash_util.hpp"
 
 namespace doris {
diff --git a/be/src/runtime/descriptor_helper.h b/be/src/runtime/descriptor_helper.h
index ba3f17c57a..15c0d68599 100644
--- a/be/src/runtime/descriptor_helper.h
+++ b/be/src/runtime/descriptor_helper.h
@@ -17,10 +17,11 @@
 
 #pragma once
 
+#include <gen_cpp/Descriptors_types.h>
+#include <gen_cpp/Types_types.h>
+
 #include <vector>
 
-#include "gen_cpp/Descriptors_types.h"
-#include "gen_cpp/Types_types.h"
 #include "runtime/define_primitive_type.h"
 #include "runtime/primitive_type.h"
 
diff --git a/be/src/runtime/descriptors.cpp b/be/src/runtime/descriptors.cpp
index 1658a3bf6c..c13e2b53d5 100644
--- a/be/src/runtime/descriptors.cpp
+++ b/be/src/runtime/descriptors.cpp
@@ -20,19 +20,20 @@
 
 #include "runtime/descriptors.h"
 
+#include <fmt/format.h>
+#include <gen_cpp/Descriptors_types.h>
 #include <gen_cpp/Types_types.h>
+#include <gen_cpp/descriptors.pb.h>
+#include <stddef.h>
 
+#include <algorithm>
 #include <boost/algorithm/string/join.hpp>
-#include <ios>
-#include <sstream>
+#include <memory>
 
 #include "common/object_pool.h"
-#include "gen_cpp/Descriptors_types.h"
-#include "gen_cpp/descriptors.pb.h"
 #include "util/string_util.h"
-#include "vec/columns/column_nullable.h"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/data_types/data_type_factory.hpp"
-#include "vec/data_types/data_type_nullable.h"
 
 namespace doris {
 using boost::algorithm::join;
diff --git a/be/src/runtime/descriptors.h b/be/src/runtime/descriptors.h
index e03aaac7f4..8fd6654e13 100644
--- a/be/src/runtime/descriptors.h
+++ b/be/src/runtime/descriptors.h
@@ -20,38 +20,37 @@
 
 #pragma once
 
-#include <google/protobuf/repeated_field.h>
-#include <google/protobuf/stubs/common.h>
+#include <gen_cpp/Descriptors_types.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <google/protobuf/stubs/port.h>
+#include <stdint.h>
 
 #include <ostream>
+#include <string>
 #include <unordered_map>
+#include <utility>
 #include <vector>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/global_types.h"
 #include "common/status.h"
-#include "gen_cpp/Descriptors_types.h"     // for TTupleId
-#include "gen_cpp/FrontendService_types.h" // for TTupleId
-#include "gen_cpp/Types_types.h"
 #include "runtime/types.h"
 #include "vec/data_types/data_type.h"
 
-namespace doris::vectorized {
-struct ColumnWithTypeAndName;
-}
+namespace google {
+namespace protobuf {
+template <typename Element>
+class RepeatedField;
+} // namespace protobuf
+} // namespace google
 
 namespace doris {
 
 class ObjectPool;
-class TDescriptorTable;
-class TSlotDescriptor;
-class TTupleDescriptor;
-class RuntimeState;
-class SchemaScanner;
-class OlapTableSchemaParam;
 class PTupleDescriptor;
 class PSlotDescriptor;
-class PInternalServiceImpl;
-class TabletSchema;
 
 // Location information for null indicator bit for particular slot.
 // For non-nullable slots, the byte_offset will be 0 and the bit_mask will be 0.
@@ -80,8 +79,6 @@ struct NullIndicatorOffset {
 
 std::ostream& operator<<(std::ostream& os, const NullIndicatorOffset& null_indicator);
 
-class TupleDescriptor;
-
 class SlotDescriptor {
 public:
     // virtual ~SlotDescriptor() {};
diff --git a/be/src/runtime/exec_env.cpp b/be/src/runtime/exec_env.cpp
index bde3de101d..8f63d6fe6a 100644
--- a/be/src/runtime/exec_env.cpp
+++ b/be/src/runtime/exec_env.cpp
@@ -17,7 +17,7 @@
 
 #include "runtime/exec_env.h"
 
-#include "gen_cpp/HeartbeatService_types.h"
+#include <gen_cpp/HeartbeatService_types.h>
 
 namespace doris {
 
diff --git a/be/src/runtime/exec_env.h b/be/src/runtime/exec_env.h
index 6862f191db..cc9cecad59 100644
--- a/be/src/runtime/exec_env.h
+++ b/be/src/runtime/exec_env.h
@@ -17,11 +17,14 @@
 
 #pragma once
 
-#include <gen_cpp/FrontendService.h>
+#include <stddef.h>
 
+#include <map>
+#include <memory>
+#include <string>
 #include <unordered_map>
+#include <vector>
 
-#include "common/config.h"
 #include "common/status.h"
 #include "olap/options.h"
 #include "util/threadpool.h"
@@ -36,12 +39,8 @@ class TaskScheduler;
 }
 class BfdParser;
 class BrokerMgr;
-
 template <class T>
 class BrpcClientCache;
-
-class DataStreamMgr;
-class EvHttpServer;
 class ExternalScanContextMgr;
 class FragmentMgr;
 class ResultCache;
@@ -50,28 +49,22 @@ class NewLoadStreamMgr;
 class MemTrackerLimiter;
 class MemTracker;
 class StorageEngine;
-class PriorityThreadPool;
-class PriorityWorkStealingThreadPool;
 class ResultBufferMgr;
 class ResultQueueMgr;
 class TMasterInfo;
 class LoadChannelMgr;
-class WebPageHandler;
 class StreamLoadExecutor;
 class RoutineLoadTaskExecutor;
 class SmallFileMgr;
-class StoragePolicyMgr;
 class BlockSpillManager;
-
 class BackendServiceClient;
 class TPaloBrokerServiceClient;
 class PBackendService_Stub;
 class PFunctionService_Stub;
-
 template <class T>
 class ClientCache;
-
 class HeartbeatFlags;
+class FrontendServiceClient;
 
 // Execution environment for queries/plan fragments.
 // Contains all required global structures, and handles to
diff --git a/be/src/runtime/exec_env_init.cpp b/be/src/runtime/exec_env_init.cpp
index e1c7cf39a9..b69b388ee1 100644
--- a/be/src/runtime/exec_env_init.cpp
+++ b/be/src/runtime/exec_env_init.cpp
@@ -15,15 +15,33 @@
 // specific language governing permissions and limitations
 // under the License.
 
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
+#include <gen_cpp/HeartbeatService_types.h>
+#include <gen_cpp/Metrics_types.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/resource.h>
+
+#include <limits>
+#include <map>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
 #include "common/config.h"
 #include "common/logging.h"
-#include "gen_cpp/BackendService.h"
-#include "gen_cpp/HeartbeatService_types.h"
-#include "gen_cpp/TPaloBrokerService.h"
+#include "common/status.h"
+#include "olap/olap_define.h"
+#include "olap/options.h"
 #include "olap/page_cache.h"
 #include "olap/rowset/segment_v2/inverted_index_cache.h"
 #include "olap/segment_loader.h"
-#include "olap/storage_engine.h"
+#include "pipeline/task_queue.h"
 #include "pipeline/task_scheduler.h"
 #include "runtime/block_spill_manager.h"
 #include "runtime/broker_mgr.h"
@@ -31,28 +49,32 @@
 #include "runtime/client_cache.h"
 #include "runtime/exec_env.h"
 #include "runtime/external_scan_context_mgr.h"
-#include "runtime/fold_constant_executor.h"
 #include "runtime/fragment_mgr.h"
 #include "runtime/heartbeat_flags.h"
 #include "runtime/load_channel_mgr.h"
 #include "runtime/load_path_mgr.h"
+#include "runtime/memory/chunk_allocator.h"
 #include "runtime/memory/mem_tracker.h"
+#include "runtime/memory/mem_tracker_limiter.h"
+#include "runtime/memory/thread_mem_tracker_mgr.h"
 #include "runtime/result_buffer_mgr.h"
 #include "runtime/result_queue_mgr.h"
 #include "runtime/routine_load/routine_load_task_executor.h"
 #include "runtime/small_file_mgr.h"
 #include "runtime/stream_load/new_load_stream_mgr.h"
 #include "runtime/stream_load/stream_load_executor.h"
+#include "runtime/thread_context.h"
 #include "service/point_query_executor.h"
 #include "util/bfd_parser.h"
+#include "util/bit_util.h"
 #include "util/brpc_client_cache.h"
+#include "util/cpu_info.h"
 #include "util/doris_metrics.h"
 #include "util/mem_info.h"
 #include "util/metrics.h"
 #include "util/parse_util.h"
 #include "util/pretty_printer.h"
-#include "util/priority_thread_pool.hpp"
-#include "util/priority_work_stealing_thread_pool.hpp"
+#include "util/threadpool.h"
 #include "vec/exec/scan/scanner_scheduler.h"
 #include "vec/runtime/vdata_stream_mgr.h"
 
@@ -62,6 +84,8 @@
 #endif
 
 namespace doris {
+class PBackendService_Stub;
+class PFunctionService_Stub;
 
 DEFINE_GAUGE_METRIC_PROTOTYPE_2ARG(scanner_thread_pool_queue_size, MetricUnit::NOUNIT);
 DEFINE_GAUGE_METRIC_PROTOTYPE_2ARG(send_batch_thread_pool_thread_num, MetricUnit::NOUNIT);
diff --git a/be/src/runtime/export_task_mgr.h b/be/src/runtime/export_task_mgr.h
index 51ea5a2d9c..90dd2017cb 100644
--- a/be/src/runtime/export_task_mgr.h
+++ b/be/src/runtime/export_task_mgr.h
@@ -17,12 +17,13 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+
 #include <mutex>
 #include <unordered_set>
 #include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/Types_types.h"
 #include "runtime/runtime_state.h"
 #include "util/lru_cache.hpp"
 
diff --git a/be/src/runtime/external_scan_context_mgr.cpp b/be/src/runtime/external_scan_context_mgr.cpp
index 8fa2183eba..9bb095d9a6 100644
--- a/be/src/runtime/external_scan_context_mgr.cpp
+++ b/be/src/runtime/external_scan_context_mgr.cpp
@@ -17,13 +17,19 @@
 
 #include "runtime/external_scan_context_mgr.h"
 
-#include <chrono>
-#include <functional>
+#include <glog/logging.h>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <ostream>
 #include <vector>
 
+#include "common/config.h"
+#include "runtime/exec_env.h"
 #include "runtime/fragment_mgr.h"
 #include "runtime/result_queue_mgr.h"
 #include "util/doris_metrics.h"
+#include "util/metrics.h"
+#include "util/thread.h"
 #include "util/uid_util.h"
 
 namespace doris {
diff --git a/be/src/runtime/external_scan_context_mgr.h b/be/src/runtime/external_scan_context_mgr.h
index 98ec60c4fe..93ae8d360a 100644
--- a/be/src/runtime/external_scan_context_mgr.h
+++ b/be/src/runtime/external_scan_context_mgr.h
@@ -17,23 +17,23 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <stdint.h>
 #include <time.h>
 
 #include <map>
 #include <memory>
 #include <mutex>
 #include <string>
-#include <thread>
 #include <utility>
 
 #include "common/status.h"
-#include "gen_cpp/Types_types.h"
 #include "gutil/ref_counted.h"
-#include "runtime/exec_env.h"
 #include "util/countdown_latch.h"
-#include "util/thread.h"
 
 namespace doris {
+class ExecEnv;
+class Thread;
 
 struct ScanContext {
 public:
diff --git a/be/src/runtime/fold_constant_executor.cpp b/be/src/runtime/fold_constant_executor.cpp
index aebda95c9f..2beb14b266 100644
--- a/be/src/runtime/fold_constant_executor.cpp
+++ b/be/src/runtime/fold_constant_executor.cpp
@@ -17,22 +17,47 @@
 
 #include "runtime/fold_constant_executor.h"
 
-#include <map>
+#include <fmt/format.h>
+#include <gen_cpp/Descriptors_types.h>
+#include <gen_cpp/Exprs_types.h>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/internal_service.pb.h>
+#include <gen_cpp/types.pb.h>
+#include <glog/logging.h>
+#include <stdint.h>
 
-#include "common/object_pool.h"
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <map>
+#include <ostream>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/status.h"
-#include "gen_cpp/PaloInternalService_types.h"
-#include "gen_cpp/internal_service.pb.h"
+#include "runtime/datetime_value.h"
+#include "runtime/decimalv2_value.h"
 #include "runtime/define_primitive_type.h"
+#include "runtime/descriptors.h"
 #include "runtime/exec_env.h"
 #include "runtime/large_int_value.h"
 #include "runtime/memory/mem_tracker.h"
+#include "runtime/primitive_type.h"
 #include "runtime/runtime_state.h"
 #include "runtime/thread_context.h"
+#include "util/binary_cast.hpp"
+#include "util/defer_op.h"
+#include "util/runtime_profile.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
 #include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/exprs/vexpr.h"
 #include "vec/exprs/vexpr_context.h"
+#include "vec/runtime/vdatetime_value.h"
 
 using std::string;
 using std::map;
diff --git a/be/src/runtime/fold_constant_executor.h b/be/src/runtime/fold_constant_executor.h
index edc1e4ada3..4a2ce4e96e 100644
--- a/be/src/runtime/fold_constant_executor.h
+++ b/be/src/runtime/fold_constant_executor.h
@@ -17,21 +17,27 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+
+#include <memory>
+#include <string>
+
 #include "common/object_pool.h"
 #include "common/status.h"
-#include "gen_cpp/PaloInternalService_types.h"
-#include "gen_cpp/internal_service.pb.h"
-#include "runtime/define_primitive_type.h"
-#include "runtime/exec_env.h"
+#include "runtime/memory/mem_tracker.h"
 #include "runtime/runtime_state.h"
-#include "util/runtime_profile.h"
+#include "runtime/types.h"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/data_types/data_type.h"
 
 namespace doris {
 
 class TFoldConstantParams;
-class TExpr;
 class TQueryGlobals;
+class PConstantExprResult;
+class RuntimeProfile;
+class TQueryOptions;
 
 // This class used to fold constant expr from fe
 class FoldConstantExecutor {
diff --git a/be/src/runtime/fragment_mgr.cpp b/be/src/runtime/fragment_mgr.cpp
index 07ef9f3c0a..a113fa3792 100644
--- a/be/src/runtime/fragment_mgr.cpp
+++ b/be/src/runtime/fragment_mgr.cpp
@@ -18,44 +18,75 @@
 #include "runtime/fragment_mgr.h"
 
 #include <bvar/latency_recorder.h>
+#include <fmt/format.h>
+#include <gen_cpp/DorisExternalService_types.h>
+#include <gen_cpp/FrontendService.h>
+#include <gen_cpp/FrontendService_types.h>
 #include <gen_cpp/HeartbeatService_types.h>
-#include <gperftools/profiler.h>
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Planner_types.h>
+#include <gen_cpp/QueryPlanExtra_types.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/internal_service.pb.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+#include <opentelemetry/trace/span.h>
+#include <opentelemetry/trace/tracer.h>
+#include <pthread.h>
+#include <stddef.h>
+#include <thrift/Thrift.h>
 #include <thrift/protocol/TDebugProtocol.h>
+#include <thrift/transport/TTransportException.h>
 
+#include <atomic>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <map>
 #include <memory>
 #include <sstream>
+#include <utility>
 
+#include "common/config.h"
+#include "common/logging.h"
 #include "common/object_pool.h"
-#include "gen_cpp/FrontendService.h"
-#include "gen_cpp/PaloInternalService_types.h"
-#include "gen_cpp/PlanNodes_types.h"
-#include "gen_cpp/QueryPlanExtra_types.h"
-#include "gen_cpp/Types_types.h"
+#include "common/utils.h"
 #include "gutil/strings/substitute.h"
 #include "io/fs/stream_load_pipe.h"
 #include "opentelemetry/trace/scope.h"
 #include "pipeline/pipeline_fragment_context.h"
 #include "pipeline/task_scheduler.h"
 #include "runtime/client_cache.h"
-#include "runtime/datetime_value.h"
 #include "runtime/descriptors.h"
 #include "runtime/exec_env.h"
+#include "runtime/memory/mem_tracker_limiter.h"
 #include "runtime/plan_fragment_executor.h"
+#include "runtime/primitive_type.h"
+#include "runtime/query_fragments_ctx.h"
 #include "runtime/runtime_filter_mgr.h"
+#include "runtime/runtime_state.h"
 #include "runtime/stream_load/new_load_stream_mgr.h"
 #include "runtime/stream_load/stream_load_context.h"
+#include "runtime/stream_load/stream_load_executor.h"
+#include "runtime/task_group/task_group.h"
 #include "runtime/task_group/task_group_manager.h"
 #include "runtime/thread_context.h"
+#include "runtime/types.h"
 #include "service/backend_options.h"
 #include "util/doris_metrics.h"
+#include "util/hash_util.hpp"
 #include "util/mem_info.h"
 #include "util/network_util.h"
-#include "util/stopwatch.hpp"
+#include "util/pretty_printer.h"
+#include "util/runtime_profile.h"
 #include "util/telemetry/telemetry.h"
+#include "util/thread.h"
 #include "util/threadpool.h"
 #include "util/thrift_util.h"
 #include "util/uid_util.h"
 #include "util/url_coding.h"
+#include "vec/runtime/shared_hash_table_controller.h"
+#include "vec/runtime/vdatetime_value.h"
 
 namespace doris {
 
@@ -78,7 +109,6 @@ std::string to_load_error_http_path(const std::string& file_name) {
 using apache::thrift::TException;
 using apache::thrift::transport::TTransportException;
 
-class RuntimeProfile;
 class FragmentExecState {
 public:
     using report_status_callback_impl = std::function<void(const ReportStatusRequest)>;
diff --git a/be/src/runtime/fragment_mgr.h b/be/src/runtime/fragment_mgr.h
index 698986f13b..d5a724787e 100644
--- a/be/src/runtime/fragment_mgr.h
+++ b/be/src/runtime/fragment_mgr.h
@@ -17,23 +17,26 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/types.pb.h>
+#include <stdint.h>
+
+#include <condition_variable>
 #include <functional>
+#include <iosfwd>
 #include <memory>
 #include <mutex>
-#include <thread>
+#include <string>
 #include <unordered_map>
 #include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/DorisExternalService_types.h"
-#include "gen_cpp/Types_types.h"
-#include "gen_cpp/internal_service.pb.h"
 #include "gutil/ref_counted.h"
 #include "http/rest_monitor_iface.h"
 #include "runtime_filter_mgr.h"
 #include "util/countdown_latch.h"
+#include "util/hash_util.hpp" // IWYU pragma: keep
 #include "util/metrics.h"
-#include "util/thread.h"
 
 namespace butil {
 class IOBufAsZeroCopyInputStream;
@@ -44,20 +47,21 @@ namespace doris {
 namespace pipeline {
 class PipelineFragmentContext;
 }
-
-namespace io {
-class StreamLoadPipe;
-}
-
 class QueryFragmentsCtx;
 class ExecEnv;
 class FragmentExecState;
-class PlanFragmentExecutor;
 class ThreadPool;
 class TExecPlanFragmentParams;
-class TExecPlanFragmentParamsList;
-class TUniqueId;
-class RuntimeFilterMergeController;
+class PExecPlanFragmentStartRequest;
+class PMergeFilterRequest;
+class PPublishFilterRequest;
+class RuntimeProfile;
+class RuntimeState;
+class TPipelineFragmentParams;
+class TPipelineInstanceParams;
+class TScanColumnDesc;
+class TScanOpenParams;
+class Thread;
 
 std::string to_load_error_http_path(const std::string& file_name);
 
diff --git a/be/src/runtime/heartbeat_flags.h b/be/src/runtime/heartbeat_flags.h
index 7412265096..77c6b942ec 100644
--- a/be/src/runtime/heartbeat_flags.h
+++ b/be/src/runtime/heartbeat_flags.h
@@ -17,9 +17,9 @@
 
 #pragma once
 
-#include <atomic>
+#include <gen_cpp/HeartbeatService_constants.h>
 
-#include "gen_cpp/HeartbeatService_constants.h"
+#include <atomic>
 
 namespace doris {
 
diff --git a/be/src/runtime/jsonb_value.cpp b/be/src/runtime/jsonb_value.cpp
index 93abcce640..62ab9102c5 100644
--- a/be/src/runtime/jsonb_value.cpp
+++ b/be/src/runtime/jsonb_value.cpp
@@ -17,9 +17,12 @@
 
 #include "runtime/jsonb_value.h"
 
-#include <cstring>
+#include <string_view>
 
 #include "util/jsonb_error.h"
+#include "util/jsonb_stream.h"
+#include "util/jsonb_utils.h"
+#include "util/jsonb_writer.h"
 
 namespace doris {
 
diff --git a/be/src/runtime/jsonb_value.h b/be/src/runtime/jsonb_value.h
index 997bca3e05..88abb379d9 100644
--- a/be/src/runtime/jsonb_value.h
+++ b/be/src/runtime/jsonb_value.h
@@ -18,14 +18,15 @@
 #ifndef DORIS_BE_RUNTIME_JSON_VALUE_H
 #define DORIS_BE_RUNTIME_JSON_VALUE_H
 
-#include "udf/udf.h"
-#include "util/cpu_info.h"
+#include <glog/logging.h>
+
+#include <cstddef>
+#include <ostream>
+#include <string>
+
+#include "common/status.h"
 #include "util/hash_util.hpp"
-#include "util/jsonb_error.h"
 #include "util/jsonb_parser_simd.h"
-#include "util/jsonb_utils.h"
-#include "util/sse_util.hpp"
-#include "vec/common/string_ref.h"
 
 namespace doris {
 
diff --git a/be/src/runtime/large_int_value.cpp b/be/src/runtime/large_int_value.cpp
index f08eff26c8..b9f354b823 100644
--- a/be/src/runtime/large_int_value.cpp
+++ b/be/src/runtime/large_int_value.cpp
@@ -19,6 +19,7 @@
 
 #include <string>
 
+#include "util/hash_util.hpp"
 #include "util/string_parser.hpp"
 
 namespace doris {
diff --git a/be/src/runtime/large_int_value.h b/be/src/runtime/large_int_value.h
index cd9fb0de53..7fa2e558ed 100644
--- a/be/src/runtime/large_int_value.h
+++ b/be/src/runtime/large_int_value.h
@@ -18,12 +18,10 @@
 #pragma once
 
 #include <fmt/format.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
+#include <stdint.h>
 
+#include <cstddef>
 #include <iostream>
-#include <sstream>
 #include <string>
 
 namespace doris {
diff --git a/be/src/runtime/load_channel.cpp b/be/src/runtime/load_channel.cpp
index de3d8d3365..3df55c9eff 100644
--- a/be/src/runtime/load_channel.cpp
+++ b/be/src/runtime/load_channel.cpp
@@ -17,11 +17,11 @@
 
 #include "runtime/load_channel.h"
 
-#include "olap/lru_cache.h"
-#include "runtime/exec_env.h"
+#include <gen_cpp/internal_service.pb.h>
+#include <glog/logging.h>
+
 #include "runtime/memory/mem_tracker.h"
 #include "runtime/tablets_channel.h"
-#include "runtime/thread_context.h"
 
 namespace doris {
 
diff --git a/be/src/runtime/load_channel.h b/be/src/runtime/load_channel.h
index 2fad48c79d..2abf822121 100644
--- a/be/src/runtime/load_channel.h
+++ b/be/src/runtime/load_channel.h
@@ -17,22 +17,31 @@
 
 #pragma once
 
+#include <stdint.h>
+#include <time.h>
+
+#include <algorithm>
+#include <atomic>
 #include <functional>
 #include <map>
+#include <memory>
 #include <mutex>
 #include <ostream>
+#include <string>
 #include <unordered_map>
 #include <unordered_set>
+#include <utility>
+#include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/internal_service.pb.h"
 #include "runtime/memory/mem_tracker.h"
 #include "runtime/tablets_channel.h"
+#include "util/spinlock.h"
 #include "util/uid_util.h"
 
 namespace doris {
 
-class Cache;
+class PTabletWriterOpenRequest;
 
 // A LoadChannel manages tablets channels for all indexes
 // corresponding to a certain load job
diff --git a/be/src/runtime/load_channel_mgr.cpp b/be/src/runtime/load_channel_mgr.cpp
index 413d7077cf..f280e87abc 100644
--- a/be/src/runtime/load_channel_mgr.cpp
+++ b/be/src/runtime/load_channel_mgr.cpp
@@ -17,20 +17,33 @@
 
 #include "runtime/load_channel_mgr.h"
 
+#include <fmt/format.h>
+#include <gen_cpp/internal_service.pb.h>
+
+#include <algorithm>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <ctime>
 #include <functional>
 #include <map>
 #include <memory>
+#include <ostream>
 #include <queue>
+#include <string>
 #include <tuple>
 #include <vector>
 
+#include "common/config.h"
+#include "common/logging.h"
+#include "runtime/exec_env.h"
 #include "runtime/load_channel.h"
 #include "runtime/memory/mem_tracker.h"
 #include "util/doris_metrics.h"
 #include "util/mem_info.h"
+#include "util/metrics.h"
 #include "util/perf_counters.h"
-#include "util/stopwatch.hpp"
-#include "util/time.h"
+#include "util/pretty_printer.h"
+#include "util/thread.h"
 
 namespace doris {
 
diff --git a/be/src/runtime/load_channel_mgr.h b/be/src/runtime/load_channel_mgr.h
index a825b74bca..69112724b5 100644
--- a/be/src/runtime/load_channel_mgr.h
+++ b/be/src/runtime/load_channel_mgr.h
@@ -17,24 +17,30 @@
 
 #pragma once
 
-#include <ctime>
+#include <stdint.h>
+
+#include <condition_variable>
 #include <memory>
 #include <mutex>
-#include <thread>
 #include <unordered_map>
+#include <utility>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/status.h"
-#include "gen_cpp/internal_service.pb.h"
 #include "gutil/ref_counted.h"
 #include "olap/lru_cache.h"
 #include "runtime/load_channel.h"
+#include "runtime/memory/mem_tracker_limiter.h"
+#include "runtime/thread_context.h"
 #include "util/countdown_latch.h"
-#include "util/thread.h"
 #include "util/uid_util.h"
 
 namespace doris {
 
-class Cache;
+class PTabletWriterCancelRequest;
+class PTabletWriterOpenRequest;
+class Thread;
 
 // LoadChannelMgr -> LoadChannel -> TabletsChannel -> DeltaWriter
 // All dispatched load data for this backend is routed from this class
diff --git a/be/src/runtime/load_path_mgr.cpp b/be/src/runtime/load_path_mgr.cpp
index 239aaffcd9..c9a5f18183 100644
--- a/be/src/runtime/load_path_mgr.cpp
+++ b/be/src/runtime/load_path_mgr.cpp
@@ -17,18 +17,31 @@
 
 #include "runtime/load_path_mgr.h"
 
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <string.h>
 #include <sys/stat.h>
-#include <sys/types.h>
-#include <unistd.h>
 
+#include <algorithm>
 #include <boost/algorithm/string/join.hpp>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <memory>
+#include <ostream>
 #include <string>
 
-#include "gen_cpp/Types_types.h"
+#include "common/config.h"
+#include "io/fs/file_system.h"
 #include "io/fs/local_file_system.h"
 #include "olap/olap_define.h"
-#include "olap/storage_engine.h"
+#include "olap/options.h"
 #include "runtime/exec_env.h"
+#include "util/thread.h"
 
 namespace doris {
 using namespace ErrorCode;
diff --git a/be/src/runtime/load_path_mgr.h b/be/src/runtime/load_path_mgr.h
index 49a9457bb1..73703517c6 100644
--- a/be/src/runtime/load_path_mgr.h
+++ b/be/src/runtime/load_path_mgr.h
@@ -17,7 +17,8 @@
 
 #pragma once
 
-#include <pthread.h>
+#include <stdint.h>
+#include <time.h>
 
 #include <mutex>
 #include <string>
@@ -25,13 +26,13 @@
 
 #include "common/status.h"
 #include "gutil/ref_counted.h"
-#include "util/thread.h"
-#include "util/uid_util.h"
+#include "util/countdown_latch.h"
 
 namespace doris {
 
 class TUniqueId;
 class ExecEnv;
+class Thread;
 
 // In every directory, '.trash' directory is used to save data need to delete
 // daemon thread is check no used directory to delete
diff --git a/be/src/runtime/map_value.h b/be/src/runtime/map_value.h
index d275316cf6..26c46692bc 100644
--- a/be/src/runtime/map_value.h
+++ b/be/src/runtime/map_value.h
@@ -17,9 +17,7 @@
 
 #pragma once
 
-#include <type_traits>
-
-#include "runtime/primitive_type.h"
+#include <stdint.h>
 
 namespace doris {
 
diff --git a/be/src/runtime/memory/chunk_allocator.cpp b/be/src/runtime/memory/chunk_allocator.cpp
index a99b8b0873..3c8a8d0946 100644
--- a/be/src/runtime/memory/chunk_allocator.cpp
+++ b/be/src/runtime/memory/chunk_allocator.cpp
@@ -17,18 +17,21 @@
 
 #include "runtime/memory/chunk_allocator.h"
 
+#include <glog/logging.h>
 #include <sanitizer/asan_interface.h>
+#include <stdlib.h>
 
-#include <list>
 #include <mutex>
 
+#include "common/config.h"
+#include "common/status.h"
 #include "runtime/memory/chunk.h"
-#include "runtime/memory/mem_tracker.h"
 #include "runtime/memory/system_allocator.h"
 #include "runtime/thread_context.h"
 #include "util/bit_util.h"
 #include "util/cpu_info.h"
 #include "util/doris_metrics.h"
+#include "util/metrics.h"
 #include "util/runtime_profile.h"
 #include "util/spinlock.h"
 
diff --git a/be/src/runtime/memory/chunk_allocator.h b/be/src/runtime/memory/chunk_allocator.h
index 8602815a57..467317c191 100644
--- a/be/src/runtime/memory/chunk_allocator.h
+++ b/be/src/runtime/memory/chunk_allocator.h
@@ -23,12 +23,13 @@
 #include <memory>
 #include <vector>
 
+#include "runtime/memory/mem_tracker_limiter.h"
+
 namespace doris {
 
 struct Chunk;
 class ChunkArena;
 class MetricEntity;
-class MemTrackerLimiter;
 class Status;
 
 // Used to allocate memory with power-of-two length.
diff --git a/be/src/runtime/memory/jemalloc_hook.cpp b/be/src/runtime/memory/jemalloc_hook.cpp
index bfc358c827..ec40879e4b 100644
--- a/be/src/runtime/memory/jemalloc_hook.cpp
+++ b/be/src/runtime/memory/jemalloc_hook.cpp
@@ -15,8 +15,15 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <features.h>
+#include <stdint.h>
+#include <stdlib.h>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "jemalloc/jemalloc.h"
 #include "runtime/thread_context.h"
+#include "util/sse_util.hpp"
 
 #ifndef __THROW
 #if __cplusplus
diff --git a/be/src/runtime/memory/mem_tracker.cpp b/be/src/runtime/memory/mem_tracker.cpp
index 78cb46d3bc..75e59385c2 100644
--- a/be/src/runtime/memory/mem_tracker.cpp
+++ b/be/src/runtime/memory/mem_tracker.cpp
@@ -22,9 +22,10 @@
 
 #include <fmt/format.h>
 
+#include <mutex>
+
+#include "runtime/memory/mem_tracker_limiter.h"
 #include "runtime/thread_context.h"
-#include "util/string_util.h"
-#include "util/time.h"
 
 namespace doris {
 
diff --git a/be/src/runtime/memory/mem_tracker.h b/be/src/runtime/memory/mem_tracker.h
index 8553ce834e..60ac996a82 100644
--- a/be/src/runtime/memory/mem_tracker.h
+++ b/be/src/runtime/memory/mem_tracker.h
@@ -19,6 +19,20 @@
 // and modified by Doris
 #pragma once
 
+#include <gen_cpp/Metrics_types.h>
+#include <stdint.h>
+
+#include <atomic>
+// IWYU pragma: no_include <bits/std_abs.h>
+#include <cmath> // IWYU pragma: keep
+#include <list>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <vector>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "util/pretty_printer.h"
 #include "util/runtime_profile.h"
 
@@ -159,4 +173,4 @@ protected:
     std::list<MemTracker*>::iterator _tracker_group_it;
 };
 
-} // namespace doris
\ No newline at end of file
+} // namespace doris
diff --git a/be/src/runtime/memory/mem_tracker_limiter.cpp b/be/src/runtime/memory/mem_tracker_limiter.cpp
index 509f190358..d9974bef7f 100644
--- a/be/src/runtime/memory/mem_tracker_limiter.cpp
+++ b/be/src/runtime/memory/mem_tracker_limiter.cpp
@@ -18,13 +18,17 @@
 #include "runtime/memory/mem_tracker_limiter.h"
 
 #include <fmt/format.h>
+#include <gen_cpp/types.pb.h>
+#include <stdlib.h>
 
-#include <boost/stacktrace.hpp>
+#include <functional>
+#include <mutex>
 #include <queue>
+#include <utility>
 
+#include "runtime/exec_env.h"
 #include "runtime/fragment_mgr.h"
 #include "runtime/load_channel_mgr.h"
-#include "runtime/runtime_state.h"
 #include "service/backend_options.h"
 #include "util/mem_info.h"
 #include "util/perf_counters.h"
diff --git a/be/src/runtime/memory/mem_tracker_limiter.h b/be/src/runtime/memory/mem_tracker_limiter.h
index e1e2dd0b14..7665a029f8 100644
--- a/be/src/runtime/memory/mem_tracker_limiter.h
+++ b/be/src/runtime/memory/mem_tracker_limiter.h
@@ -17,16 +17,30 @@
 
 #pragma once
 
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <stdint.h>
+
 #include <atomic>
+// IWYU pragma: no_include <bits/std_abs.h>
+#include <cmath> // IWYU pragma: keep
+#include <list>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <unordered_map>
+#include <vector>
 
 #include "common/config.h"
+#include "common/status.h"
 #include "runtime/memory/mem_tracker.h"
+#include "util/runtime_profile.h"
 #include "util/string_util.h"
+#include "util/uid_util.h"
 
 namespace doris {
 
-class RuntimeState;
-
 // Track and limit the memory usage of process and query.
 // Contains an limit, arranged into a tree structure.
 //
diff --git a/be/src/runtime/memory/system_allocator.cpp b/be/src/runtime/memory/system_allocator.cpp
index 34e379d9f1..fc72213cda 100644
--- a/be/src/runtime/memory/system_allocator.cpp
+++ b/be/src/runtime/memory/system_allocator.cpp
@@ -17,13 +17,17 @@
 
 #include "runtime/memory/system_allocator.h"
 
+#include <fmt/format.h>
 #include <stdlib.h>
 #include <string.h>
-#include <sys/mman.h>
 
-#include "common/config.h"
+#include <new>
+#include <string>
+
 #include "common/logging.h"
+#include "runtime/memory/mem_tracker_limiter.h"
 #include "runtime/thread_context.h"
+#include "util/sse_util.hpp"
 
 namespace doris {
 
diff --git a/be/src/runtime/memory/thread_mem_tracker_mgr.cpp b/be/src/runtime/memory/thread_mem_tracker_mgr.cpp
index b166163227..f0ffaebaad 100644
--- a/be/src/runtime/memory/thread_mem_tracker_mgr.cpp
+++ b/be/src/runtime/memory/thread_mem_tracker_mgr.cpp
@@ -17,12 +17,10 @@
 
 #include "runtime/memory/thread_mem_tracker_mgr.h"
 
-#include <chrono>
-#include <thread>
+#include <gen_cpp/types.pb.h>
 
 #include "runtime/exec_env.h"
 #include "runtime/fragment_mgr.h"
-#include "service/backend_options.h"
 
 namespace doris {
 
diff --git a/be/src/runtime/memory/thread_mem_tracker_mgr.h b/be/src/runtime/memory/thread_mem_tracker_mgr.h
index 3665ec3d53..1192bdb2a1 100644
--- a/be/src/runtime/memory/thread_mem_tracker_mgr.h
+++ b/be/src/runtime/memory/thread_mem_tracker_mgr.h
@@ -17,10 +17,18 @@
 
 #pragma once
 
-#include <bthread/bthread.h>
 #include <fmt/format.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <stdint.h>
 
-#include "gutil/macros.h"
+#include <algorithm>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <vector>
+
+#include "common/config.h"
 #include "runtime/exec_env.h"
 #include "runtime/memory/mem_tracker.h"
 #include "runtime/memory/mem_tracker_limiter.h"
diff --git a/be/src/runtime/message_body_sink.cpp b/be/src/runtime/message_body_sink.cpp
index 62aa309c24..36a498d27d 100644
--- a/be/src/runtime/message_body_sink.cpp
+++ b/be/src/runtime/message_body_sink.cpp
@@ -17,11 +17,14 @@
 
 #include "runtime/message_body_sink.h"
 
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
 #include <fcntl.h>
+#include <glog/logging.h>
+#include <string.h>
+#include <unistd.h>
 
-#include <algorithm>
-
-#include "util/runtime_profile.h"
+#include <ostream>
 
 namespace doris {
 
diff --git a/be/src/runtime/message_body_sink.h b/be/src/runtime/message_body_sink.h
index 2c4df30976..a396cc3648 100644
--- a/be/src/runtime/message_body_sink.h
+++ b/be/src/runtime/message_body_sink.h
@@ -17,6 +17,11 @@
 
 #pragma once
 
+#include <stddef.h>
+
+#include <memory>
+#include <string>
+
 #include "common/status.h"
 #include "util/byte_buffer.h"
 
diff --git a/be/src/runtime/plan_fragment_executor.cpp b/be/src/runtime/plan_fragment_executor.cpp
index 701a22601e..10747ea7d3 100644
--- a/be/src/runtime/plan_fragment_executor.cpp
+++ b/be/src/runtime/plan_fragment_executor.cpp
@@ -20,18 +20,33 @@
 
 #include "runtime/plan_fragment_executor.h"
 
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Planner_types.h>
 #include <gen_cpp/version.h>
-#include <thrift/protocol/TDebugProtocol.h>
-
-#include <unordered_map>
+#include <opentelemetry/nostd/shared_ptr.h>
+#include <opentelemetry/trace/span.h>
+#include <opentelemetry/trace/span_context.h>
+#include <opentelemetry/trace/tracer.h>
+#include <pthread.h>
+#include <stdint.h>
+#include <stdlib.h>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <ostream>
+#include <typeinfo>
+#include <utility>
 
+#include "common/config.h"
+#include "common/logging.h"
 #include "exec/data_sink.h"
 #include "exec/exec_node.h"
 #include "exec/scan_node.h"
 #include "runtime/descriptors.h"
 #include "runtime/exec_env.h"
-#include "runtime/memory/mem_tracker.h"
-#include "runtime/result_buffer_mgr.h"
+#include "runtime/memory/mem_tracker_limiter.h"
+#include "runtime/query_fragments_ctx.h"
+#include "runtime/query_statistics.h"
 #include "runtime/result_queue_mgr.h"
 #include "runtime/runtime_filter_mgr.h"
 #include "runtime/thread_context.h"
@@ -39,6 +54,8 @@
 #include "util/defer_op.h"
 #include "util/pretty_printer.h"
 #include "util/telemetry/telemetry.h"
+#include "util/threadpool.h"
+#include "util/time.h"
 #include "util/uid_util.h"
 #include "vec/core/block.h"
 #include "vec/exec/scan/new_es_scan_node.h"
@@ -47,6 +64,7 @@
 #include "vec/exec/scan/new_odbc_scan_node.h"
 #include "vec/exec/scan/new_olap_scan_node.h"
 #include "vec/exec/scan/vmeta_scan_node.h"
+#include "vec/exec/scan/vscan_node.h"
 #include "vec/exec/vexchange_node.h"
 #include "vec/runtime/vdata_stream_mgr.h"
 
diff --git a/be/src/runtime/plan_fragment_executor.h b/be/src/runtime/plan_fragment_executor.h
index 46521004fb..61f03b624d 100644
--- a/be/src/runtime/plan_fragment_executor.h
+++ b/be/src/runtime/plan_fragment_executor.h
@@ -20,17 +20,22 @@
 
 #pragma once
 
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/types.pb.h>
+
 #include <condition_variable>
 #include <functional>
 #include <future>
+#include <map>
+#include <memory>
+#include <mutex>
+#include <string>
 #include <vector>
 
-#include "common/object_pool.h"
 #include "common/status.h"
-#include "runtime/query_fragments_ctx.h"
-#include "runtime/query_statistics.h"
 #include "runtime/runtime_state.h"
-#include "vec/core/block.h"
+#include "util/runtime_profile.h"
 
 namespace doris {
 
@@ -38,14 +43,14 @@ class QueryFragmentsCtx;
 class ExecNode;
 class RowDescriptor;
 class DataSink;
-class DataStreamMgr;
-class RuntimeProfile;
-class RuntimeState;
-class TNetworkAddress;
-class TPlanExecRequest;
-class TPlanFragment;
-class TPlanFragmentExecParams;
-class TPlanExecParams;
+class DescriptorTbl;
+class ExecEnv;
+class ObjectPool;
+class QueryStatistics;
+
+namespace vectorized {
+class Block;
+} // namespace vectorized
 
 // PlanFragmentExecutor handles all aspects of the execution of a single plan fragment,
 // including setup and tear-down, both in the success and error case.
diff --git a/be/src/runtime/primitive_type.cpp b/be/src/runtime/primitive_type.cpp
index 6c3e9fb5f1..8f07934a91 100644
--- a/be/src/runtime/primitive_type.cpp
+++ b/be/src/runtime/primitive_type.cpp
@@ -17,13 +17,11 @@
 
 #include "runtime/primitive_type.h"
 
-#include "gen_cpp/Types_types.h"
-#include "runtime/collection_value.h"
+#include <gen_cpp/Types_types.h>
+
+#include <vector>
+
 #include "runtime/define_primitive_type.h"
-#include "runtime/jsonb_value.h"
-#include "runtime/map_value.h"
-#include "runtime/struct_value.h"
-#include "vec/common/string_ref.h"
 
 namespace doris {
 
diff --git a/be/src/runtime/primitive_type.h b/be/src/runtime/primitive_type.h
index d261aa4582..dfc121c330 100644
--- a/be/src/runtime/primitive_type.h
+++ b/be/src/runtime/primitive_type.h
@@ -17,13 +17,20 @@
 
 #pragma once
 
+#include <gen_cpp/Opcodes_types.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <stdint.h>
+
 #include <string>
 
-#include "gen_cpp/PaloInternalService_types.h"
+#include "olap/decimal12.h"
 #include "runtime/define_primitive_type.h"
 #include "vec/columns/column_decimal.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
 #include "vec/core/types.h"
+#include "vec/runtime/vdatetime_value.h"
 
 namespace doris {
 
diff --git a/be/src/runtime/query_fragments_ctx.h b/be/src/runtime/query_fragments_ctx.h
index bbf5be6890..a5ff251dce 100644
--- a/be/src/runtime/query_fragments_ctx.h
+++ b/be/src/runtime/query_fragments_ctx.h
@@ -17,14 +17,15 @@
 
 #pragma once
 
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/Types_types.h>
+
 #include <atomic>
 #include <memory>
 #include <string>
 
 #include "common/config.h"
 #include "common/object_pool.h"
-#include "gen_cpp/PaloInternalService_types.h" // for TQueryOptions
-#include "gen_cpp/Types_types.h"               // for TUniqueId
 #include "runtime/datetime_value.h"
 #include "runtime/exec_env.h"
 #include "runtime/memory/mem_tracker_limiter.h"
diff --git a/be/src/runtime/query_statistics.cpp b/be/src/runtime/query_statistics.cpp
index 4ef8e4a7ce..f9bd7867f4 100644
--- a/be/src/runtime/query_statistics.cpp
+++ b/be/src/runtime/query_statistics.cpp
@@ -17,6 +17,7 @@
 
 #include "runtime/query_statistics.h"
 
+#include <gen_cpp/data.pb.h>
 #include <glog/logging.h>
 
 namespace doris {
diff --git a/be/src/runtime/query_statistics.h b/be/src/runtime/query_statistics.h
index c1b2e4f028..6b5127cf4f 100644
--- a/be/src/runtime/query_statistics.h
+++ b/be/src/runtime/query_statistics.h
@@ -17,15 +17,20 @@
 
 #pragma once
 
-#include <mutex>
+#include <stdint.h>
+
+#include <map>
+#include <mutex>
+#include <unordered_map>
+#include <utility>
 
-#include "gen_cpp/data.pb.h"
 #include "util/spinlock.h"
 
 namespace doris {
 
-class QueryStatistics;
 class QueryStatisticsRecvr;
+class PNodeStatistics;
+class PQueryStatistics;
 
 class NodeStatistics {
 public:
diff --git a/be/src/runtime/record_batch_queue.cpp b/be/src/runtime/record_batch_queue.cpp
index 46cac2e316..1f76102bb9 100644
--- a/be/src/runtime/record_batch_queue.cpp
+++ b/be/src/runtime/record_batch_queue.cpp
@@ -17,6 +17,8 @@
 
 #include "runtime/record_batch_queue.h"
 
+#include "util/spinlock.h"
+
 namespace doris {
 
 void RecordBatchQueue::update_status(const Status& status) {
diff --git a/be/src/runtime/record_batch_queue.h b/be/src/runtime/record_batch_queue.h
index 008b8f10d3..7ababc9b61 100644
--- a/be/src/runtime/record_batch_queue.h
+++ b/be/src/runtime/record_batch_queue.h
@@ -17,8 +17,12 @@
 
 #pragma once
 
+#include <sys/types.h>
 #include <util/spinlock.h>
 
+#include <memory>
+#include <mutex>
+
 #include "common/status.h"
 #include "util/blocking_queue.hpp"
 
diff --git a/be/src/runtime/result_buffer_mgr.cpp b/be/src/runtime/result_buffer_mgr.cpp
index 02229efe23..14e844bfeb 100644
--- a/be/src/runtime/result_buffer_mgr.cpp
+++ b/be/src/runtime/result_buffer_mgr.cpp
@@ -17,12 +17,20 @@
 
 #include "runtime/result_buffer_mgr.h"
 
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/types.pb.h>
+#include <glog/logging.h>
+#include <stdint.h>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
 #include <memory>
+#include <ostream>
+#include <utility>
 
-#include "gen_cpp/PaloInternalService_types.h"
-#include "gen_cpp/types.pb.h"
 #include "runtime/buffer_control_block.h"
 #include "util/doris_metrics.h"
+#include "util/metrics.h"
+#include "util/thread.h"
 
 namespace doris {
 
diff --git a/be/src/runtime/result_buffer_mgr.h b/be/src/runtime/result_buffer_mgr.h
index 250551f42a..d8cf8e5787 100644
--- a/be/src/runtime/result_buffer_mgr.h
+++ b/be/src/runtime/result_buffer_mgr.h
@@ -17,24 +17,26 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+
+#include <ctime>
 #include <map>
+#include <memory>
 #include <mutex>
-#include <thread>
 #include <unordered_map>
 #include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/Types_types.h"
 #include "gutil/ref_counted.h"
+#include "util/countdown_latch.h"
 #include "util/hash_util.hpp"
-#include "util/thread.h"
 
 namespace doris {
 
-class TFetchDataResult;
 class BufferControlBlock;
 struct GetResultBatchCtx;
 class PUniqueId;
+class Thread;
 
 // manage all result buffer control block in one backend
 class ResultBufferMgr {
diff --git a/be/src/runtime/result_queue_mgr.cpp b/be/src/runtime/result_queue_mgr.cpp
index 4f30f91345..8090a3e6ee 100644
--- a/be/src/runtime/result_queue_mgr.cpp
+++ b/be/src/runtime/result_queue_mgr.cpp
@@ -17,11 +17,16 @@
 
 #include "runtime/result_queue_mgr.h"
 
+#include <gen_cpp/Types_types.h>
+
+#include <utility>
+
 #include "common/config.h"
 #include "common/status.h"
-#include "gen_cpp/Types_types.h"
-#include "runtime/exec_env.h"
+#include "runtime/record_batch_queue.h"
 #include "util/doris_metrics.h"
+#include "util/hash_util.hpp"
+#include "util/metrics.h"
 
 namespace doris {
 
diff --git a/be/src/runtime/result_queue_mgr.h b/be/src/runtime/result_queue_mgr.h
index 28dcabfaa5..d7e9f66898 100644
--- a/be/src/runtime/result_queue_mgr.h
+++ b/be/src/runtime/result_queue_mgr.h
@@ -17,15 +17,14 @@
 
 #pragma once
 
-#include <map>
+#include <gen_cpp/Types_types.h>
+
 #include <memory>
 #include <mutex>
 #include <unordered_map>
 
 #include "common/status.h"
-#include "runtime/primitive_type.h"
-#include "runtime/record_batch_queue.h"
-#include "util/hash_util.hpp"
+#include "util/hash_util.hpp" // IWYU pragma: keep
 
 namespace arrow {
 
@@ -34,10 +33,9 @@ class RecordBatch;
 
 namespace doris {
 
-class TUniqueId;
-class TScanRowBatch;
 class RecordBatchQueue;
-typedef std::shared_ptr<RecordBatchQueue> BlockQueueSharedPtr;
+
+using BlockQueueSharedPtr = std::shared_ptr<RecordBatchQueue>;
 
 class ResultQueueMgr {
 public:
diff --git a/be/src/runtime/result_writer.h b/be/src/runtime/result_writer.h
index bd5a11cc8f..7298cce13b 100644
--- a/be/src/runtime/result_writer.h
+++ b/be/src/runtime/result_writer.h
@@ -17,18 +17,16 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include <string>
+
 #include "common/status.h"
-#include "gen_cpp/PlanNodes_types.h"
 
 namespace doris {
 
-class Status;
 class RuntimeState;
 
-namespace vectorized {
-class Block;
-}
-
 // abstract class of the result writer
 class ResultWriter {
 public:
diff --git a/be/src/runtime/routine_load/data_consumer.cpp b/be/src/runtime/routine_load/data_consumer.cpp
index 88c00044e1..cdbff1c1d1 100644
--- a/be/src/runtime/routine_load/data_consumer.cpp
+++ b/be/src/runtime/routine_load/data_consumer.cpp
@@ -17,18 +17,28 @@
 
 #include "runtime/routine_load/data_consumer.h"
 
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/internal_service.pb.h>
+#include <librdkafka/rdkafkacpp.h>
+
 #include <algorithm>
-#include <functional>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
 #include <string>
+#include <thread>
+#include <utility>
 #include <vector>
 
+#include "common/config.h"
 #include "common/status.h"
-#include "gen_cpp/internal_service.pb.h"
 #include "gutil/strings/split.h"
+#include "runtime/exec_env.h"
 #include "runtime/small_file_mgr.h"
 #include "service/backend_options.h"
+#include "util/blocking_queue.hpp"
 #include "util/defer_op.h"
 #include "util/stopwatch.hpp"
+#include "util/string_util.h"
 #include "util/uid_util.h"
 
 namespace doris {
diff --git a/be/src/runtime/routine_load/data_consumer.h b/be/src/runtime/routine_load/data_consumer.h
index 4ad3d6821f..2a3f65f199 100644
--- a/be/src/runtime/routine_load/data_consumer.h
+++ b/be/src/runtime/routine_load/data_consumer.h
@@ -17,19 +17,27 @@
 
 #pragma once
 
-#include <ctime>
-#include <mutex>
-#include <unordered_map>
+#include <stdint.h>
 
+#include <ctime>
+#include <map>
+#include <memory>
+#include <mutex>
+#include <ostream>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "common/logging.h"
+#include "common/status.h"
 #include "librdkafka/rdkafkacpp.h"
 #include "runtime/stream_load/stream_load_context.h"
-#include "util/blocking_queue.hpp"
 #include "util/uid_util.h"
 
 namespace doris {
 
-class KafkaConsumerPipe;
-class Status;
+template <typename T>
+class BlockingQueue;
 
 class DataConsumer {
 public:
@@ -77,6 +85,7 @@ protected:
 };
 
 class PIntegerPair;
+
 class KafkaEventCb : public RdKafka::EventCb {
 public:
     void event_cb(RdKafka::Event& event) {
diff --git a/be/src/runtime/routine_load/data_consumer_group.cpp b/be/src/runtime/routine_load/data_consumer_group.cpp
index 2d22fa93ab..d27c86fdf0 100644
--- a/be/src/runtime/routine_load/data_consumer_group.cpp
+++ b/be/src/runtime/routine_load/data_consumer_group.cpp
@@ -16,11 +16,20 @@
 // under the License.
 #include "runtime/routine_load/data_consumer_group.h"
 
+#include <gen_cpp/PlanNodes_types.h>
+#include <stddef.h>
+
+#include <map>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/logging.h"
 #include "io/fs/kafka_consumer_pipe.h"
-#include "librdkafka/rdkafka.h"
 #include "librdkafka/rdkafkacpp.h"
 #include "runtime/routine_load/data_consumer.h"
 #include "runtime/stream_load/stream_load_context.h"
+#include "util/stopwatch.hpp"
 
 namespace doris {
 
diff --git a/be/src/runtime/routine_load/data_consumer_group.h b/be/src/runtime/routine_load/data_consumer_group.h
index bd866f0468..6ede3d9f66 100644
--- a/be/src/runtime/routine_load/data_consumer_group.h
+++ b/be/src/runtime/routine_load/data_consumer_group.h
@@ -17,11 +17,25 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include <functional>
+#include <memory>
+#include <mutex>
+#include <vector>
+
+#include "common/status.h"
 #include "runtime/routine_load/data_consumer.h"
 #include "util/blocking_queue.hpp"
 #include "util/priority_thread_pool.hpp"
+#include "util/uid_util.h"
+
+namespace RdKafka {
+class Message;
+} // namespace RdKafka
 
 namespace doris {
+class StreamLoadContext;
 
 // data consumer group saves a group of data consumers.
 // These data consumers share the same stream load pipe.
diff --git a/be/src/runtime/routine_load/data_consumer_pool.cpp b/be/src/runtime/routine_load/data_consumer_pool.cpp
index 579edfd201..c79e25634d 100644
--- a/be/src/runtime/routine_load/data_consumer_pool.cpp
+++ b/be/src/runtime/routine_load/data_consumer_pool.cpp
@@ -17,8 +17,24 @@
 
 #include "runtime/routine_load/data_consumer_pool.h"
 
+#include <gen_cpp/Types_types.h>
+
+#include <algorithm>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <ctime>
+#include <iterator>
+#include <map>
+#include <ostream>
+#include <vector>
+
 #include "common/config.h"
+#include "common/logging.h"
+#include "common/status.h"
+#include "runtime/routine_load/data_consumer.h"
 #include "runtime/routine_load/data_consumer_group.h"
+#include "runtime/stream_load/stream_load_context.h"
+#include "util/uid_util.h"
 
 namespace doris {
 
diff --git a/be/src/runtime/routine_load/data_consumer_pool.h b/be/src/runtime/routine_load/data_consumer_pool.h
index 6f8cf79f74..15baa2be6e 100644
--- a/be/src/runtime/routine_load/data_consumer_pool.h
+++ b/be/src/runtime/routine_load/data_consumer_pool.h
@@ -17,14 +17,14 @@
 
 #pragma once
 
-#include <ctime>
+#include <stdint.h>
+
+#include <list>
 #include <memory>
 #include <mutex>
 
 #include "gutil/ref_counted.h"
-#include "runtime/routine_load/data_consumer.h"
 #include "util/countdown_latch.h"
-#include "util/lru_cache.hpp"
 #include "util/thread.h"
 
 namespace doris {
@@ -32,6 +32,7 @@ namespace doris {
 class DataConsumer;
 class DataConsumerGroup;
 class Status;
+class StreamLoadContext;
 
 // DataConsumerPool saves all available data consumer
 // to be reused
diff --git a/be/src/runtime/routine_load/routine_load_task_executor.cpp b/be/src/runtime/routine_load/routine_load_task_executor.cpp
index 3aa361d157..4db6d42e18 100644
--- a/be/src/runtime/routine_load/routine_load_task_executor.cpp
+++ b/be/src/runtime/routine_load/routine_load_task_executor.cpp
@@ -17,18 +17,43 @@
 
 #include "runtime/routine_load/routine_load_task_executor.h"
 
-#include <thread>
+#include <gen_cpp/BackendService_types.h>
+#include <gen_cpp/FrontendService_types.h>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/Status_types.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/internal_service.pb.h>
+#include <librdkafka/rdkafkacpp.h>
+#include <stddef.h>
 
+#include <algorithm>
+#include <future>
+#include <map>
+#include <ostream>
+#include <thread>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/config.h"
+#include "common/logging.h"
 #include "common/status.h"
-#include "gen_cpp/BackendService_types.h"
-#include "gen_cpp/FrontendService_types.h"
-#include "gen_cpp/Types_types.h"
+#include "common/utils.h"
 #include "io/fs/kafka_consumer_pipe.h"
-#include "olap/iterators.h"
+#include "io/fs/stream_load_pipe.h"
 #include "runtime/exec_env.h"
+#include "runtime/message_body_sink.h"
+#include "runtime/routine_load/data_consumer.h"
 #include "runtime/routine_load/data_consumer_group.h"
+#include "runtime/stream_load/new_load_stream_mgr.h"
 #include "runtime/stream_load/stream_load_context.h"
+#include "runtime/stream_load/stream_load_executor.h"
+#include "service/backend_options.h"
 #include "util/defer_op.h"
+#include "util/doris_metrics.h"
+#include "util/metrics.h"
+#include "util/slice.h"
+#include "util/time.h"
 #include "util/uid_util.h"
 
 namespace doris {
diff --git a/be/src/runtime/routine_load/routine_load_task_executor.h b/be/src/runtime/routine_load/routine_load_task_executor.h
index 8b46e7bd8d..b2ddfae2b7 100644
--- a/be/src/runtime/routine_load/routine_load_task_executor.h
+++ b/be/src/runtime/routine_load/routine_load_task_executor.h
@@ -17,14 +17,16 @@
 
 #pragma once
 
+#include <stdint.h>
+
 #include <functional>
-#include <map>
 #include <memory>
 #include <mutex>
+#include <string>
+#include <unordered_map>
+#include <vector>
 
-#include "gen_cpp/internal_service.pb.h"
 #include "runtime/routine_load/data_consumer_pool.h"
-#include "util/doris_metrics.h"
 #include "util/priority_thread_pool.hpp"
 #include "util/uid_util.h"
 
@@ -34,6 +36,8 @@ class ExecEnv;
 class Status;
 class StreamLoadContext;
 class TRoutineLoadTask;
+class PIntegerPair;
+class PKafkaMetaProxyRequest;
 
 // A routine load task executor will receive routine load
 // tasks from FE, put it to a fixed thread pool.
diff --git a/be/src/runtime/runtime_filter_mgr.cpp b/be/src/runtime/runtime_filter_mgr.cpp
index b3ffd54032..187948d1a5 100644
--- a/be/src/runtime/runtime_filter_mgr.cpp
+++ b/be/src/runtime/runtime_filter_mgr.cpp
@@ -17,16 +17,26 @@
 
 #include "runtime/runtime_filter_mgr.h"
 
-#include <string>
+#include <brpc/controller.h>
+#include <butil/iobuf.h>
+#include <butil/iobuf_inl.h>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/internal_service.pb.h>
+#include <gen_cpp/types.pb.h>
+#include <stddef.h>
 
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/logging.h"
 #include "exprs/bloom_filter_func.h"
 #include "exprs/runtime_filter.h"
-#include "gen_cpp/internal_service.pb.h"
 #include "runtime/exec_env.h"
 #include "runtime/memory/mem_tracker.h"
 #include "runtime/runtime_state.h"
 #include "runtime/thread_context.h"
-#include "service/brpc.h"
 #include "util/brpc_client_cache.h"
 
 namespace doris {
diff --git a/be/src/runtime/runtime_filter_mgr.h b/be/src/runtime/runtime_filter_mgr.h
index 8d2a19a5ff..6bf9454e32 100644
--- a/be/src/runtime/runtime_filter_mgr.h
+++ b/be/src/runtime/runtime_filter_mgr.h
@@ -17,19 +17,22 @@
 
 #pragma once
 
-#include <condition_variable>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Types_types.h>
+#include <stdint.h>
+
+#include <functional>
 #include <map>
 #include <memory>
 #include <mutex>
-#include <thread>
+#include <string>
 #include <unordered_map>
+#include <unordered_set>
+#include <vector>
 
 #include "common/object_pool.h"
 #include "common/status.h"
-#include "exprs/runtime_filter.h"
-#include "gen_cpp/PaloInternalService_types.h"
-#include "gen_cpp/PlanNodes_types.h"
-#include "runtime/runtime_state.h"
 #include "util/uid_util.h"
 
 namespace butil {
@@ -37,12 +40,12 @@ class IOBufAsZeroCopyInputStream;
 }
 
 namespace doris {
-class TUniqueId;
-class RuntimeFilter;
-class FragmentExecState;
-class PlanFragmentExecutor;
 class PPublishFilterRequest;
 class PMergeFilterRequest;
+class IRuntimeFilter;
+class MemTracker;
+class RuntimeState;
+enum class RuntimeFilterRole;
 
 /// producer:
 /// Filter filter;
diff --git a/be/src/runtime/runtime_predicate.cpp b/be/src/runtime/runtime_predicate.cpp
index 52be66dfe3..d2ee35c67e 100644
--- a/be/src/runtime/runtime_predicate.cpp
+++ b/be/src/runtime/runtime_predicate.cpp
@@ -17,7 +17,12 @@
 
 #include "runtime/runtime_predicate.h"
 
+#include <stdint.h>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "olap/accept_null_predicate.h"
+#include "olap/column_predicate.h"
 #include "olap/predicate_creator.h"
 
 namespace doris {
diff --git a/be/src/runtime/runtime_predicate.h b/be/src/runtime/runtime_predicate.h
index 0d2b8e23c2..9dd48279ac 100644
--- a/be/src/runtime/runtime_predicate.h
+++ b/be/src/runtime/runtime_predicate.h
@@ -17,30 +17,28 @@
 
 #pragma once
 
-#include <atomic>
+#include <functional>
+#include <memory>
 #include <mutex>
 #include <shared_mutex>
 #include <string>
 
-#include "common/config.h"
-#include "common/object_pool.h"
+#include "common/status.h"
 #include "exec/olap_common.h"
-#include "gen_cpp/PaloInternalService_types.h" // for TQueryOptions
-#include "gen_cpp/Types_types.h"               // for TUniqueId
-#include "olap/column_predicate.h"
 #include "olap/tablet_schema.h"
-#include "runtime/datetime_value.h"
-#include "runtime/exec_env.h"
-#include "util/threadpool.h"
+#include "runtime/define_primitive_type.h"
+#include "runtime/primitive_type.h"
+#include "util/binary_cast.hpp"
+#include "vec/common/arena.h"
 #include "vec/core/field.h"
 #include "vec/core/types.h"
+#include "vec/runtime/vdatetime_value.h"
 
 namespace doris {
+class ColumnPredicate;
 
 namespace vectorized {
 
-class Arena;
-
 class RuntimePredicate {
 public:
     RuntimePredicate() = default;
diff --git a/be/src/runtime/runtime_state.cpp b/be/src/runtime/runtime_state.cpp
index 052daa7515..69a1a560d9 100644
--- a/be/src/runtime/runtime_state.cpp
+++ b/be/src/runtime/runtime_state.cpp
@@ -21,22 +21,24 @@
 #include "runtime/runtime_state.h"
 
 #include <fmt/format.h>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/Types_types.h>
 
-#include <boost/algorithm/string/join.hpp>
-#include <sstream>
 #include <string>
 
+#include "common/config.h"
 #include "common/logging.h"
 #include "common/object_pool.h"
 #include "common/status.h"
-#include "exec/exec_node.h"
 #include "runtime/exec_env.h"
 #include "runtime/load_path_mgr.h"
-#include "runtime/memory/mem_tracker.h"
+#include "runtime/memory/mem_tracker_limiter.h"
+#include "runtime/memory/thread_mem_tracker_mgr.h"
 #include "runtime/runtime_filter_mgr.h"
-#include "util/pretty_printer.h"
+#include "runtime/thread_context.h"
 #include "util/timezone_utils.h"
 #include "util/uid_util.h"
+#include "vec/runtime/vdatetime_value.h"
 
 namespace doris {
 using namespace ErrorCode;
diff --git a/be/src/runtime/runtime_state.h b/be/src/runtime/runtime_state.h
index ae98d7e68a..105acf5918 100644
--- a/be/src/runtime/runtime_state.h
+++ b/be/src/runtime/runtime_state.h
@@ -20,14 +20,24 @@
 
 #pragma once
 
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/segment_v2.pb.h>
+#include <stdint.h>
+
+#include <atomic>
 #include <fstream>
+#include <functional>
+#include <memory>
+#include <mutex>
+#include <string>
+#include <utility>
+#include <vector>
 
 #include "cctz/time_zone.h"
-#include "common/global_types.h"
-#include "common/object_pool.h"
-#include "gen_cpp/PaloInternalService_types.h" // for TQueryOptions
-#include "gen_cpp/Types_types.h"               // for TUniqueId
-#include "runtime/query_fragments_ctx.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/status.h"
 #include "util/runtime_profile.h"
 #include "util/telemetry/telemetry.h"
 
@@ -35,15 +45,10 @@ namespace doris {
 
 class DescriptorTbl;
 class ObjectPool;
-class Status;
 class ExecEnv;
-class DateTimeValue;
-class MemTracker;
-class DataStreamRecvr;
-class ResultBufferMgr;
-class BufferedBlockMgr;
-class RowDescriptor;
 class RuntimeFilterMgr;
+class MemTrackerLimiter;
+class QueryFragmentsCtx;
 
 // A collection of items that are part of the global state of a
 // query and shared across all execution nodes of that query.
diff --git a/be/src/runtime/small_file_mgr.cpp b/be/src/runtime/small_file_mgr.cpp
index cfa367bbc4..19a05f9c9d 100644
--- a/be/src/runtime/small_file_mgr.cpp
+++ b/be/src/runtime/small_file_mgr.cpp
@@ -17,19 +17,29 @@
 
 #include "runtime/small_file_mgr.h"
 
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
+#include <gen_cpp/HeartbeatService_types.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
 #include <stdint.h>
 #include <stdio.h>
 
+#include <cstring>
+#include <memory>
 #include <sstream>
+#include <utility>
+#include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/HeartbeatService.h"
 #include "gutil/strings/split.h"
 #include "http/http_client.h"
+#include "io/fs/file_system.h"
 #include "io/fs/local_file_system.h"
 #include "runtime/exec_env.h"
 #include "util/doris_metrics.h"
 #include "util/md5.h"
+#include "util/metrics.h"
 #include "util/string_util.h"
 
 namespace doris {
diff --git a/be/src/runtime/small_file_mgr.h b/be/src/runtime/small_file_mgr.h
index 89e0a85d89..74665a7655 100644
--- a/be/src/runtime/small_file_mgr.h
+++ b/be/src/runtime/small_file_mgr.h
@@ -22,12 +22,8 @@
 #include <mutex>
 #include <string>
 #include <unordered_map>
-#include <utility>
-#include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/Types_types.h"
-#include "runtime/client_cache.h"
 
 namespace doris {
 
diff --git a/be/src/runtime/snapshot_loader.cpp b/be/src/runtime/snapshot_loader.cpp
index 50e0be587e..4db803b10e 100644
--- a/be/src/runtime/snapshot_loader.cpp
+++ b/be/src/runtime/snapshot_loader.cpp
@@ -17,25 +17,43 @@
 
 #include "runtime/snapshot_loader.h"
 
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
+#include <fmt/format.h>
+#include <gen_cpp/FrontendService.h>
+#include <gen_cpp/FrontendService_types.h>
+#include <gen_cpp/HeartbeatService_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Status_types.h>
+#include <gen_cpp/Types_types.h>
 #include <stdint.h>
+#include <stdio.h>
+#include <unistd.h>
+
+#include <algorithm>
+#include <cstring>
+#include <filesystem>
+#include <istream>
+#include <utility>
 
 #include "common/logging.h"
-#include "gen_cpp/FrontendService.h"
-#include "gen_cpp/FrontendService_types.h"
-#include "gen_cpp/HeartbeatService_types.h"
-#include "gen_cpp/PaloBrokerService_types.h"
-#include "gen_cpp/TPaloBrokerService.h"
 #include "io/fs/broker_file_system.h"
+#include "io/fs/file_system.h"
 #include "io/fs/hdfs_file_system.h"
 #include "io/fs/local_file_system.h"
+#include "io/fs/path.h"
+#include "io/fs/remote_file_system.h"
 #include "io/fs/s3_file_system.h"
 #include "io/hdfs_builder.h"
+#include "olap/data_dir.h"
 #include "olap/snapshot_manager.h"
 #include "olap/storage_engine.h"
 #include "olap/tablet.h"
-#include "runtime/broker_mgr.h"
+#include "olap/tablet_manager.h"
+#include "runtime/client_cache.h"
 #include "runtime/exec_env.h"
 #include "util/s3_uri.h"
+#include "util/s3_util.h"
 #include "util/thrift_rpc_helper.h"
 
 namespace doris {
diff --git a/be/src/runtime/snapshot_loader.h b/be/src/runtime/snapshot_loader.h
index 3a5ea5925b..9e7a22d7a3 100644
--- a/be/src/runtime/snapshot_loader.h
+++ b/be/src/runtime/snapshot_loader.h
@@ -17,19 +17,21 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
 #include <stdint.h>
 
 #include <map>
+#include <memory>
 #include <string>
 #include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/Types_types.h"
-#include "io/fs/remote_file_system.h"
 #include "olap/tablet.h"
-#include "runtime/client_cache.h"
 
 namespace doris {
+namespace io {
+class RemoteFileSystem;
+} // namespace io
 
 struct FileStat {
     std::string name;
@@ -37,7 +39,6 @@ struct FileStat {
     int64_t size;
 };
 class ExecEnv;
-class StorageBackend;
 
 /*
  * Upload:
diff --git a/be/src/runtime/stream_load/new_load_stream_mgr.cpp b/be/src/runtime/stream_load/new_load_stream_mgr.cpp
index c7a57362fe..e5428bc21d 100644
--- a/be/src/runtime/stream_load/new_load_stream_mgr.cpp
+++ b/be/src/runtime/stream_load/new_load_stream_mgr.cpp
@@ -17,6 +17,9 @@
 
 #include "runtime/stream_load/new_load_stream_mgr.h"
 
+#include "util/doris_metrics.h"
+#include "util/metrics.h"
+
 namespace doris {
 
 DEFINE_GAUGE_METRIC_PROTOTYPE_2ARG(new_stream_load_pipe_count, MetricUnit::NOUNIT);
diff --git a/be/src/runtime/stream_load/new_load_stream_mgr.h b/be/src/runtime/stream_load/new_load_stream_mgr.h
index bb245232dc..65425a70cd 100644
--- a/be/src/runtime/stream_load/new_load_stream_mgr.h
+++ b/be/src/runtime/stream_load/new_load_stream_mgr.h
@@ -17,17 +17,21 @@
 
 #pragma once
 
+#include <iterator>
 #include <memory>
 #include <mutex>
+#include <ostream>
 #include <unordered_map>
+#include <utility>
 
+#include "common/logging.h"
 #include "common/status.h"
-#include "util/doris_metrics.h"
 #include "util/uid_util.h"
 
 namespace doris {
 
 class StreamLoadContext;
+
 // used to register all streams in process so that other module can get this stream
 class NewLoadStreamMgr {
 public:
diff --git a/be/src/runtime/stream_load/stream_load_context.cpp b/be/src/runtime/stream_load/stream_load_context.cpp
index 1f3ecd4372..f381ba097d 100644
--- a/be/src/runtime/stream_load/stream_load_context.cpp
+++ b/be/src/runtime/stream_load/stream_load_context.cpp
@@ -17,8 +17,19 @@
 
 #include "runtime/stream_load/stream_load_context.h"
 
+#include <gen_cpp/BackendService_types.h>
+#include <rapidjson/document.h>
+#include <rapidjson/encodings.h>
+#include <rapidjson/prettywriter.h>
+#include <rapidjson/rapidjson.h>
+#include <rapidjson/stringbuffer.h>
+#include <rapidjson/writer.h>
+
+#include <new>
 #include <sstream>
 
+#include "common/logging.h"
+
 namespace doris {
 using namespace ErrorCode;
 
diff --git a/be/src/runtime/stream_load/stream_load_context.h b/be/src/runtime/stream_load/stream_load_context.h
index d12148b527..0fa1d1f0e6 100644
--- a/be/src/runtime/stream_load/stream_load_context.h
+++ b/be/src/runtime/stream_load/stream_load_context.h
@@ -17,27 +17,31 @@
 
 #pragma once
 
-#include <rapidjson/prettywriter.h>
+#include <gen_cpp/BackendService_types.h>
+#include <gen_cpp/FrontendService_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+#include <stdint.h>
 
 #include <future>
-#include <sstream>
+#include <map>
+#include <memory>
+#include <string>
+#include <utility>
+#include <vector>
 
-#include "common/logging.h"
 #include "common/status.h"
 #include "common/utils.h"
-#include "gen_cpp/BackendService_types.h"
-#include "gen_cpp/FrontendService_types.h"
-#include "io/fs/stream_load_pipe.h"
 #include "runtime/exec_env.h"
-#include "runtime/message_body_sink.h"
-#include "runtime/stream_load/new_load_stream_mgr.h"
 #include "runtime/stream_load/stream_load_executor.h"
-#include "service/backend_options.h"
-#include "util/string_util.h"
 #include "util/time.h"
 #include "util/uid_util.h"
 
 namespace doris {
+namespace io {
+class StreamLoadPipe;
+} // namespace io
 
 // kafka related info
 class KafkaLoadInfo {
diff --git a/be/src/runtime/stream_load/stream_load_executor.cpp b/be/src/runtime/stream_load/stream_load_executor.cpp
index 9fe991f153..ba1c1f4517 100644
--- a/be/src/runtime/stream_load/stream_load_executor.cpp
+++ b/be/src/runtime/stream_load/stream_load_executor.cpp
@@ -18,16 +18,34 @@
 #include "runtime/stream_load/stream_load_executor.h"
 
 #include <gen_cpp/FrontendService.h>
+#include <gen_cpp/FrontendService_types.h>
 #include <gen_cpp/HeartbeatService_types.h>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <stdint.h>
 
+#include <future>
+#include <map>
+#include <ostream>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "common/config.h"
 #include "common/status.h"
 #include "common/utils.h"
 #include "runtime/client_cache.h"
 #include "runtime/exec_env.h"
 #include "runtime/fragment_mgr.h"
+#include "runtime/message_body_sink.h"
+#include "runtime/runtime_state.h"
+#include "runtime/stream_load/new_load_stream_mgr.h"
 #include "runtime/stream_load/stream_load_context.h"
 #include "util/doris_metrics.h"
 #include "util/thrift_rpc_helper.h"
+#include "util/time.h"
+#include "util/uid_util.h"
 
 namespace doris {
 using namespace ErrorCode;
diff --git a/be/src/runtime/stream_load/stream_load_executor.h b/be/src/runtime/stream_load/stream_load_executor.h
index b90cef0a06..fc9e6b17e4 100644
--- a/be/src/runtime/stream_load/stream_load_executor.h
+++ b/be/src/runtime/stream_load/stream_load_executor.h
@@ -19,14 +19,13 @@
 
 #include <memory>
 
-#include "gen_cpp/FrontendService_types.h"
-
 namespace doris {
 
 class ExecEnv;
 class StreamLoadContext;
 class Status;
 class TTxnCommitAttachment;
+class TLoadTxnCommitRequest;
 
 class StreamLoadExecutor {
 public:
diff --git a/be/src/runtime/stream_load/stream_load_recorder.cpp b/be/src/runtime/stream_load/stream_load_recorder.cpp
index e1d082ed18..bb055e43bd 100644
--- a/be/src/runtime/stream_load/stream_load_recorder.cpp
+++ b/be/src/runtime/stream_load/stream_load_recorder.cpp
@@ -17,12 +17,19 @@
 
 #include "runtime/stream_load/stream_load_recorder.h"
 
+#include <glog/logging.h>
+#include <rocksdb/iterator.h>
+#include <rocksdb/status.h>
+
+#include <algorithm>
+#include <memory>
+#include <ostream>
+
 #include "common/config.h"
 #include "common/status.h"
 #include "rocksdb/db.h"
 #include "rocksdb/options.h"
 #include "rocksdb/slice.h"
-#include "rocksdb/slice_transform.h"
 #include "rocksdb/utilities/db_ttl.h"
 #include "util/time.h"
 
diff --git a/be/src/runtime/stream_load/stream_load_recorder.h b/be/src/runtime/stream_load/stream_load_recorder.h
index 9d80360b91..5d446dd2af 100644
--- a/be/src/runtime/stream_load/stream_load_recorder.h
+++ b/be/src/runtime/stream_load/stream_load_recorder.h
@@ -15,11 +15,17 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <stdint.h>
+
+#include <atomic>
 #include <map>
 #include <string>
+#include <vector>
 
-#include "olap/utils.h"
-#include "rocksdb/utilities/db_ttl.h"
+namespace rocksdb {
+class ColumnFamilyHandle;
+class DBWithTTL;
+} // namespace rocksdb
 
 #pragma once
 
diff --git a/be/src/runtime/struct_value.h b/be/src/runtime/struct_value.h
index d02ddc994c..cb29d74b4c 100644
--- a/be/src/runtime/struct_value.h
+++ b/be/src/runtime/struct_value.h
@@ -17,9 +17,7 @@
 
 #pragma once
 
-#include <type_traits>
-
-#include "runtime/primitive_type.h"
+#include <stdint.h>
 
 namespace doris {
 
diff --git a/be/src/runtime/tablets_channel.cpp b/be/src/runtime/tablets_channel.cpp
index 57f795f912..b7f6f24953 100644
--- a/be/src/runtime/tablets_channel.cpp
+++ b/be/src/runtime/tablets_channel.cpp
@@ -17,14 +17,32 @@
 
 #include "runtime/tablets_channel.h"
 
+#include <fmt/format.h>
+#include <gen_cpp/internal_service.pb.h>
+#include <gen_cpp/types.pb.h>
+#include <time.h>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <initializer_list>
+#include <set>
+#include <thread>
+#include <utility>
+
+#include "common/logging.h"
 #include "exec/tablet_info.h"
 #include "olap/delta_writer.h"
-#include "olap/memtable.h"
 #include "olap/storage_engine.h"
+#include "olap/txn_manager.h"
 #include "runtime/load_channel.h"
 #include "util/doris_metrics.h"
+#include "util/metrics.h"
+#include "vec/core/block.h"
 
 namespace doris {
+class SlotDescriptor;
 
 DEFINE_GAUGE_METRIC_PROTOTYPE_2ARG(tablet_writer_count, MetricUnit::NOUNIT);
 
diff --git a/be/src/runtime/tablets_channel.h b/be/src/runtime/tablets_channel.h
index 3a29ddd620..954f1543b0 100644
--- a/be/src/runtime/tablets_channel.h
+++ b/be/src/runtime/tablets_channel.h
@@ -17,23 +17,44 @@
 
 #pragma once
 
+#include <glog/logging.h>
+
+#include <atomic>
 #include <cstdint>
 #include <functional>
 #include <map>
+#include <mutex>
+#include <ostream>
 #include <shared_mutex>
+#include <string>
 #include <unordered_map>
 #include <unordered_set>
-#include <utility>
 #include <vector>
 
-#include "gen_cpp/PaloInternalService_types.h"
-#include "gen_cpp/Types_types.h"
-#include "gen_cpp/internal_service.pb.h"
-#include "runtime/descriptors.h"
+#include "common/status.h"
 #include "util/bitmap.h"
+#include "util/spinlock.h"
 #include "util/uid_util.h"
 
+namespace google {
+namespace protobuf {
+template <typename Element>
+class RepeatedField;
+template <typename Key, typename T>
+class Map;
+template <typename T>
+class RepeatedPtrField;
+} // namespace protobuf
+} // namespace google
+
 namespace doris {
+class PSlaveTabletNodes;
+class PSuccessSlaveTabletNodeIds;
+class PTabletError;
+class PTabletInfo;
+class PTabletWriterOpenRequest;
+class PUniqueId;
+class TupleDescriptor;
 
 struct TabletsChannelKey {
     UniqueId id;
diff --git a/be/src/runtime/task_group/task_group.cpp b/be/src/runtime/task_group/task_group.cpp
index 783470be45..ea3a453d9c 100644
--- a/be/src/runtime/task_group/task_group.cpp
+++ b/be/src/runtime/task_group/task_group.cpp
@@ -17,10 +17,16 @@
 
 #include "task_group.h"
 
-#include <charconv>
+#include <fmt/format.h>
+#include <gen_cpp/PaloInternalService_types.h>
 
-#include "gen_cpp/PaloInternalService_types.h"
-#include "pipeline/pipeline_task.h"
+#include <charconv>
+#include <map>
+#include <mutex>
+#include <ostream>
+#include <utility>
+
+#include "common/logging.h"
 
 namespace doris {
 namespace taskgroup {
diff --git a/be/src/runtime/task_group/task_group.h b/be/src/runtime/task_group/task_group.h
index d6eed42a19..9a3aa117c7 100644
--- a/be/src/runtime/task_group/task_group.h
+++ b/be/src/runtime/task_group/task_group.h
@@ -17,11 +17,16 @@
 
 #pragma once
 
+#include <stddef.h>
+#include <stdint.h>
+
 #include <atomic>
+#include <memory>
 #include <queue>
 #include <shared_mutex>
+#include <string>
 
-#include "olap/olap_define.h"
+#include "common/status.h"
 
 namespace doris {
 
@@ -29,7 +34,6 @@ namespace pipeline {
 class PipelineTask;
 }
 
-class QueryFragmentsCtx;
 class TPipelineResourceGroup;
 
 namespace taskgroup {
diff --git a/be/src/runtime/task_group/task_group_manager.cpp b/be/src/runtime/task_group/task_group_manager.cpp
index cb7de92a6e..d393264983 100644
--- a/be/src/runtime/task_group/task_group_manager.cpp
+++ b/be/src/runtime/task_group/task_group_manager.cpp
@@ -17,6 +17,11 @@
 
 #include "task_group_manager.h"
 
+#include <memory>
+#include <mutex>
+
+#include "runtime/task_group/task_group.h"
+
 namespace doris::taskgroup {
 
 TaskGroupManager::TaskGroupManager() = default;
diff --git a/be/src/runtime/task_group/task_group_manager.h b/be/src/runtime/task_group/task_group_manager.h
index d5e32ac64e..0f415498a0 100644
--- a/be/src/runtime/task_group/task_group_manager.h
+++ b/be/src/runtime/task_group/task_group_manager.h
@@ -16,6 +16,8 @@
 // under the License.
 #pragma once
 
+#include <stdint.h>
+
 #include <shared_mutex>
 #include <unordered_map>
 
diff --git a/be/src/runtime/thread_context.cpp b/be/src/runtime/thread_context.cpp
index 9e0fbf50f4..2e8845b206 100644
--- a/be/src/runtime/thread_context.cpp
+++ b/be/src/runtime/thread_context.cpp
@@ -19,8 +19,11 @@
 
 #include "common/signal_handler.h"
 #include "runtime/runtime_state.h"
+#include "util/doris_metrics.h" // IWYU pragma: keep
+#include "util/uid_util.h"
 
 namespace doris {
+class MemTracker;
 
 DEFINE_STATIC_THREAD_LOCAL(ThreadContext, ThreadContextPtr, _ptr);
 
diff --git a/be/src/runtime/thread_context.h b/be/src/runtime/thread_context.h
index 04e8915378..983ad2fd0c 100644
--- a/be/src/runtime/thread_context.h
+++ b/be/src/runtime/thread_context.h
@@ -18,16 +18,22 @@
 #pragma once
 
 #include <bthread/bthread.h>
+#include <bthread/types.h>
+#include <gen_cpp/Types_types.h>
+#include <stdint.h>
 
+#include <memory>
+#include <ostream>
 #include <string>
 #include <thread>
 
 #include "common/logging.h"
-// #include "gen_cpp/PaloInternalService_types.h" // for TQueryType
 #include "gutil/macros.h"
+#include "runtime/exec_env.h"
+#include "runtime/memory/mem_tracker_limiter.h"
 #include "runtime/memory/thread_mem_tracker_mgr.h"
 #include "runtime/threadlocal.h"
-#include "util/defer_op.h"
+#include "util/defer_op.h" // IWYU pragma: keep
 
 // Used to observe the memory usage of the specified code segment
 #ifdef USE_MEM_TRACKER
@@ -70,8 +76,9 @@
 
 namespace doris {
 
-class TUniqueId;
 class ThreadContext;
+class MemTracker;
+class RuntimeState;
 
 extern bthread_key_t btls_key;
 
diff --git a/be/src/runtime/threadlocal.cc b/be/src/runtime/threadlocal.cc
index 689931b8b6..69b5760773 100644
--- a/be/src/runtime/threadlocal.cc
+++ b/be/src/runtime/threadlocal.cc
@@ -19,7 +19,9 @@
 #include <pthread.h>
 
 #include <memory>
+#include <mutex>
 #include <ostream>
+#include <string>
 
 #include "common/logging.h"
 #include "util/errno.h"
diff --git a/be/src/runtime/threadlocal.h b/be/src/runtime/threadlocal.h
index 4abbf14371..3842b2fcad 100644
--- a/be/src/runtime/threadlocal.h
+++ b/be/src/runtime/threadlocal.h
@@ -41,6 +41,8 @@
 
 #pragma once
 
+#include <stddef.h>
+
 #include "gutil/port.h"
 
 #define BLOCK_STATIC_THREAD_LOCAL(T, t, ...) \
diff --git a/be/src/runtime/types.cpp b/be/src/runtime/types.cpp
index 9ec81e9441..e0cf887b04 100644
--- a/be/src/runtime/types.cpp
+++ b/be/src/runtime/types.cpp
@@ -20,9 +20,16 @@
 
 #include "runtime/types.h"
 
-#include <ostream>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/types.pb.h>
+#include <stddef.h>
 
+#include <ostream>
+#include <utility>
+
+#include "olap/olap_define.h"
 #include "runtime/primitive_type.h"
+
 namespace doris {
 
 TypeDescriptor::TypeDescriptor(const std::vector<TTypeNode>& types, int* idx)
diff --git a/be/src/runtime/types.h b/be/src/runtime/types.h
index 9cbd8d70d6..82df1d01d6 100644
--- a/be/src/runtime/types.h
+++ b/be/src/runtime/types.h
@@ -20,21 +20,21 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/types.pb.h>
+#include <glog/logging.h>
+
+#include <iosfwd>
 #include <string>
 #include <vector>
 
 #include "common/config.h"
-#include "gen_cpp/Types_types.h"
-#include "gen_cpp/types.pb.h"
-#include "olap/olap_define.h"
 #include "runtime/define_primitive_type.h"
 
 namespace doris {
 
 extern const int HLL_COLUMN_DEFAULT_LEN;
 
-struct TPrimitiveType;
-
 // Describes a type. Includes the enum, children types, and any type-specific metadata
 // (e.g. precision and scale for decimals).
 // TODO for 2.3: rename to TypeDescriptor
diff --git a/be/src/runtime/user_function_cache.cpp b/be/src/runtime/user_function_cache.cpp
index 312e9ca1f2..ac87920cd8 100644
--- a/be/src/runtime/user_function_cache.cpp
+++ b/be/src/runtime/user_function_cache.cpp
@@ -17,17 +17,27 @@
 
 #include "runtime/user_function_cache.h"
 
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
+#include <glog/logging.h>
+#include <stdio.h>
+#include <string.h>
+#include <unistd.h>
+
 #include <atomic>
+#include <memory>
+#include <ostream>
 #include <regex>
+#include <utility>
 #include <vector>
 
 #include "common/config.h"
 #include "common/status.h"
 #include "gutil/strings/split.h"
 #include "http/http_client.h"
+#include "io/fs/file_system.h"
 #include "io/fs/local_file_system.h"
 #include "util/dynamic_util.h"
-#include "util/jni-util.h"
 #include "util/md5.h"
 #include "util/spinlock.h"
 #include "util/string_util.h"
diff --git a/be/src/runtime/user_function_cache.h b/be/src/runtime/user_function_cache.h
index b3c4aa7e80..1507642167 100644
--- a/be/src/runtime/user_function_cache.h
+++ b/be/src/runtime/user_function_cache.h
@@ -17,6 +17,8 @@
 
 #pragma once
 
+#include <stdint.h>
+
 #include <mutex>
 #include <string>
 #include <unordered_map>
diff --git a/be/src/service/backend_options.h b/be/src/service/backend_options.h
index 87ca36b17a..9c6f0df484 100644
--- a/be/src/service/backend_options.h
+++ b/be/src/service/backend_options.h
@@ -18,7 +18,6 @@
 #pragma once
 
 #include <butil/macros.h>
-#include <gutil/macros.h>
 
 #include <string>
 #include <vector>
diff --git a/be/src/service/backend_service.h b/be/src/service/backend_service.h
index 35a9ba8abc..727d2348a7 100644
--- a/be/src/service/backend_service.h
+++ b/be/src/service/backend_service.h
@@ -18,13 +18,8 @@
 #pragma once
 
 #include <gen_cpp/BackendService.h>
-#include <gen_cpp/DorisExternalService_types.h>
-#include <gen_cpp/TDorisExternalService.h>
 #include <stdint.h>
-#include <thrift/protocol/TDebugProtocol.h>
-#include <time.h>
 
-#include <map>
 #include <memory>
 #include <string>
 #include <vector>
@@ -39,23 +34,10 @@ class ThriftServer;
 class TAgentResult;
 class TAgentTaskRequest;
 class TAgentPublishRequest;
-class TPlanExecRequest;
-class TPlanExecParams;
 class TExecPlanFragmentParams;
 class TExecPlanFragmentResult;
-class TInsertResult;
-class TReportExecStatusArgs;
-class TReportExecStatusParams;
-class TReportExecStatusResult;
-class TCancelPlanFragmentArgs;
 class TCancelPlanFragmentResult;
-class TTransmitDataArgs;
 class TTransmitDataResult;
-class TNetworkAddress;
-class TClientRequest;
-class TExecRequest;
-class TSessionState;
-class TQueryOptions;
 class TExportTaskRequest;
 class TExportStatusResult;
 class TStreamLoadRecordResult;
diff --git a/be/src/service/brpc_service.cpp b/be/src/service/brpc_service.cpp
index 35989f29dc..80138b99ee 100644
--- a/be/src/service/brpc_service.cpp
+++ b/be/src/service/brpc_service.cpp
@@ -19,8 +19,9 @@
 
 #include <brpc/server.h>
 #include <brpc/ssl_options.h>
-#include <bthread/errno.h>
 #include <butil/endpoint.h>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
 #include <gflags/gflags_declare.h>
 #include <string.h>
 
diff --git a/be/src/service/doris_main.cpp b/be/src/service/doris_main.cpp
index be10089eae..c65f25da2e 100644
--- a/be/src/service/doris_main.cpp
+++ b/be/src/service/doris_main.cpp
@@ -15,9 +15,9 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <bthread/errno.h>
 #include <butil/macros.h>
-#include <errno.h>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
 #include <fcntl.h>
 #include <gperftools/malloc_extension.h> // IWYU pragma: keep
 #include <libgen.h>
diff --git a/be/src/service/internal_service.h b/be/src/service/internal_service.h
index 1b1e56d0cf..6cfbae4fa7 100644
--- a/be/src/service/internal_service.h
+++ b/be/src/service/internal_service.h
@@ -23,7 +23,6 @@
 #include <string>
 
 #include "common/status.h"
-#include "runtime/cache/result_cache.h"
 #include "util/priority_thread_pool.hpp"
 
 namespace google {
@@ -33,10 +32,6 @@ class RpcController;
 } // namespace protobuf
 } // namespace google
 
-namespace brpc {
-class Controller;
-}
-
 namespace doris {
 
 class ExecEnv;
diff --git a/be/src/service/point_query_executor.h b/be/src/service/point_query_executor.h
index 69fa6850e9..bc8e770049 100644
--- a/be/src/service/point_query_executor.h
+++ b/be/src/service/point_query_executor.h
@@ -21,7 +21,6 @@
 #include <butil/macros.h>
 #include <butil/time.h>
 #include <gen_cpp/Metrics_types.h>
-#include <gen_cpp/internal_service.pb.h>
 #include <parallel_hashmap/phmap.h>
 #include <stdint.h>
 #include <string.h>
@@ -53,8 +52,6 @@
 
 namespace doris {
 
-class RowCache;
-class Cache;
 class PTabletKeyLookupRequest;
 class PTabletKeyLookupResponse;
 class RuntimeState;
diff --git a/be/src/tools/meta_tool.cpp b/be/src/tools/meta_tool.cpp
index 452402852f..e4ac6493da 100644
--- a/be/src/tools/meta_tool.cpp
+++ b/be/src/tools/meta_tool.cpp
@@ -15,6 +15,8 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <gen_cpp/olap_file.pb.h>
+#include <gen_cpp/segment_v2.pb.h>
 #include <gflags/gflags.h>
 
 #include <filesystem>
@@ -25,8 +27,6 @@
 #include <string>
 
 #include "common/status.h"
-#include "gen_cpp/olap_file.pb.h"
-#include "gen_cpp/segment_v2.pb.h"
 #include "gutil/strings/numbers.h"
 #include "gutil/strings/split.h"
 #include "gutil/strings/substitute.h"
diff --git a/be/src/udf/udf.h b/be/src/udf/udf.h
index bec07e8363..5b429b85d1 100644
--- a/be/src/udf/udf.h
+++ b/be/src/udf/udf.h
@@ -20,11 +20,7 @@
 
 #pragma once
 
-#include <string.h>
-
 #include <cstdint>
-#include <functional>
-#include <iostream>
 #include <memory>
 #include <string>
 #include <vector>
@@ -35,17 +31,7 @@ namespace doris {
 
 struct ColumnPtrWrapper;
 struct StringRef;
-class BitmapValue;
-class DecimalV2Value;
-class CollectionValue;
-struct TypeDescriptor;
 class RuntimeState;
-// All input and output values will be one of the structs below. The struct is a simple
-// object containing a boolean to store if the value is nullptr and the value itself. The
-// value is unspecified if the nullptr boolean is set.
-struct AnyVal;
-struct StringRef;
-struct DateTimeVal;
 
 // The FunctionContext is passed to every UDF/UDA and is the interface for the UDF to the
 // rest of the system. It contains APIs to examine the system state, report errors
diff --git a/be/src/util/aligned_new.h b/be/src/util/aligned_new.h
index f56a56eb51..2800b19e4f 100644
--- a/be/src/util/aligned_new.h
+++ b/be/src/util/aligned_new.h
@@ -22,7 +22,8 @@
 
 #include <memory>
 
-#include "common/compiler_util.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/logging.h"
 
 namespace doris {
diff --git a/be/src/util/arrow/block_convertor.cpp b/be/src/util/arrow/block_convertor.cpp
index 3fbfcbd0ed..2b426ca2f2 100644
--- a/be/src/util/arrow/block_convertor.cpp
+++ b/be/src/util/arrow/block_convertor.cpp
@@ -17,37 +17,49 @@
 
 #include "util/arrow/block_convertor.h"
 
-#include <arrow/array.h>
+#include <arrow/array/builder_base.h>
+#include <arrow/array/builder_binary.h>
+#include <arrow/array/builder_decimal.h>
 #include <arrow/array/builder_nested.h>
 #include <arrow/array/builder_primitive.h>
-#include <arrow/buffer.h>
-#include <arrow/builder.h>
-#include <arrow/io/memory.h>
-#include <arrow/ipc/writer.h>
-#include <arrow/memory_pool.h>
 #include <arrow/record_batch.h>
 #include <arrow/status.h>
 #include <arrow/type.h>
-#include <arrow/visit_array_inline.h>
+#include <arrow/util/decimal.h>
 #include <arrow/visit_type_inline.h>
 #include <arrow/visitor.h>
+#include <glog/logging.h>
+#include <stdint.h>
 
-#include <cstdlib>
+#include <algorithm>
 #include <ctime>
 #include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
+#include <vector>
 
-#include "gutil/strings/substitute.h"
-#include "runtime/descriptor_helper.h"
-#include "runtime/descriptors.h"
-#include "runtime/jsonb_value.h"
+#include "gutil/integral_types.h"
 #include "runtime/large_int_value.h"
 #include "util/arrow/utils.h"
+#include "util/jsonb_utils.h"
 #include "util/types.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
-#include "vec/columns/column_map.h"
+#include "vec/columns/column_decimal.h"
+#include "vec/columns/column_nullable.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_array.h"
-#include "vec/data_types/data_type_map.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/runtime/vdatetime_value.h"
+
+namespace arrow {
+class Array;
+} // namespace arrow
 
 namespace doris {
 
diff --git a/be/src/util/arrow/block_convertor.h b/be/src/util/arrow/block_convertor.h
index c76ae73c9f..9fb710fc75 100644
--- a/be/src/util/arrow/block_convertor.h
+++ b/be/src/util/arrow/block_convertor.h
@@ -36,9 +36,6 @@ class Schema;
 
 namespace doris {
 
-class ObjectPool;
-class RowDescriptor;
-
 Status convert_to_arrow_batch(const vectorized::Block& block,
                               const std::shared_ptr<arrow::Schema>& schema, arrow::MemoryPool* pool,
                               std::shared_ptr<arrow::RecordBatch>* result);
diff --git a/be/src/util/arrow/row_batch.cpp b/be/src/util/arrow/row_batch.cpp
index 51c490c1ae..110977c6e6 100644
--- a/be/src/util/arrow/row_batch.cpp
+++ b/be/src/util/arrow/row_batch.cpp
@@ -17,30 +17,26 @@
 
 #include "util/arrow/row_batch.h"
 
-#include <arrow/array.h>
-#include <arrow/array/builder_primitive.h>
 #include <arrow/buffer.h>
-#include <arrow/builder.h>
 #include <arrow/io/memory.h>
 #include <arrow/ipc/writer.h>
-#include <arrow/memory_pool.h>
 #include <arrow/record_batch.h>
+#include <arrow/result.h>
 #include <arrow/status.h>
 #include <arrow/type.h>
-#include <arrow/visit_array_inline.h>
-#include <arrow/visit_type_inline.h>
-#include <arrow/visitor.h>
+#include <glog/logging.h>
+#include <stdint.h>
 
+#include <algorithm>
 #include <cstdlib>
-#include <ctime>
 #include <memory>
+#include <utility>
+#include <vector>
 
 #include "gutil/strings/substitute.h"
-#include "runtime/descriptor_helper.h"
+#include "runtime/define_primitive_type.h"
 #include "runtime/descriptors.h"
-#include "runtime/large_int_value.h"
-#include "util/arrow/utils.h"
-#include "util/types.h"
+#include "runtime/types.h"
 
 namespace doris {
 
diff --git a/be/src/util/arrow/row_batch.h b/be/src/util/arrow/row_batch.h
index 25203ec170..8946502f83 100644
--- a/be/src/util/arrow/row_batch.h
+++ b/be/src/util/arrow/row_batch.h
@@ -18,6 +18,7 @@
 #pragma once
 
 #include <memory>
+#include <string>
 
 #include "common/status.h"
 
@@ -27,7 +28,6 @@
 
 namespace arrow {
 
-class MemoryPool;
 class RecordBatch;
 class Schema;
 
@@ -35,7 +35,6 @@ class Schema;
 
 namespace doris {
 
-class ObjectPool;
 class RowDescriptor;
 
 // Convert Doris RowDescriptor to Arrow Schema.
diff --git a/be/src/util/arrow/utils.cpp b/be/src/util/arrow/utils.cpp
index ed2bdfab2a..2ef5301673 100644
--- a/be/src/util/arrow/utils.cpp
+++ b/be/src/util/arrow/utils.cpp
@@ -18,9 +18,7 @@
 #include "util/arrow/utils.h"
 
 #include <arrow/pretty_print.h>
-#include <arrow/record_batch.h>
-#include <arrow/type.h>
-#include <arrow/type_fwd.h>
+#include <arrow/status.h>
 
 #include "gutil/strings/substitute.h"
 
diff --git a/be/src/util/arrow/utils.h b/be/src/util/arrow/utils.h
index 6e8be0c625..f6ecff1a3a 100644
--- a/be/src/util/arrow/utils.h
+++ b/be/src/util/arrow/utils.h
@@ -18,7 +18,6 @@
 #pragma once
 
 #include <iostream>
-#include <memory>
 
 #include "common/status.h"
 
diff --git a/be/src/util/bfd_parser.cpp b/be/src/util/bfd_parser.cpp
index 98f0a46d62..8213d468c2 100644
--- a/be/src/util/bfd_parser.cpp
+++ b/be/src/util/bfd_parser.cpp
@@ -17,7 +17,12 @@
 
 #include "util/bfd_parser.h"
 
+#include <stdio.h>
+#include <stdlib.h>
+
+#include <algorithm>
 #include <memory>
+#include <ostream>
 
 #include "common/logging.h"
 
diff --git a/be/src/util/bit_util.h b/be/src/util/bit_util.h
index ede72dd1d5..0a60eb5429 100644
--- a/be/src/util/bit_util.h
+++ b/be/src/util/bit_util.h
@@ -24,7 +24,8 @@
 #include <endian.h>
 #endif
 
-#include "common/compiler_util.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "gutil/bits.h"
 #include "util/cpu_info.h"
 #include "util/sse_util.hpp"
diff --git a/be/src/util/bitmap.h b/be/src/util/bitmap.h
index 931b3becc3..b523427bee 100644
--- a/be/src/util/bitmap.h
+++ b/be/src/util/bitmap.h
@@ -20,6 +20,15 @@
 
 #pragma once
 
+#include <glog/logging.h>
+#include <stdint.h>
+#include <string.h>
+
+#include <algorithm>
+#include <string>
+#include <vector>
+
+#include "gutil/port.h"
 #include "gutil/strings/fastmem.h"
 #include "util/bit_util.h"
 
diff --git a/be/src/util/block_compression.cpp b/be/src/util/block_compression.cpp
index 4ff091b677..f909d4c452 100644
--- a/be/src/util/block_compression.cpp
+++ b/be/src/util/block_compression.cpp
@@ -17,15 +17,25 @@
 
 #include "util/block_compression.h"
 
+#include <gen_cpp/parquet_types.h>
+#include <gen_cpp/segment_v2.pb.h>
+#include <glog/logging.h>
+#include <limits.h>
 #include <lz4/lz4.h>
 #include <lz4/lz4frame.h>
 #include <snappy/snappy-sinksource.h>
 #include <snappy/snappy.h>
+#include <stdint.h>
+#include <zconf.h>
 #include <zlib.h>
 #include <zstd.h>
 #include <zstd_errors.h>
 
+#include <algorithm>
 #include <limits>
+#include <mutex>
+#include <new>
+#include <ostream>
 
 #include "gutil/strings/substitute.h"
 #include "util/defer_op.h"
diff --git a/be/src/util/block_compression.h b/be/src/util/block_compression.h
index 60c0c22f8e..c4991a9191 100644
--- a/be/src/util/block_compression.h
+++ b/be/src/util/block_compression.h
@@ -17,15 +17,20 @@
 
 #pragma once
 
+#include <gen_cpp/parquet_types.h>
+
 #include <cstddef>
 #include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/parquet_types.h"
-#include "gen_cpp/segment_v2.pb.h"
 #include "util/slice.h"
 
 namespace doris {
+class faststring;
+
+namespace segment_v2 {
+enum CompressionTypePB : int;
+} // namespace segment_v2
 
 // This class is used to encapsulate Compression/Decompression algorithm.
 // This class only used to compress a block data, which means all data
diff --git a/be/src/util/brpc_client_cache.cpp b/be/src/util/brpc_client_cache.cpp
index df89585e73..b9135e8014 100644
--- a/be/src/util/brpc_client_cache.cpp
+++ b/be/src/util/brpc_client_cache.cpp
@@ -17,8 +17,13 @@
 
 #include "util/brpc_client_cache.h"
 
-namespace doris {
+#include <gen_cpp/function_service.pb.h> // IWYU pragma: keep
+#include <gen_cpp/internal_service.pb.h> // IWYU pragma: keep
 
+#include "util/doris_metrics.h"
+#include "util/metrics.h"
+
+namespace doris {
 DEFINE_GAUGE_METRIC_PROTOTYPE_2ARG(brpc_endpoint_stub_count, MetricUnit::NOUNIT);
 
 DEFINE_GAUGE_METRIC_PROTOTYPE_2ARG(brpc_function_endpoint_stub_count, MetricUnit::NOUNIT);
diff --git a/be/src/util/brpc_client_cache.h b/be/src/util/brpc_client_cache.h
index 364f11e6ea..9c48c15529 100644
--- a/be/src/util/brpc_client_cache.h
+++ b/be/src/util/brpc_client_cache.h
@@ -17,20 +17,37 @@
 
 #pragma once
 
+#include <brpc/adaptive_connection_type.h>
+#include <brpc/adaptive_protocol_type.h>
+#include <brpc/channel.h>
+#include <brpc/controller.h>
+#include <butil/endpoint.h>
+#include <fmt/format.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/types.pb.h>
+#include <glog/logging.h>
+#include <google/protobuf/service.h>
 #include <parallel_hashmap/phmap.h>
+#include <stddef.h>
 
+#include <functional>
 #include <memory>
 #include <mutex>
-#include <type_traits>
+#include <ostream>
+#include <string>
+#include <utility>
+#include <vector>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/config.h"
-#include "gen_cpp/Types_types.h" // TNetworkAddress
-#include "gen_cpp/function_service.pb.h"
-#include "gen_cpp/internal_service.pb.h"
-#include "service/brpc.h"
-#include "util/doris_metrics.h"
 #include "util/network_util.h"
 
+namespace doris {
+class PBackendService_Stub;
+class PFunctionService_Stub;
+} // namespace doris
+
 template <typename T>
 using StubMap = phmap::parallel_flat_hash_map<
         std::string, std::shared_ptr<T>, std::hash<std::string>, std::equal_to<std::string>,
diff --git a/be/src/util/cgroup_util.cpp b/be/src/util/cgroup_util.cpp
index 15e5e8791c..b1cb9cf8ea 100644
--- a/be/src/util/cgroup_util.cpp
+++ b/be/src/util/cgroup_util.cpp
@@ -21,9 +21,11 @@
 #include <cfloat>
 #include <fstream>
 #include <iomanip>
-#include <iostream>
+#include <memory>
 #include <utility>
+#include <vector>
 
+#include "gutil/stringprintf.h"
 #include "gutil/strings/escaping.h"
 #include "gutil/strings/split.h"
 #include "gutil/strings/substitute.h"
diff --git a/be/src/util/cidr.cpp b/be/src/util/cidr.cpp
index 267998f51b..573074bae1 100644
--- a/be/src/util/cidr.cpp
+++ b/be/src/util/cidr.cpp
@@ -18,9 +18,15 @@
 #include "util/cidr.h"
 
 #include <arpa/inet.h>
+#include <sys/socket.h>
+
+#include <algorithm>
+#include <cstddef>
+#include <exception>
+#include <iterator>
+#include <ostream>
 
 #include "common/logging.h"
-#include "gutil/strings/split.h"
 
 namespace doris {
 
diff --git a/be/src/util/cidr.h b/be/src/util/cidr.h
index 599725554c..6204369bb1 100644
--- a/be/src/util/cidr.h
+++ b/be/src/util/cidr.h
@@ -17,9 +17,10 @@
 
 #pragma once
 
-#include <sys/socket.h>
+#include <sys/un.h>
 
 #include <array>
+#include <cstdint>
 #include <string>
 
 namespace doris {
diff --git a/be/src/util/cityhash102/city.cc b/be/src/util/cityhash102/city.cc
index 765d2a88cd..8c3d449d14 100644
--- a/be/src/util/cityhash102/city.cc
+++ b/be/src/util/cityhash102/city.cc
@@ -29,9 +29,9 @@
 
 #include "city.h"
 
+// IWYU pragma: no_include <crc32intrin.h>
 #include <string.h> // for memcpy and memset
-
-#include <algorithm>
+#include <sys/types.h>
 
 #include "config.h"
 
@@ -44,12 +44,14 @@
 
 #ifdef _MSC_VER
 #include <stdlib.h>
+
 #define bswap_32(x) _byteswap_ulong(x)
 #define bswap_64(x) _byteswap_uint64(x)
 
 #elif defined(__APPLE__)
 // Mac OS X / Darwin features
 #include <libkern/OSByteOrder.h>
+
 #define bswap_32(x) OSSwapInt32(x)
 #define bswap_64(x) OSSwapInt64(x)
 
@@ -346,7 +348,7 @@ uint128 CityHash128(const char* s, size_t len) {
 
 #if defined(__SSE4_2__) || defined(__aarch64__)
 #include "citycrc.h"
-#include "util/sse_util.hpp"
+#include "util/sse_util.hpp" // IWYU pragma: keep
 
 namespace CityHash_v1_0_2 {
 
diff --git a/be/src/util/coding.h b/be/src/util/coding.h
index 16b89507b2..ee9ade4729 100644
--- a/be/src/util/coding.h
+++ b/be/src/util/coding.h
@@ -8,9 +8,12 @@
 
 #pragma once
 
-#include <string>
+#ifndef __APPLE__
+#include <endian.h>
+#endif
+#include <stdint.h>
+#include <string.h>
 
-#include "gutil/endian.h"
 #include "olap/olap_common.h"
 #include "util/slice.h"
 
diff --git a/be/src/util/container_util.hpp b/be/src/util/container_util.hpp
index 9f11c46983..3c99dc646d 100644
--- a/be/src/util/container_util.hpp
+++ b/be/src/util/container_util.hpp
@@ -20,10 +20,11 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+
 #include <map>
 #include <unordered_map>
 
-#include "gen_cpp/Types_types.h"
 #include "util/hash_util.hpp"
 
 namespace doris {
diff --git a/be/src/util/core_local.cpp b/be/src/util/core_local.cpp
index 2077d87908..77efc18bfe 100644
--- a/be/src/util/core_local.cpp
+++ b/be/src/util/core_local.cpp
@@ -18,10 +18,15 @@
 #include "util/core_local.h"
 
 #include <cstdlib>
+#include <iterator>
+#include <ostream>
 #include <vector>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/logging.h"
 #include "util/spinlock.h"
+#include "util/sse_util.hpp"
 
 namespace doris {
 
diff --git a/be/src/util/core_local.h b/be/src/util/core_local.h
index 0ebfea0779..e23a6820d9 100644
--- a/be/src/util/core_local.h
+++ b/be/src/util/core_local.h
@@ -17,15 +17,20 @@
 
 #pragma once
 
+#include <butil/macros.h>
+#include <sched.h>
+#include <stddef.h>
+
 #include <deque>
 #include <map>
-#include <memory>
 #include <mutex>
+#include <new>
 #include <thread>
+#include <utility>
 #include <vector>
 
-#include "common/compiler_util.h"
-#include "gutil/macros.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 
 namespace doris {
 
diff --git a/be/src/util/countdown_latch.h b/be/src/util/countdown_latch.h
index 4647f920fc..43db2bf189 100644
--- a/be/src/util/countdown_latch.h
+++ b/be/src/util/countdown_latch.h
@@ -17,6 +17,7 @@
 
 #pragma once
 
+// IWYU pragma: no_include <bits/chrono.h>
 #include <chrono>
 #include <condition_variable>
 #include <mutex>
diff --git a/be/src/util/cpu_info.cpp b/be/src/util/cpu_info.cpp
index 92badaed6b..60bdf1cd17 100644
--- a/be/src/util/cpu_info.cpp
+++ b/be/src/util/cpu_info.cpp
@@ -21,8 +21,6 @@
 #include "util/cpu_info.h"
 
 #if defined(__GNUC__) && (defined(__x86_64__) || defined(__i386__))
-/* GCC-compatible compiler, targeting x86/x86-64 */
-#include <x86intrin.h>
 #elif defined(__GNUC__) && defined(__ARM_NEON__)
 /* GCC-compatible compiler, targeting ARM with NEON */
 #include <arm_neon.h>
@@ -43,24 +41,25 @@
 #include <sys/sysctl.h>
 #endif
 
+#include <gen_cpp/Metrics_types.h>
 #include <sched.h>
 #include <stdlib.h>
-#include <string.h>
 #include <unistd.h>
 
 #include <algorithm>
-#include <boost/algorithm/string.hpp>
+#include <boost/algorithm/string/predicate.hpp>
+#include <boost/algorithm/string/trim.hpp>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
 #include <filesystem>
 #include <fstream>
-#include <iostream>
-#include <sstream>
 
 #include "common/config.h"
 #include "common/env_config.h"
 #include "gflags/gflags.h"
+#include "gutil/stringprintf.h"
 #include "gutil/strings/substitute.h"
 #include "util/pretty_printer.h"
-#include "util/string_parser.hpp"
 
 using boost::algorithm::contains;
 using boost::algorithm::trim;
diff --git a/be/src/util/cpu_info.h b/be/src/util/cpu_info.h
index d09c89775f..0e03ad3928 100644
--- a/be/src/util/cpu_info.h
+++ b/be/src/util/cpu_info.h
@@ -21,6 +21,7 @@
 #pragma once
 
 #include <glog/logging.h>
+#include <stdint.h>
 
 #include <memory>
 #include <string>
diff --git a/be/src/util/crc32c.cpp b/be/src/util/crc32c.cpp
index 99884c3b5e..7ad31e7f22 100644
--- a/be/src/util/crc32c.cpp
+++ b/be/src/util/crc32c.cpp
@@ -18,8 +18,11 @@
 // the following code are modified from RocksDB:
 // https://github.com/facebook/rocksdb/blob/master/util/crc32c.cc
 
+// IWYU pragma: no_include <crc32intrin.h>
+#include <stddef.h>
+#include <stdint.h>
+
 #include "util/coding.h"
-#include "util/sse_util.hpp"
 
 namespace doris {
 namespace crc32c {
diff --git a/be/src/util/date_func.cpp b/be/src/util/date_func.cpp
index c77be510f8..e7ae30e003 100644
--- a/be/src/util/date_func.cpp
+++ b/be/src/util/date_func.cpp
@@ -17,7 +17,12 @@
 
 #include "util/date_func.h"
 
-#include <iomanip>
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <string.h>
+#include <time.h>
+
+#include <ostream>
 
 #include "vec/runtime/vdatetime_value.h"
 
diff --git a/be/src/util/date_func.h b/be/src/util/date_func.h
index 4378fd32e1..cc9aa7b6da 100644
--- a/be/src/util/date_func.h
+++ b/be/src/util/date_func.h
@@ -18,12 +18,9 @@
 #pragma once
 
 #include <stdint.h>
-#include <time.h>
 
 #include <string>
 
-#include "olap/field.h"
-
 namespace doris {
 
 uint64_t timestamp_from_datetime(const std::string& datetime_str);
diff --git a/be/src/util/debug_util.cpp b/be/src/util/debug_util.cpp
index 9179c70a60..bfb4826514 100644
--- a/be/src/util/debug_util.cpp
+++ b/be/src/util/debug_util.cpp
@@ -17,14 +17,14 @@
 
 #include "util/debug_util.h"
 
-#include <iomanip>
-#include <sstream>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/version.h>
+#include <stdint.h>
 
-#include "common/logging.h"
-#include "gen_cpp/Opcodes_types.h"
-#include "gen_cpp/types.pb.h"
-#include "gen_cpp/version.h"
-#include "util/cpu_info.h"
+#include <iomanip>
+#include <map>
+#include <sstream> // IWYU pragma: keep
+#include <utility>
 
 namespace doris {
 
diff --git a/be/src/util/debug_util.h b/be/src/util/debug_util.h
index 956a38a403..fbc0c221f4 100644
--- a/be/src/util/debug_util.h
+++ b/be/src/util/debug_util.h
@@ -17,21 +17,14 @@
 
 #pragma once
 
-#include <ostream>
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Types_types.h>
+
 #include <string>
 
-#include "gen_cpp/AgentService_types.h"
-#include "gen_cpp/Descriptors_types.h"
-#include "gen_cpp/Exprs_types.h"
-#include "gen_cpp/Opcodes_types.h"
-#include "gen_cpp/PlanNodes_types.h"
-#include "gen_cpp/RuntimeProfile_types.h"
-#include "gen_cpp/Types_types.h"
-
 namespace doris {
 
-class PUniqueId;
-
 std::string print_plan_node_type(const TPlanNodeType::type& type);
 std::string print_tstmt_type(const TStmtType::type& type);
 std::string print_query_state(const QueryState::type& type);
diff --git a/be/src/util/disk_info.cpp b/be/src/util/disk_info.cpp
index b3c70706bb..86ddb447f2 100644
--- a/be/src/util/disk_info.cpp
+++ b/be/src/util/disk_info.cpp
@@ -17,17 +17,23 @@
 
 #include "util/disk_info.h"
 
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
 #include <sys/stat.h>
 #include <sys/sysmacros.h>
 #include <sys/types.h>
-#include <sys/vfs.h>
-#include <unistd.h>
 
-#include <boost/algorithm/string.hpp>
-#include <boost/algorithm/string/join.hpp>
+#include <algorithm>
+#include <boost/algorithm/string/classification.hpp>
+#include <boost/algorithm/string/detail/classification.hpp>
+#include <boost/algorithm/string/trim.hpp>
 #include <fstream>
-#include <iostream>
-#include <sstream>
+#include <iterator>
+#include <memory>
+#include <utility>
 
 #include "gutil/strings/split.h"
 #include "io/fs/local_file_system.h"
diff --git a/be/src/util/disk_info.h b/be/src/util/disk_info.h
index f284f55433..f415cb958d 100644
--- a/be/src/util/disk_info.h
+++ b/be/src/util/disk_info.h
@@ -17,9 +17,12 @@
 
 #pragma once
 
+#include <sys/types.h>
+
 #include <map>
 #include <set>
 #include <string>
+#include <vector>
 
 #include "common/logging.h"
 #include "common/status.h"
diff --git a/be/src/util/doris_metrics.cpp b/be/src/util/doris_metrics.cpp
index a40d1691cb..97ca3de938 100644
--- a/be/src/util/doris_metrics.cpp
+++ b/be/src/util/doris_metrics.cpp
@@ -17,14 +17,26 @@
 
 #include "util/doris_metrics.h"
 
-#include <sys/types.h>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
+#include <glog/logging.h>
+#include <inttypes.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
 #include <unistd.h>
 
+#include <functional>
+#include <ostream>
+
+#include "common/status.h"
 #include "io/fs/local_file_system.h"
-#include "util/debug_util.h"
 #include "util/system_metrics.h"
 
 namespace doris {
+namespace io {
+struct FileInfo;
+} // namespace io
 
 DEFINE_COUNTER_METRIC_PROTOTYPE_3ARG(fragment_requests_total, MetricUnit::REQUESTS,
                                      "Total fragment requests received.");
diff --git a/be/src/util/doris_metrics.h b/be/src/util/doris_metrics.h
index ec44f58e53..86b52e32be 100644
--- a/be/src/util/doris_metrics.h
+++ b/be/src/util/doris_metrics.h
@@ -17,9 +17,9 @@
 
 #pragma once
 
+#include <memory>
 #include <set>
 #include <string>
-#include <unordered_map>
 #include <vector>
 
 #include "util/metrics.h"
diff --git a/be/src/util/dynamic_util.cpp b/be/src/util/dynamic_util.cpp
index d6ff812493..1f09c11b6d 100644
--- a/be/src/util/dynamic_util.cpp
+++ b/be/src/util/dynamic_util.cpp
@@ -22,8 +22,6 @@
 
 #include <dlfcn.h>
 
-#include <sstream>
-
 namespace doris {
 
 Status dynamic_lookup(void* handle, const char* symbol, void** fn_ptr) {
diff --git a/be/src/util/easy_json.cc b/be/src/util/easy_json.cc
index 96564b1d16..4df9e88fce 100644
--- a/be/src/util/easy_json.cc
+++ b/be/src/util/easy_json.cc
@@ -18,6 +18,7 @@
 #include "util/easy_json.h"
 
 #include <glog/logging.h>
+#include <rapidjson/allocators.h>
 #include <rapidjson/document.h>
 #include <rapidjson/rapidjson.h>
 #include <rapidjson/stringbuffer.h>
diff --git a/be/src/util/encryption_util.cpp b/be/src/util/encryption_util.cpp
index b490a30fc3..ab1ff3e857 100644
--- a/be/src/util/encryption_util.cpp
+++ b/be/src/util/encryption_util.cpp
@@ -22,6 +22,7 @@
 #include <openssl/ossl_typ.h>
 #include <sys/types.h>
 
+#include <algorithm>
 #include <cstring>
 #include <string>
 
diff --git a/be/src/util/error_util.cc b/be/src/util/error_util.cc
index f1b006b6fd..f13161eed8 100644
--- a/be/src/util/error_util.cc
+++ b/be/src/util/error_util.cc
@@ -20,7 +20,8 @@
 
 #include "util/error_util.h"
 
-#include <errno.h>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
 
 #include <cstring>
 #include <sstream>
diff --git a/be/src/util/faststring.h b/be/src/util/faststring.h
index 354a8965ec..139b7bd92b 100644
--- a/be/src/util/faststring.h
+++ b/be/src/util/faststring.h
@@ -17,12 +17,13 @@
 
 #pragma once
 
+#include <butil/macros.h>
+
 #include <cstdint>
 #include <cstring>
 #include <string>
 
 #include "gutil/dynamic_annotations.h"
-#include "gutil/macros.h"
 #include "gutil/port.h"
 #include "gutil/strings/fastmem.h"
 #include "util/slice.h"
diff --git a/be/src/util/frame_of_reference_coding.cpp b/be/src/util/frame_of_reference_coding.cpp
index 0e7bb57b64..d98726c6e9 100644
--- a/be/src/util/frame_of_reference_coding.cpp
+++ b/be/src/util/frame_of_reference_coding.cpp
@@ -17,8 +17,14 @@
 
 #include "util/frame_of_reference_coding.h"
 
+#include <glog/logging.h>
+#include <sys/types.h>
+
 #include <algorithm>
 #include <cstring>
+#include <iostream>
+#include <iterator>
+#include <limits>
 
 #include "util/bit_util.h"
 #include "util/coding.h"
diff --git a/be/src/util/frame_of_reference_coding.h b/be/src/util/frame_of_reference_coding.h
index da3d716e75..06fd2a25f0 100644
--- a/be/src/util/frame_of_reference_coding.h
+++ b/be/src/util/frame_of_reference_coding.h
@@ -17,14 +17,13 @@
 
 #pragma once
 
+#include <stdint.h>
+
 #include <cstdlib>
-#include <iostream>
-#include <limits>
+#include <vector>
 
 #include "olap/olap_common.h"
 #include "olap/uint24.h"
-#include "util/bit_stream_utils.h"
-#include "util/bit_stream_utils.inline.h"
 #include "util/faststring.h"
 
 namespace doris {
diff --git a/be/src/util/hash_util.hpp b/be/src/util/hash_util.hpp
index 2b9dbe0b2a..e8c70e290c 100644
--- a/be/src/util/hash_util.hpp
+++ b/be/src/util/hash_util.hpp
@@ -20,13 +20,14 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
 #include <xxh3.h>
 #include <zlib.h>
 
 #include <functional>
 
-#include "common/compiler_util.h"
-#include "gen_cpp/Types_types.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "runtime/define_primitive_type.h"
 #include "util/cpu_info.h"
 #include "util/murmur_hash3.h"
diff --git a/be/src/util/hdfs_util.cpp b/be/src/util/hdfs_util.cpp
index d8be4b96c6..643053292e 100644
--- a/be/src/util/hdfs_util.cpp
+++ b/be/src/util/hdfs_util.cpp
@@ -17,11 +17,11 @@
 
 #include "util/hdfs_util.h"
 
-#include <util/string_util.h>
+#include <ostream>
 
-#include "common/config.h"
 #include "common/logging.h"
 #include "io/fs/err_utils.h"
+#include "io/hdfs_builder.h"
 
 namespace doris {
 namespace io {
diff --git a/be/src/util/hdfs_util.h b/be/src/util/hdfs_util.h
index 2e56181df7..39dc8bc416 100644
--- a/be/src/util/hdfs_util.h
+++ b/be/src/util/hdfs_util.h
@@ -17,16 +17,14 @@
 
 #pragma once
 
-#include <map>
-#include <memory>
 #include <string>
 
-#include "common/status.h"
 #include "io/fs/hdfs.h"
 #include "io/fs/path.h"
-#include "io/hdfs_builder.h"
 
 namespace doris {
+class HDFSCommonBuilder;
+
 namespace io {
 
 class HDFSHandle {
diff --git a/be/src/util/histogram.cpp b/be/src/util/histogram.cpp
index 21664eacd5..8cdcc42b7d 100644
--- a/be/src/util/histogram.cpp
+++ b/be/src/util/histogram.cpp
@@ -19,9 +19,11 @@
 
 #include <stdio.h>
 
+#include <algorithm>
 #include <cinttypes>
 #include <cmath>
 #include <limits>
+#include <utility>
 
 namespace doris {
 
diff --git a/be/src/util/histogram.h b/be/src/util/histogram.h
index c8dd6b8ed8..3055c75994 100644
--- a/be/src/util/histogram.h
+++ b/be/src/util/histogram.h
@@ -17,6 +17,9 @@
 
 #pragma once
 
+#include <stddef.h>
+#include <stdint.h>
+
 #include <atomic>
 #include <map>
 #include <string>
diff --git a/be/src/util/jni-util.cpp b/be/src/util/jni-util.cpp
index 955efb78e7..411ef99dc2 100644
--- a/be/src/util/jni-util.cpp
+++ b/be/src/util/jni-util.cpp
@@ -17,20 +17,22 @@
 
 #include "util/jni-util.h"
 
+#include <fmt/format.h>
+#include <glog/logging.h>
 #include <jni.h>
 #include <jni_md.h>
-#include <stdlib.h>
 
+#include <algorithm>
 #include <cstdlib>
 #include <filesystem>
+#include <iterator>
+#include <memory>
 #include <mutex>
 #include <sstream>
 #include <string>
 #include <vector>
 
-#include "common/config.h"
 #include "gutil/strings/substitute.h"
-#include "util/defer_op.h"
 #include "util/jni_native_method.h"
 #include "util/libjvm_loader.h"
 
diff --git a/be/src/util/jni-util.h b/be/src/util/jni-util.h
index ec5f6abf6e..f43d5577d1 100644
--- a/be/src/util/jni-util.h
+++ b/be/src/util/jni-util.h
@@ -17,10 +17,16 @@
 
 #pragma once
 
+#include <butil/macros.h>
 #include <jni.h>
+#include <limits.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <string>
 
 #include "common/status.h"
-#include "gutil/macros.h"
+#include "jni_md.h"
 #include "util/thrift_util.h"
 
 #ifdef USE_HADOOP_HDFS
@@ -29,6 +35,7 @@ extern "C" JNIEnv* getJNIEnv(void);
 #endif
 
 namespace doris {
+class JniUtil;
 
 #define RETURN_ERROR_IF_EXC(env)                                     \
     do {                                                             \
diff --git a/be/src/util/jni_native_method.cpp b/be/src/util/jni_native_method.cpp
index 185293440f..50f43c4d9a 100644
--- a/be/src/util/jni_native_method.cpp
+++ b/be/src/util/jni_native_method.cpp
@@ -17,6 +17,9 @@
 
 #include "jni_native_method.h"
 
+#include <stdlib.h>
+
+#include "jni.h"
 #include "vec/columns/column_string.h"
 
 namespace doris {
diff --git a/be/src/util/jni_native_method.h b/be/src/util/jni_native_method.h
index 75551870cf..d208cb0d95 100644
--- a/be/src/util/jni_native_method.h
+++ b/be/src/util/jni_native_method.h
@@ -19,6 +19,8 @@
 
 #include <jni.h>
 
+#include "jni_md.h"
+
 namespace doris {
 
 /**
diff --git a/be/src/util/jsonb_document.h b/be/src/util/jsonb_document.h
index fb98175b60..19106cdea4 100644
--- a/be/src/util/jsonb_document.h
+++ b/be/src/util/jsonb_document.h
@@ -74,6 +74,9 @@
 #include <limits>
 #include <type_traits>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+
 // #include "util/string_parser.hpp"
 
 namespace doris {
diff --git a/be/src/util/key_util.h b/be/src/util/key_util.h
index 13cb5c1768..ccd67b9cdd 100644
--- a/be/src/util/key_util.h
+++ b/be/src/util/key_util.h
@@ -17,13 +17,14 @@
 
 #pragma once
 
+#include <gen_cpp/segment_v2.pb.h>
+
 #include <cstdint>
 #include <iterator>
 #include <string>
 #include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/segment_v2.pb.h"
 #include "util/debug_util.h"
 #include "util/faststring.h"
 #include "util/slice.h"
diff --git a/be/src/util/libjvm_loader.cpp b/be/src/util/libjvm_loader.cpp
index 6175da6081..02375f8142 100644
--- a/be/src/util/libjvm_loader.cpp
+++ b/be/src/util/libjvm_loader.cpp
@@ -24,6 +24,8 @@
 #include <mutex>
 
 #include "common/status.h"
+#include "jni.h"
+#include "jni_md.h"
 
 _JNI_IMPORT_OR_EXPORT_ jint JNICALL JNI_GetCreatedJavaVMs(JavaVM** vm_buf, jsize bufLen,
                                                           jsize* numVMs) {
diff --git a/be/src/util/libjvm_loader.h b/be/src/util/libjvm_loader.h
index 2982cbea7f..cb55c22221 100644
--- a/be/src/util/libjvm_loader.h
+++ b/be/src/util/libjvm_loader.h
@@ -17,8 +17,7 @@
 
 #pragma once
 
-#include <dlfcn.h>
-#include <jni.h>
+#include <jni.h> // IWYU pragma: keep
 
 #include <memory>
 #include <string>
diff --git a/be/src/util/mem_info.cpp b/be/src/util/mem_info.cpp
index da52948dbc..918c3c85fc 100644
--- a/be/src/util/mem_info.cpp
+++ b/be/src/util/mem_info.cpp
@@ -24,21 +24,29 @@
 #include <sys/sysctl.h>
 #endif
 
-#include <stdlib.h>
-#include <string.h>
-#include <unistd.h>
+#include <fmt/format.h>
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/segment_v2.pb.h>
+#include <jemalloc/jemalloc.h>
 
+#include <algorithm>
+#include <boost/algorithm/string/trim.hpp>
 #include <fstream>
-#include <iostream>
-#include <sstream>
+#include <unordered_map>
+#include <vector>
 
 #include "common/config.h"
+#include "common/status.h"
 #include "gutil/strings/split.h"
 #include "olap/page_cache.h"
 #include "olap/segment_loader.h"
+#include "runtime/memory/chunk_allocator.h"
+#include "runtime/memory/mem_tracker_limiter.h"
 #include "util/cgroup_util.h"
+#include "util/defer_op.h"
 #include "util/parse_util.h"
 #include "util/pretty_printer.h"
+#include "util/stopwatch.hpp"
 #include "util/string_parser.hpp"
 
 namespace doris {
diff --git a/be/src/util/mem_info.h b/be/src/util/mem_info.h
index b13368222a..b25e194916 100644
--- a/be/src/util/mem_info.h
+++ b/be/src/util/mem_info.h
@@ -20,8 +20,10 @@
 
 #pragma once
 
-#include <gperftools/malloc_extension.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <atomic>
 #include <string>
 
 #include "common/logging.h"
@@ -31,7 +33,6 @@
 #include <gperftools/malloc_extension.h>
 #endif
 #include "util/perf_counters.h"
-#include "util/pretty_printer.h"
 
 namespace doris {
 
diff --git a/be/src/util/metrics.cpp b/be/src/util/metrics.cpp
index a9cae294d6..de2d43c854 100644
--- a/be/src/util/metrics.cpp
+++ b/be/src/util/metrics.cpp
@@ -17,8 +17,15 @@
 
 #include "util/metrics.h"
 
+#include <glog/logging.h>
+#include <rapidjson/encodings.h>
 #include <rapidjson/stringbuffer.h>
 #include <rapidjson/writer.h>
+
+#include <initializer_list>
+
+#include "common/config.h"
+
 namespace doris {
 
 std::ostream& operator<<(std::ostream& os, MetricType type) {
diff --git a/be/src/util/metrics.h b/be/src/util/metrics.h
index 648f96809f..6007562eb0 100644
--- a/be/src/util/metrics.h
+++ b/be/src/util/metrics.h
@@ -19,18 +19,20 @@
 
 #include <rapidjson/document.h>
 #include <rapidjson/rapidjson.h>
+#include <stddef.h>
+#include <stdint.h>
 
 #include <atomic>
 #include <functional>
-#include <iomanip>
+#include <map>
+#include <memory>
 #include <mutex>
-#include <ostream>
-#include <set>
 #include <sstream>
 #include <string>
 #include <unordered_map>
+#include <utility>
+#include <vector>
 
-#include "common/config.h"
 #include "util/core_local.h"
 #include "util/histogram.h"
 #include "util/spinlock.h"
@@ -39,8 +41,6 @@ namespace doris {
 
 namespace rj = RAPIDJSON_NAMESPACE;
 
-class MetricRegistry;
-
 enum class MetricType { COUNTER, GAUGE, HISTOGRAM, SUMMARY, UNTYPED };
 
 enum class MetricUnit {
diff --git a/be/src/util/mustache/mustache.cc b/be/src/util/mustache/mustache.cc
index a3947e8463..e72967152d 100644
--- a/be/src/util/mustache/mustache.cc
+++ b/be/src/util/mustache/mustache.cc
@@ -12,10 +12,23 @@
 
 #include "mustache.h"
 
+#include <rapidjson/allocators.h>
+#include <rapidjson/document.h>
+#include <rapidjson/encodings.h>
 #include <rapidjson/prettywriter.h>
+#include <rapidjson/rapidjson.h>
+#include <strings.h>
 
-#include <boost/algorithm/string.hpp>
-#include <fstream>
+#include <algorithm>
+#include <boost/algorithm/string/classification.hpp>
+#include <boost/algorithm/string/detail/classification.hpp>
+#include <boost/algorithm/string/join.hpp>
+#include <boost/algorithm/string/predicate_facade.hpp>
+#include <boost/algorithm/string/split.hpp>
+#include <boost/algorithm/string/trim.hpp>
+#include <boost/iterator/iterator_facade.hpp>
+#include <boost/type_index/type_index_facade.hpp>
+#include <fstream> // IWYU pragma: keep
 #include <iostream>
 #include <stack>
 #include <vector>
diff --git a/be/src/util/mustache/mustache.h b/be/src/util/mustache/mustache.h
index e4b861f32f..90b2e0f3ee 100644
--- a/be/src/util/mustache/mustache.h
+++ b/be/src/util/mustache/mustache.h
@@ -13,6 +13,7 @@
 #pragma once
 
 #include <sstream>
+#include <string>
 
 #include "rapidjson/document.h"
 
diff --git a/be/src/util/mysql_row_buffer.cpp b/be/src/util/mysql_row_buffer.cpp
index e86613713b..b3c0f0ccf2 100644
--- a/be/src/util/mysql_row_buffer.cpp
+++ b/be/src/util/mysql_row_buffer.cpp
@@ -17,16 +17,24 @@
 
 #include "util/mysql_row_buffer.h"
 
+#include <assert.h>
 #include <fmt/format.h>
-#include <stdio.h>
-#include <stdlib.h>
+#include <string.h>
+#include <sys/types.h>
+
+#include <algorithm>
+#include <new>
+#include <ostream>
+#include <string>
 
 #include "common/logging.h"
 #include "date_func.h"
 #include "gutil/strings/numbers.h"
+#include "olap/olap_common.h"
+#include "runtime/decimalv2_value.h"
 #include "runtime/large_int_value.h"
 #include "util/mysql_global.h"
-#include "vec/runtime/vdatetime_value.h"
+#include "vec/runtime/vdatetime_value.h" // IWYU pragma: keep
 
 namespace doris {
 
diff --git a/be/src/util/network_util.cpp b/be/src/util/network_util.cpp
index 78e399a6a9..4b53e7f9b6 100644
--- a/be/src/util/network_util.cpp
+++ b/be/src/util/network_util.cpp
@@ -21,11 +21,17 @@
 #include "util/network_util.h"
 
 #include <arpa/inet.h>
-#include <common/logging.h>
+// IWYU pragma: no_include <bits/local_lim.h>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
+#include <fmt/format.h>
+#include <gen_cpp/Types_types.h>
 #include <ifaddrs.h>
-#include <limits.h>
 #include <netdb.h>
+#include <netinet/in.h>
+#include <string.h>
 #include <sys/socket.h>
+#include <unistd.h>
 
 #include <sstream>
 
diff --git a/be/src/util/network_util.h b/be/src/util/network_util.h
index 27e97a23a0..195d909f40 100644
--- a/be/src/util/network_util.h
+++ b/be/src/util/network_util.h
@@ -20,10 +20,13 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <sys/un.h>
+
+#include <string>
 #include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/Types_types.h"
 
 namespace doris {
 
diff --git a/be/src/util/os_info.cpp b/be/src/util/os_info.cpp
index db14642806..0b42e29bbb 100644
--- a/be/src/util/os_info.cpp
+++ b/be/src/util/os_info.cpp
@@ -20,13 +20,7 @@
 
 #include "util/os_info.h"
 
-#include <stdlib.h>
-#include <string.h>
-#include <unistd.h>
-
 #include <fstream>
-#include <iostream>
-#include <sstream>
 
 namespace doris {
 
diff --git a/be/src/util/os_util.cpp b/be/src/util/os_util.cpp
index b59b12bee9..0b942c5f04 100644
--- a/be/src/util/os_util.cpp
+++ b/be/src/util/os_util.cpp
@@ -25,21 +25,16 @@
 #include <sys/resource.h>
 #include <unistd.h>
 
-#include <cstddef>
-#include <ostream>
+#include <algorithm>
 #include <string>
-#include <utility>
 #include <vector>
 
 #include "gutil/macros.h"
 #include "gutil/strings/numbers.h"
 #include "gutil/strings/split.h"
-#include "gutil/strings/stringpiece.h"
 #include "gutil/strings/substitute.h"
-#include "gutil/strings/util.h"
 #include "io/fs/fs_utils.h"
 #include "io/fs/local_file_system.h"
-#include "util/faststring.h"
 
 using std::string;
 using std::vector;
diff --git a/be/src/util/os_util.h b/be/src/util/os_util.h
index 42b3d61572..f830136b68 100644
--- a/be/src/util/os_util.h
+++ b/be/src/util/os_util.h
@@ -22,7 +22,6 @@
 
 #include <cstdint>
 #include <string>
-#include <type_traits>
 
 #include "common/status.h"
 
diff --git a/be/src/util/parse_util.h b/be/src/util/parse_util.h
index 14c8cb85ba..22ad2940b3 100644
--- a/be/src/util/parse_util.h
+++ b/be/src/util/parse_util.h
@@ -20,6 +20,8 @@
 
 #pragma once
 
+#include <stdint.h>
+
 #include <string>
 
 namespace doris {
diff --git a/be/src/util/path_util.cpp b/be/src/util/path_util.cpp
index 3de2f106c3..d3e860044f 100644
--- a/be/src/util/path_util.cpp
+++ b/be/src/util/path_util.cpp
@@ -17,12 +17,9 @@
 
 #include "util/path_util.h"
 
-#include <cstring>
-#include <memory>
 // Use the POSIX version of dirname(3). See `man 3 dirname`
 #include <libgen.h>
 
-#include "common/logging.h"
 #include "gutil/strings/split.h"
 #include "gutil/strings/stringpiece.h"
 #include "gutil/strings/strip.h"
diff --git a/be/src/util/perf_counters.cpp b/be/src/util/perf_counters.cpp
index d802cc77c8..e1b4bb7de9 100644
--- a/be/src/util/perf_counters.cpp
+++ b/be/src/util/perf_counters.cpp
@@ -21,19 +21,20 @@
 #include "util/perf_counters.h"
 
 #include <linux/perf_event.h>
-#include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
 #include <sys/syscall.h>
+#include <unistd.h>
 
 #include <boost/algorithm/string/trim.hpp>
-#include <fstream>
+#include <fstream> // IWYU pragma: keep
 #include <iomanip>
 #include <iostream>
-#include <sstream>
+#include <unordered_map>
+#include <utility>
 
+#include "gutil/stringprintf.h"
 #include "gutil/strings/substitute.h"
-#include "util/debug_util.h"
 #include "util/pretty_printer.h"
 #include "util/string_parser.hpp"
 #include "util/string_util.h"
diff --git a/be/src/util/perf_counters.h b/be/src/util/perf_counters.h
index 1a19f6025b..9f0936193d 100644
--- a/be/src/util/perf_counters.h
+++ b/be/src/util/perf_counters.h
@@ -20,12 +20,13 @@
 
 #pragma once
 
+#include <gen_cpp/Metrics_types.h>
+#include <stdint.h>
+
 #include <iostream>
 #include <string>
 #include <vector>
 
-#include "util/debug_util.h"
-
 // This is a utility class that aggregates counters from the kernel.  These counters
 // come from different sources.
 //   - perf counter syscall (/usr/include/linux/perf_event.h")
diff --git a/be/src/util/pprof_utils.cpp b/be/src/util/pprof_utils.cpp
index 2025a0a4a7..18f62fb1de 100644
--- a/be/src/util/pprof_utils.cpp
+++ b/be/src/util/pprof_utils.cpp
@@ -17,7 +17,12 @@
 
 #include "util/pprof_utils.h"
 
-#include <fstream>
+#include <stdio.h>
+#include <unistd.h>
+
+#include <cstdlib>
+#include <fstream> // IWYU pragma: keep
+#include <memory>
 
 #include "agent/utils.h"
 #include "gutil/strings/substitute.h"
diff --git a/be/src/util/pprof_utils.h b/be/src/util/pprof_utils.h
index bef4c64dcb..461ec56280 100644
--- a/be/src/util/pprof_utils.h
+++ b/be/src/util/pprof_utils.h
@@ -17,8 +17,10 @@
 
 #pragma once
 
+#include <stdint.h>
+
 #include <iostream>
-#include <sstream>
+#include <string>
 
 #include "common/status.h"
 
diff --git a/be/src/util/pretty_printer.h b/be/src/util/pretty_printer.h
index 501d15d42a..2f48eae33a 100644
--- a/be/src/util/pretty_printer.h
+++ b/be/src/util/pretty_printer.h
@@ -20,12 +20,13 @@
 
 #pragma once
 
+#include <gen_cpp/RuntimeProfile_types.h>
+
 #include <boost/algorithm/string.hpp>
 #include <cmath>
 #include <iomanip>
 #include <sstream>
 
-#include "gen_cpp/RuntimeProfile_types.h"
 #include "util/binary_cast.hpp"
 #include "util/cpu_info.h"
 
diff --git a/be/src/util/progress_updater.cpp b/be/src/util/progress_updater.cpp
index c3f3670324..320b37ed31 100644
--- a/be/src/util/progress_updater.cpp
+++ b/be/src/util/progress_updater.cpp
@@ -20,6 +20,8 @@
 
 #include "util/progress_updater.h"
 
+#include <ostream>
+
 #include "common/logging.h"
 
 namespace doris {
diff --git a/be/src/util/progress_updater.h b/be/src/util/progress_updater.h
index 5f241673af..f37de88335 100644
--- a/be/src/util/progress_updater.h
+++ b/be/src/util/progress_updater.h
@@ -20,6 +20,8 @@
 
 #pragma once
 
+#include <stdint.h>
+
 #include <string>
 
 namespace doris {
diff --git a/be/src/util/quantile_state.cpp b/be/src/util/quantile_state.cpp
index d991f77953..40aba9006f 100644
--- a/be/src/util/quantile_state.cpp
+++ b/be/src/util/quantile_state.cpp
@@ -19,9 +19,13 @@
 #include <string.h>
 
 #include <cmath>
+#include <ostream>
+#include <utility>
 
 #include "common/logging.h"
 #include "util/coding.h"
+#include "util/slice.h"
+#include "util/tdigest.h"
 
 namespace doris {
 
diff --git a/be/src/util/quantile_state.h b/be/src/util/quantile_state.h
index c3af66c495..6f3a92b2a4 100644
--- a/be/src/util/quantile_state.h
+++ b/be/src/util/quantile_state.h
@@ -17,16 +17,17 @@
 
 #pragma once
 
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
 #include <memory>
-#include <string>
 #include <vector>
 
 #include "slice.h"
-#include "tdigest.h"
 
 namespace doris {
 
-struct Slice;
 class TDigest;
 
 const static int QUANTILE_STATE_EXPLICIT_NUM = 2048;
diff --git a/be/src/util/radix_sort.h b/be/src/util/radix_sort.h
index a3738ac3d1..463af1c772 100644
--- a/be/src/util/radix_sort.h
+++ b/be/src/util/radix_sort.h
@@ -37,7 +37,8 @@
 #include <cstdlib>
 #include <type_traits>
 
-#include "common/compiler_util.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 
 namespace doris {
 
diff --git a/be/src/util/runtime_profile.cpp b/be/src/util/runtime_profile.cpp
index 32e76c94eb..186c61fa56 100644
--- a/be/src/util/runtime_profile.cpp
+++ b/be/src/util/runtime_profile.cpp
@@ -20,17 +20,19 @@
 
 #include "util/runtime_profile.h"
 
+#include <gen_cpp/RuntimeProfile_types.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+#include <opentelemetry/trace/span.h>
+#include <opentelemetry/trace/tracer.h>
+#include <rapidjson/encodings.h>
+#include <rapidjson/stringbuffer.h>
+#include <rapidjson/writer.h>
+
 #include <iomanip>
 #include <iostream>
-#include <thread>
 
-#include "common/config.h"
 #include "common/object_pool.h"
 #include "util/container_util.hpp"
-#include "util/cpu_info.h"
-#include "util/debug_util.h"
-#include "util/thrift_util.h"
-#include "util/url_coding.h"
 
 namespace doris {
 
diff --git a/be/src/util/runtime_profile.h b/be/src/util/runtime_profile.h
index c80507110e..cd37f9fd43 100644
--- a/be/src/util/runtime_profile.h
+++ b/be/src/util/runtime_profile.h
@@ -20,24 +20,32 @@
 
 #pragma once
 
-#include <sys/resource.h>
-#include <sys/time.h>
+#include <gen_cpp/Metrics_types.h>
+#include <glog/logging.h>
+#include <stdint.h>
 
+#include <algorithm>
 #include <atomic>
 #include <functional>
 #include <iostream>
+#include <map>
+#include <memory>
 #include <mutex>
-#include <thread>
+#include <set>
+#include <string>
+#include <utility>
+#include <vector>
 
-#include "gen_cpp/RuntimeProfile_types.h"
-#include "rapidjson/stringbuffer.h"
-#include "rapidjson/writer.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "util/binary_cast.hpp"
 #include "util/pretty_printer.h"
 #include "util/stopwatch.hpp"
 #include "util/telemetry/telemetry.h"
 
 namespace doris {
+class TRuntimeProfileNode;
+class TRuntimeProfileTree;
 
 // Some macro magic to generate unique ids using __COUNTER__
 #define CONCAT_IMPL(x, y) x##y
@@ -109,10 +117,6 @@ public:
         TUnit::type _type;
     };
 
-    class DerivedCounter;
-    class EventSequence;
-    class HighWaterMarkCounter;
-
     /// A counter that keeps track of the highest value seen (reporting that
     /// as value()) and the current value.
     class HighWaterMarkCounter : public Counter {
diff --git a/be/src/util/s3_uri.cpp b/be/src/util/s3_uri.cpp
index c8f5350d12..08041fece8 100644
--- a/be/src/util/s3_uri.cpp
+++ b/be/src/util/s3_uri.cpp
@@ -17,9 +17,9 @@
 
 #include "util/s3_uri.h"
 
+#include <algorithm>
 #include <vector>
 
-#include "common/logging.h"
 #include "gutil/strings/split.h"
 #include "gutil/strings/strip.h"
 
diff --git a/be/src/util/s3_util.cpp b/be/src/util/s3_util.cpp
index 1c01297548..02508904aa 100644
--- a/be/src/util/s3_util.cpp
+++ b/be/src/util/s3_util.cpp
@@ -17,10 +17,21 @@
 
 #include "util/s3_util.h"
 
+#include <aws/core/auth/AWSAuthSigner.h>
 #include <aws/core/auth/AWSCredentials.h>
+#include <aws/core/client/ClientConfiguration.h>
+#include <aws/core/utils/logging/LogLevel.h>
+#include <aws/core/utils/logging/LogSystemInterface.h>
+#include <aws/core/utils/memory/stl/AWSStringStream.h>
 #include <aws/s3/S3Client.h>
 #include <util/string_util.h>
 
+#include <atomic>
+#include <cstdlib>
+#include <functional>
+#include <ostream>
+#include <utility>
+
 #include "common/config.h"
 #include "common/logging.h"
 #include "s3_uri.h"
diff --git a/be/src/util/s3_util.h b/be/src/util/s3_util.h
index 382a203d5b..6107fae4a7 100644
--- a/be/src/util/s3_util.h
+++ b/be/src/util/s3_util.h
@@ -18,10 +18,14 @@
 #pragma once
 
 #include <aws/core/Aws.h>
+#include <fmt/format.h>
+#include <stdint.h>
 
 #include <map>
 #include <memory>
+#include <mutex>
 #include <string>
+#include <unordered_map>
 
 #include "common/status.h"
 #include "gutil/hash/hash.h"
diff --git a/be/src/util/slice.h b/be/src/util/slice.h
index 1a7ee78a01..8049de003d 100644
--- a/be/src/util/slice.h
+++ b/be/src/util/slice.h
@@ -25,6 +25,7 @@
 #include <iostream>
 #include <map>
 #include <string>
+#include <utility>
 #include <vector>
 
 namespace doris {
diff --git a/be/src/util/sm3.cpp b/be/src/util/sm3.cpp
index 25f59c1f6b..2d4349abc8 100644
--- a/be/src/util/sm3.cpp
+++ b/be/src/util/sm3.cpp
@@ -17,6 +17,9 @@
 
 #include "util/sm3.h"
 
+#include <openssl/evp.h>
+#include <stdint.h>
+
 #include "common/logging.h"
 
 namespace doris {
diff --git a/be/src/util/sm3.h b/be/src/util/sm3.h
index d2e93988ce..7f0ab12a72 100644
--- a/be/src/util/sm3.h
+++ b/be/src/util/sm3.h
@@ -19,7 +19,8 @@
 
 #define SM3_DIGEST_LENGTH 32
 
-#include <openssl/evp.h>
+#include <openssl/ossl_typ.h>
+#include <stddef.h>
 
 #include <string>
 
diff --git a/be/src/util/sort_heap.h b/be/src/util/sort_heap.h
index 108017834b..f3f6e546b7 100644
--- a/be/src/util/sort_heap.h
+++ b/be/src/util/sort_heap.h
@@ -22,7 +22,8 @@
 #include <queue>
 #include <utility>
 
-#include "common/compiler_util.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 
 namespace doris {
 
diff --git a/be/src/util/sse_util.hpp b/be/src/util/sse_util.hpp
index 1d00a54fba..a14c281c20 100644
--- a/be/src/util/sse_util.hpp
+++ b/be/src/util/sse_util.hpp
@@ -23,8 +23,10 @@
 #if defined(__aarch64__)
 #include <sse2neon.h> // IWYU pragma: export
 #elif defined(__x86_64__)
+#include <emmintrin.h> // IWYU pragma: export
 #include <immintrin.h> // IWYU pragma: export
 #include <mm_malloc.h> // IWYU pragma: export
+#include <smmintrin.h> // IWYU pragma: export
 #endif
 
 namespace doris {
diff --git a/be/src/util/static_asserts.cpp b/be/src/util/static_asserts.cpp
index 5258a3fecd..2493587fb8 100644
--- a/be/src/util/static_asserts.cpp
+++ b/be/src/util/static_asserts.cpp
@@ -18,9 +18,11 @@
 // https://github.com/apache/impala/blob/branch-2.9.0/be/src/util/static-asserts.cpp
 // and modified by Doris
 
-#include "runtime/datetime_value.h"
+#include <stddef.h>
+
 #include "vec/common/string_ref.h"
 #include "vec/runtime/vdatetime_value.h"
+
 namespace doris {
 // This class is unused.  It contains static (compile time) asserts.
 // This is useful to validate struct sizes and other similar things
diff --git a/be/src/util/string_parser.hpp b/be/src/util/string_parser.hpp
index 4dbca73b23..2910cad9c9 100644
--- a/be/src/util/string_parser.hpp
+++ b/be/src/util/string_parser.hpp
@@ -21,20 +21,31 @@
 #pragma once
 
 #include <fast_float/fast_float.h>
+#include <fast_float/parse_number.h>
+#include <glog/logging.h>
+#include <stdlib.h>
 
-#include <cmath>
+// IWYU pragma: no_include <bits/std_abs.h>
+#include <cmath> // IWYU pragma: keep
 #include <cstdint>
-#include <cstring>
 #include <limits>
+#include <map>
 #include <string>
+#include <system_error>
 #include <type_traits>
+#include <utility>
 
-#include "common/compiler_util.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/status.h"
-#include "runtime/primitive_type.h"
 #include "vec/data_types/data_type_decimal.h"
 
 namespace doris {
+namespace vectorized {
+struct Int128I;
+template <typename T>
+struct Decimal;
+} // namespace vectorized
 
 // Utility functions for doing atoi/atof on non-null terminated strings.  On micro benchmarks,
 // this is significantly faster than libc (atoi/strtol and atof/strtod).
diff --git a/be/src/util/string_util.h b/be/src/util/string_util.h
index b0bfdbd7e9..c118e5ddab 100644
--- a/be/src/util/string_util.h
+++ b/be/src/util/string_util.h
@@ -20,7 +20,10 @@
 #include <strings.h>
 
 #include <algorithm>
+#include <boost/token_functions.hpp>
 #include <boost/tokenizer.hpp>
+#include <cctype>
+#include <cstddef>
 #include <map>
 #include <set>
 #include <sstream>
diff --git a/be/src/util/system_metrics.cpp b/be/src/util/system_metrics.cpp
index 2f259d343f..fa8f5a181a 100644
--- a/be/src/util/system_metrics.cpp
+++ b/be/src/util/system_metrics.cpp
@@ -17,13 +17,22 @@
 
 #include "util/system_metrics.h"
 
+#include <ctype.h>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
+#include <glog/logging.h>
+#include <inttypes.h>
 #include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
 
 #include <functional>
+#include <ostream>
+#include <unordered_map>
+#include <utility>
 
 #include "gutil/strings/split.h" // for string split
 #include "gutil/strtoint.h"      //  for atoi64
-#include "util/doris_metrics.h"
 #include "util/mem_info.h"
 #include "util/perf_counters.h"
 
diff --git a/be/src/util/system_metrics.h b/be/src/util/system_metrics.h
index 19ccfe0726..830b11e04c 100644
--- a/be/src/util/system_metrics.h
+++ b/be/src/util/system_metrics.h
@@ -17,8 +17,14 @@
 
 #pragma once
 
+#include <stddef.h>
+#include <stdint.h>
+
 #include <map>
 #include <memory>
+#include <set>
+#include <string>
+#include <vector>
 
 #include "util/metrics.h"
 
diff --git a/be/src/util/telemetry/brpc_carrier.cpp b/be/src/util/telemetry/brpc_carrier.cpp
index 5d49b40fff..a3c9275009 100644
--- a/be/src/util/telemetry/brpc_carrier.cpp
+++ b/be/src/util/telemetry/brpc_carrier.cpp
@@ -17,6 +17,8 @@
 
 #include "brpc_carrier.h"
 
+#include <brpc/http_header.h>
+
 opentelemetry::nostd::string_view doris::telemetry::RpcServerCarrier::Get(
         opentelemetry::nostd::string_view key) const noexcept {
     auto it = cntl_->http_request().GetHeader(key.data());
diff --git a/be/src/util/telemetry/brpc_carrier.h b/be/src/util/telemetry/brpc_carrier.h
index 60a7abf784..0e6b4d2d47 100644
--- a/be/src/util/telemetry/brpc_carrier.h
+++ b/be/src/util/telemetry/brpc_carrier.h
@@ -18,6 +18,16 @@
 #pragma once
 
 #include <brpc/controller.h>
+#include <opentelemetry/context/context.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+#include <opentelemetry/nostd/string_view.h>
+#include <opentelemetry/nostd/variant.h>
+#include <opentelemetry/trace/span.h>
+#include <opentelemetry/trace/span_context.h>
+#include <opentelemetry/trace/tracer.h>
+
+#include <string>
+#include <utility>
 
 #include "opentelemetry/context/propagation/global_propagator.h"
 #include "opentelemetry/context/propagation/text_map_propagator.h"
@@ -27,6 +37,12 @@
 #include "opentelemetry/trace/span_startoptions.h"
 #include "util/telemetry/telemetry.h"
 
+namespace google {
+namespace protobuf {
+class RpcController;
+} // namespace protobuf
+} // namespace google
+
 namespace doris::telemetry {
 
 class RpcServerCarrier : public opentelemetry::context::propagation::TextMapCarrier {
diff --git a/be/src/util/telemetry/open_telemetry_scop_wrapper.hpp b/be/src/util/telemetry/open_telemetry_scop_wrapper.hpp
index 4c30333476..2f1d8eb7b2 100644
--- a/be/src/util/telemetry/open_telemetry_scop_wrapper.hpp
+++ b/be/src/util/telemetry/open_telemetry_scop_wrapper.hpp
@@ -17,7 +17,8 @@
 
 #pragma once
 
-#include "common/compiler_util.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "opentelemetry/trace/provider.h"
 
 namespace doris {
diff --git a/be/src/util/telemetry/telemetry.cpp b/be/src/util/telemetry/telemetry.cpp
index 44904355f4..1b69da8356 100644
--- a/be/src/util/telemetry/telemetry.cpp
+++ b/be/src/util/telemetry/telemetry.cpp
@@ -17,7 +17,22 @@
 
 #include "telemetry.h"
 
+#include <fmt/format.h>
+#include <opentelemetry/exporters/otlp/otlp_environment.h>
+#include <opentelemetry/sdk/resource/resource.h>
+#include <opentelemetry/sdk/resource/resource_detector.h>
+#include <opentelemetry/sdk/trace/exporter.h>
+#include <opentelemetry/sdk/trace/processor.h>
+#include <opentelemetry/sdk/trace/recordable.h>
+#include <opentelemetry/trace/canonical_code.h>
+#include <opentelemetry/trace/provider.h>
+
 #include <boost/algorithm/string/case_conv.hpp>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <memory>
+#include <ostream>
+#include <utility>
 
 #include "common/config.h"
 #include "common/logging.h"
diff --git a/be/src/util/telemetry/telemetry.h b/be/src/util/telemetry/telemetry.h
index 5b20acc96e..c007269550 100644
--- a/be/src/util/telemetry/telemetry.h
+++ b/be/src/util/telemetry/telemetry.h
@@ -17,8 +17,19 @@
 
 #pragma once
 
+#include <opentelemetry/nostd/shared_ptr.h>
+#include <opentelemetry/trace/noop.h>
+#include <opentelemetry/trace/scope.h>
+#include <opentelemetry/trace/span.h>
+#include <opentelemetry/trace/span_context.h>
+#include <opentelemetry/trace/tracer.h>
+#include <opentelemetry/trace/tracer_provider.h>
+
+#include <string>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "opentelemetry/trace/provider.h"
-#include "util/telemetry/open_telemetry_scop_wrapper.hpp"
 
 /// A trace represents the execution process of a single request in the system, span represents a
 /// logical operation unit with start time and execution duration in the system, and multiple spans
diff --git a/be/src/util/thread.cpp b/be/src/util/thread.cpp
index aaee79fe7d..76b38195b3 100644
--- a/be/src/util/thread.cpp
+++ b/be/src/util/thread.cpp
@@ -21,6 +21,8 @@
 #include "thread.h"
 
 #ifndef __APPLE__
+// IWYU pragma: no_include <bits/types/struct_sched_param.h>
+#include <sched.h>
 #include <sys/prctl.h>
 #else
 #include <pthread.h>
@@ -28,22 +30,32 @@
 #include <cstdint>
 #endif
 
-#include <sys/types.h>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
+#include <sys/syscall.h>
+#include <time.h>
 #include <unistd.h>
 
+#include <algorithm>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
 #include <cstring>
 #include <functional>
 #include <limits>
 #include <map>
 #include <memory>
+#include <mutex>
+#include <ostream>
 #include <string>
+#include <vector>
 
 #include "common/logging.h"
 #include "gutil/atomicops.h"
 #include "gutil/dynamic_annotations.h"
 #include "gutil/map-util.h"
+#include "gutil/stringprintf.h"
 #include "gutil/strings/substitute.h"
-#include "olap/olap_define.h"
+#include "http/web_page_handler.h"
 #include "util/debug/sanitizer_scopes.h"
 #include "util/easy_json.h"
 #include "util/os_util.h"
diff --git a/be/src/util/thread.h b/be/src/util/thread.h
index 83a66fa71d..9514e4ef76 100644
--- a/be/src/util/thread.h
+++ b/be/src/util/thread.h
@@ -19,21 +19,20 @@
 // and modified by Doris
 
 #pragma once
-
-#ifndef __APPLE__
-#include <syscall.h>
-#endif
-
+#include <butil/macros.h>
 #include <pthread.h>
+#include <stdint.h>
 
-#include <atomic>
+#include <functional>
+#include <string>
+#include <utility>
 
 #include "common/status.h"
 #include "gutil/ref_counted.h"
-#include "http/web_page_handler.h"
 #include "util/countdown_latch.h"
 
 namespace doris {
+class WebPageHandler;
 
 class Thread : public RefCountedThreadSafe<Thread> {
 public:
diff --git a/be/src/util/threadpool.cpp b/be/src/util/threadpool.cpp
index adae72c623..93c14f4d61 100644
--- a/be/src/util/threadpool.cpp
+++ b/be/src/util/threadpool.cpp
@@ -20,14 +20,16 @@
 
 #include "util/threadpool.h"
 
+#include <algorithm>
 #include <cstdint>
 #include <limits>
 #include <ostream>
 #include <thread>
+#include <utility>
 
 #include "common/logging.h"
-#include "gutil/macros.h"
 #include "gutil/map-util.h"
+#include "gutil/port.h"
 #include "gutil/strings/substitute.h"
 #include "util/debug/sanitizer_scopes.h"
 #include "util/scoped_cleanup.h"
diff --git a/be/src/util/threadpool.h b/be/src/util/threadpool.h
index 89052274b0..d1b744b63e 100644
--- a/be/src/util/threadpool.h
+++ b/be/src/util/threadpool.h
@@ -20,26 +20,31 @@
 
 #pragma once
 
+#include <limits.h>
+#include <stddef.h>
+
+#include <boost/intrusive/detail/algo_type.hpp>
 #include <boost/intrusive/list.hpp>
 #include <boost/intrusive/list_hook.hpp>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
 #include <condition_variable>
 #include <deque>
 #include <functional>
+#include <iosfwd>
 #include <memory>
+#include <mutex>
 #include <string>
-#include <type_traits>
 #include <unordered_set>
-#include <utility>
 
 #include "common/status.h"
-#include "gutil/ref_counted.h"
-#include "util/priority_thread_pool.hpp"
 
 namespace doris {
 
 class Thread;
 class ThreadPool;
 class ThreadPoolToken;
+class PriorityThreadPool;
 
 class Runnable {
 public:
diff --git a/be/src/util/thrift_client.cpp b/be/src/util/thrift_client.cpp
index 82bf291047..d133b45157 100644
--- a/be/src/util/thrift_client.cpp
+++ b/be/src/util/thrift_client.cpp
@@ -17,8 +17,12 @@
 
 #include "util/thrift_client.h"
 
-#include <ostream>
+#include <thrift/transport/TTransport.h>
+#include <thrift/transport/TTransportException.h>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
 #include <string>
+#include <thread>
 
 #include "gutil/strings/substitute.h"
 
diff --git a/be/src/util/thrift_client.h b/be/src/util/thrift_client.h
index 0bc7c1b95c..2328f29845 100644
--- a/be/src/util/thrift_client.h
+++ b/be/src/util/thrift_client.h
@@ -17,24 +17,26 @@
 
 #pragma once
 
-#include <netinet/in.h>
-#include <netinet/tcp.h>
-#include <sys/socket.h>
-#include <sys/types.h>
-#include <thrift/Thrift.h>
 #include <thrift/protocol/TBinaryProtocol.h>
 #include <thrift/transport/TBufferTransports.h>
 #include <thrift/transport/TSocket.h>
 
-#include <ostream>
+#include <memory>
 #include <sstream>
 #include <string>
 
 #include "common/logging.h"
 #include "common/status.h"
-#include "gen_cpp/Types_types.h"
 #include "util/thrift_server.h"
 
+namespace apache {
+namespace thrift {
+namespace transport {
+class TTransport;
+} // namespace transport
+} // namespace thrift
+} // namespace apache
+
 namespace doris {
 // Super class for templatized thrift clients.
 class ThriftClientImpl {
diff --git a/be/src/util/thrift_rpc_helper.cpp b/be/src/util/thrift_rpc_helper.cpp
index d6889cd258..309d2fcedf 100644
--- a/be/src/util/thrift_rpc_helper.cpp
+++ b/be/src/util/thrift_rpc_helper.cpp
@@ -17,18 +17,33 @@
 
 #include "util/thrift_rpc_helper.h"
 
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <thrift/Thrift.h>
+#include <thrift/protocol/TBinaryProtocol.h>
+#include <thrift/transport/TTransportException.h>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
 #include <sstream>
 #include <thread>
 
 #include "common/status.h"
-#include "gen_cpp/FrontendService.h"
-#include "gen_cpp/FrontendService_types.h"
 #include "runtime/client_cache.h"
-#include "runtime/exec_env.h"
-#include "runtime/runtime_state.h"
+#include "runtime/exec_env.h" // IWYU pragma: keep
 #include "util/network_util.h"
-#include "util/runtime_profile.h"
-#include "util/thrift_util.h"
+
+namespace apache {
+namespace thrift {
+namespace protocol {
+class TProtocol;
+} // namespace protocol
+namespace transport {
+class TBufferedTransport;
+class TSocket;
+class TTransport;
+} // namespace transport
+} // namespace thrift
+} // namespace apache
 
 namespace doris {
 
diff --git a/be/src/util/thrift_rpc_helper.h b/be/src/util/thrift_rpc_helper.h
index 20765b5877..3b2bfb4551 100644
--- a/be/src/util/thrift_rpc_helper.h
+++ b/be/src/util/thrift_rpc_helper.h
@@ -17,14 +17,17 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include <functional>
+#include <string>
+
 #include "common/config.h"
 #include "common/status.h"
-#include "gen_cpp/FrontendService_types.h"
 
 namespace doris {
 
 class ExecEnv;
-class FrontendServiceClient;
 template <class T>
 class ClientConnection;
 
diff --git a/be/src/util/thrift_server.cpp b/be/src/util/thrift_server.cpp
index 7a11de2aa9..335a44a367 100644
--- a/be/src/util/thrift_server.cpp
+++ b/be/src/util/thrift_server.cpp
@@ -17,17 +17,22 @@
 
 #include "util/thrift_server.h"
 
-#include <thrift/concurrency/Thread.h>
+#include <glog/logging.h>
+#include <thrift/Thrift.h>
 #include <thrift/concurrency/ThreadFactory.h>
 #include <thrift/concurrency/ThreadManager.h>
 #include <thrift/protocol/TBinaryProtocol.h>
+#include <thrift/protocol/TProtocol.h>
 #include <thrift/server/TNonblockingServer.h>
 #include <thrift/server/TThreadPoolServer.h>
 #include <thrift/server/TThreadedServer.h>
+#include <thrift/transport/TBufferTransports.h>
 #include <thrift/transport/TNonblockingServerSocket.h>
 #include <thrift/transport/TServerSocket.h>
 #include <thrift/transport/TSocket.h>
-
+#include <thrift/transport/TTransport.h>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
 #include <condition_variable>
 #include <mutex>
 #include <sstream>
@@ -35,6 +40,16 @@
 
 #include "util/doris_metrics.h"
 
+namespace apache {
+namespace thrift {
+class TProcessor;
+
+namespace transport {
+class TServerTransport;
+} // namespace transport
+} // namespace thrift
+} // namespace apache
+
 namespace doris {
 
 DEFINE_GAUGE_METRIC_PROTOTYPE_3ARG(thrift_current_connections, MetricUnit::CONNECTIONS,
diff --git a/be/src/util/thrift_server.h b/be/src/util/thrift_server.h
index dacde385e6..18dd5afd57 100644
--- a/be/src/util/thrift_server.h
+++ b/be/src/util/thrift_server.h
@@ -17,16 +17,23 @@
 
 #pragma once
 
-#include <thrift/TProcessor.h>
 #include <thrift/server/TServer.h>
 
+#include <memory>
 #include <mutex>
+#include <string>
 #include <thread>
 #include <unordered_map>
 
 #include "common/status.h"
 #include "util/metrics.h"
 
+namespace apache {
+namespace thrift {
+class TProcessor;
+} // namespace thrift
+} // namespace apache
+
 namespace doris {
 // Utility class for all Thrift servers. Runs a TNonblockingServer(default) or a
 // TThreadPoolServer with, by default, 2 worker threads, that exposes the interface
@@ -137,6 +144,7 @@ private:
     // Helper class which monitors starting servers. Needs access to internal members, and
     // is not used outside of this class.
     class ThriftServerEventProcessor;
+
     friend class ThriftServerEventProcessor;
 
     std::shared_ptr<MetricEntity> _thrift_server_metric_entity;
diff --git a/be/src/util/thrift_util.cpp b/be/src/util/thrift_util.cpp
index f533eeb1c1..af0a206d4e 100644
--- a/be/src/util/thrift_util.cpp
+++ b/be/src/util/thrift_util.cpp
@@ -17,18 +17,27 @@
 
 #include "util/thrift_util.h"
 
-#include <thrift/Thrift.h>
-#include <thrift/concurrency/ThreadFactory.h>
-#include <thrift/concurrency/ThreadManager.h>
-#include <thrift/server/TNonblockingServer.h>
-#include <thrift/transport/TServerSocket.h>
+#include <gen_cpp/Types_types.h>
+#include <thrift/TOutput.h>
+#include <thrift/protocol/TBinaryProtocol.h>
 #include <thrift/transport/TSocket.h>
+#include <thrift/transport/TTransportException.h>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
 
-#include "gen_cpp/Data_types.h"
-#include "gen_cpp/Types_types.h"
-#include "util/hash_util.hpp"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/logging.h"
 #include "util/thrift_server.h"
 
+namespace apache {
+namespace thrift {
+namespace protocol {
+class TProtocol;
+} // namespace protocol
+} // namespace thrift
+} // namespace apache
+
 // TCompactProtocol requires some #defines to work right.  They also define UNLIKELY
 // so we need to undef this.
 // TODO: is there a better include to use?
@@ -39,6 +48,9 @@
 #define ARITHMETIC_RIGHT_SHIFT 1
 #include <thrift/protocol/TCompactProtocol.h>
 
+#include <sstream>
+#include <thread>
+
 namespace doris {
 
 ThriftSerializer::ThriftSerializer(bool compact, int initial_buffer_size)
diff --git a/be/src/util/thrift_util.h b/be/src/util/thrift_util.h
index 3d9a82421d..6b40dd6906 100644
--- a/be/src/util/thrift_util.h
+++ b/be/src/util/thrift_util.h
@@ -17,16 +17,27 @@
 
 #pragma once
 
+#include <stdint.h>
+#include <string.h>
 #include <thrift/TApplicationException.h>
-#include <thrift/protocol/TBinaryProtocol.h>
-#include <thrift/protocol/TDebugProtocol.h>
 #include <thrift/transport/TBufferTransports.h>
 
-#include <sstream>
+#include <exception>
+#include <memory>
+#include <string>
 #include <vector>
 
 #include "common/status.h"
 
+namespace apache {
+namespace thrift {
+namespace protocol {
+class TProtocol;
+class TProtocolFactory;
+} // namespace protocol
+} // namespace thrift
+} // namespace apache
+
 namespace doris {
 
 class TNetworkAddress;
diff --git a/be/src/util/time.cpp b/be/src/util/time.cpp
index b75ddeeb47..7a7f510905 100644
--- a/be/src/util/time.cpp
+++ b/be/src/util/time.cpp
@@ -20,9 +20,13 @@
 
 #include "util/time.h"
 
-#include <chrono>
+// IWYU pragma: no_include <bits/std_abs.h>
+#include <cmath> // IWYU pragma: keep
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
 #include <cstdlib>
 #include <iomanip>
+#include <ratio>
 #include <sstream>
 #include <thread>
 
diff --git a/be/src/util/time_lut.h b/be/src/util/time_lut.h
index 142700d079..3955d8c8ed 100644
--- a/be/src/util/time_lut.h
+++ b/be/src/util/time_lut.h
@@ -21,7 +21,6 @@
 #include <stdint.h>
 
 #include <atomic>
-#include <mutex>
 
 namespace doris {
 
diff --git a/be/src/util/timezone_utils.cpp b/be/src/util/timezone_utils.cpp
index b40a61d129..2b22e52126 100644
--- a/be/src/util/timezone_utils.cpp
+++ b/be/src/util/timezone_utils.cpp
@@ -18,6 +18,9 @@
 
 #include "util/timezone_utils.h"
 
+#include <cctz/time_zone.h>
+#include <re2/stringpiece.h>
+
 namespace doris {
 
 RE2 TimezoneUtils::time_zone_offset_format_reg("^[+-]{1}\\d{2}\\:\\d{2}$");
diff --git a/be/src/util/timezone_utils.h b/be/src/util/timezone_utils.h
index 7fbe0ef9e2..b74ce9e4b1 100644
--- a/be/src/util/timezone_utils.h
+++ b/be/src/util/timezone_utils.h
@@ -20,7 +20,11 @@
 
 #include <re2/re2.h>
 
-#include "cctz/time_zone.h"
+#include <string>
+
+namespace cctz {
+class time_zone;
+} // namespace cctz
 
 namespace doris {
 
diff --git a/be/src/util/trace.cpp b/be/src/util/trace.cpp
index 20a8a9f5c0..01b396ee1d 100644
--- a/be/src/util/trace.cpp
+++ b/be/src/util/trace.cpp
@@ -18,18 +18,22 @@
 #include "util/trace.h"
 
 #include <glog/logging.h>
-#include <rapidjson/rapidjson.h>
+#include <rapidjson/encodings.h>
+#include <stdlib.h>
+#include <time.h>
 
-#include <cstdint>
+#include <algorithm>
 #include <cstring>
 #include <iomanip>
 #include <iostream>
+#include <iterator>
 #include <map>
 #include <mutex>
 #include <string>
 #include <utility>
 #include <vector>
 
+#include "gutil/stringprintf.h"
 #include "gutil/strings/substitute.h"
 //#include "util/memory/arena.h"
 
diff --git a/be/src/util/trace.h b/be/src/util/trace.h
index 35bf9c2619..1ebd602b65 100644
--- a/be/src/util/trace.h
+++ b/be/src/util/trace.h
@@ -16,14 +16,16 @@
 // under the License.
 #pragma once
 
+#include <butil/macros.h>
+#include <rapidjson/stringbuffer.h>
 #include <rapidjson/writer.h>
+#include <stdint.h>
 
 #include <iosfwd>
-#include <memory>
 #include <string>
 #include <utility>
+#include <vector>
 
-#include "gutil/macros.h"
 #include "gutil/ref_counted.h"
 #include "gutil/strings/stringpiece.h"
 #include "gutil/strings/substitute.h"
diff --git a/be/src/util/trace_metrics.h b/be/src/util/trace_metrics.h
index ea014ae77a..37f43b5924 100644
--- a/be/src/util/trace_metrics.h
+++ b/be/src/util/trace_metrics.h
@@ -16,12 +16,13 @@
 // under the License.
 #pragma once
 
+#include <butil/macros.h>
+
 #include <cstdint>
 #include <map>
 #include <mutex>
 #include <string>
 
-#include "gutil/macros.h"
 #include "gutil/map-util.h"
 #include "util/spinlock.h"
 
diff --git a/be/src/util/uid_util.cpp b/be/src/util/uid_util.cpp
index 1fc6c4d457..6743c05a84 100644
--- a/be/src/util/uid_util.cpp
+++ b/be/src/util/uid_util.cpp
@@ -17,6 +17,12 @@
 
 #include "util/uid_util.h"
 
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/types.pb.h>
+#include <glog/logging.h>
+
+#include <cstdlib>
+
 #include "util/hash_util.hpp"
 
 namespace doris {
diff --git a/be/src/util/uid_util.h b/be/src/util/uid_util.h
index 498a1db503..c187bb9d6d 100644
--- a/be/src/util/uid_util.h
+++ b/be/src/util/uid_util.h
@@ -17,11 +17,19 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/types.pb.h>
+#include <stdint.h>
+
+#include <boost/random/mersenne_twister.hpp>
+#include <boost/uuid/random_generator.hpp>
+#include <boost/uuid/uuid.hpp>
+#include <boost/uuid/uuid_io.hpp>
+#include <cstring>
 #include <ostream>
 #include <string>
+#include <string_view>
 
-#include "gen_cpp/Types_types.h" // for TUniqueId
-#include "gen_cpp/types.pb.h"    // for PUniqueId
 #include "util/uuid_generator.h"
 
 namespace doris {
diff --git a/be/src/util/url_coding.cpp b/be/src/util/url_coding.cpp
index a3d64fb61b..0a5c4144c5 100644
--- a/be/src/util/url_coding.cpp
+++ b/be/src/util/url_coding.cpp
@@ -19,12 +19,9 @@
 
 #include <math.h>
 
-#include <exception>
 #include <memory>
 #include <sstream>
 
-#include "common/logging.h"
-
 namespace doris {
 
 static inline void url_encode(const char* in, int in_len, std::string* out) {
diff --git a/be/src/util/url_coding.h b/be/src/util/url_coding.h
index e1fd26e432..f4d63be5b1 100644
--- a/be/src/util/url_coding.h
+++ b/be/src/util/url_coding.h
@@ -17,7 +17,10 @@
 
 #pragma once
 
+#include <stddef.h>
+
 #include <cstdint>
+#include <iosfwd>
 #include <string>
 #include <vector>
 
diff --git a/be/src/util/url_parser.cpp b/be/src/util/url_parser.cpp
index e54281eb2c..149398b39a 100644
--- a/be/src/util/url_parser.cpp
+++ b/be/src/util/url_parser.cpp
@@ -17,8 +17,13 @@
 
 #include "util/url_parser.h"
 
+#include <ctype.h>
+#include <stdint.h>
+
+#include <algorithm>
 #include <string>
 
+#include "runtime/string_search.hpp"
 #include "vec/common/string_ref.h"
 
 namespace doris {
diff --git a/be/src/util/url_parser.h b/be/src/util/url_parser.h
index 882ae251ff..3c30a4ef3f 100644
--- a/be/src/util/url_parser.h
+++ b/be/src/util/url_parser.h
@@ -17,10 +17,10 @@
 
 #pragma once
 
-#include "runtime/string_search.hpp"
 #include "vec/common/string_ref.h"
 
 namespace doris {
+class StringSearch;
 
 // TODO: For now, our parse_url may not behave exactly like Hive
 // when given malformed URLs.
diff --git a/be/src/util/zlib.cpp b/be/src/util/zlib.cpp
index c4143ca232..7de336e1ac 100644
--- a/be/src/util/zlib.cpp
+++ b/be/src/util/zlib.cpp
@@ -17,17 +17,14 @@
 
 #include "util/zlib.h"
 
-#include <zconf.h>
+#include <butil/macros.h>
 #include <zlib.h>
 
-#include <cstdint>
 #include <cstring>
 #include <memory>
+#include <ostream>
 #include <string>
 
-#include "gutil/macros.h"
-#include "gutil/strings/substitute.h"
-
 using std::ostream;
 using std::string;
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function.h b/be/src/vec/aggregate_functions/aggregate_function.h
index 2d6b07e403..ea3e4897fe 100644
--- a/be/src/vec/aggregate_functions/aggregate_function.h
+++ b/be/src/vec/aggregate_functions/aggregate_function.h
@@ -22,6 +22,7 @@
 
 #include "vec/columns/column_complex.h"
 #include "vec/common/hash_table/phmap_fwd_decl.h"
+#include "vec/common/string_buffer.hpp"
 #include "vec/core/block.h"
 #include "vec/core/column_numbers.h"
 #include "vec/core/field.h"
diff --git a/be/src/vec/aggregate_functions/aggregate_function_approx_count_distinct.cpp b/be/src/vec/aggregate_functions/aggregate_function_approx_count_distinct.cpp
index 3083b9c7b1..3580ba72e8 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_approx_count_distinct.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_approx_count_distinct.cpp
@@ -17,7 +17,16 @@
 
 #include "vec/aggregate_functions/aggregate_function_approx_count_distinct.h"
 
+#include "util/bitmap_value.h"
 #include "vec/aggregate_functions/helpers.h"
+#include "vec/columns/column_array.h"
+#include "vec/columns/column_decimal.h"
+#include "vec/columns/column_map.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_struct.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/utils/template_helpers.hpp"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_approx_count_distinct.h b/be/src/vec/aggregate_functions/aggregate_function_approx_count_distinct.h
index 79a8c23e34..19c0efb13a 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_approx_count_distinct.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_approx_count_distinct.h
@@ -17,15 +17,34 @@
 
 #pragma once
 
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <string>
+
 #include "olap/hll.h"
-#include "udf/udf.h"
+#include "util/hash_util.hpp"
+#include "util/slice.h"
 #include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
-#include "vec/common/hash_table/hash.h"
+#include "vec/columns/columns_number.h"
 #include "vec/common/string_ref.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/io/io_helper.h"
 
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+class IColumn;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 struct AggregateFunctionApproxCountDistinctData {
diff --git a/be/src/vec/aggregate_functions/aggregate_function_avg.cpp b/be/src/vec/aggregate_functions/aggregate_function_avg.cpp
index f6fe08a9e3..2e58ccb56e 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_avg.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_avg.cpp
@@ -22,6 +22,7 @@
 
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
 #include "vec/aggregate_functions/helpers.h"
+#include "vec/core/field.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_avg.h b/be/src/vec/aggregate_functions/aggregate_function_avg.h
index 125ab1b8ee..1528881203 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_avg.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_avg.h
@@ -20,14 +20,40 @@
 
 #pragma once
 
-#include "common/status.h"
+#include <glog/logging.h>
+#include <string.h>
+
+#include <limits>
+#include <memory>
+#include <ostream>
+#include <type_traits>
+#include <vector>
+
+#include "runtime/decimalv2_value.h"
 #include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_fixed_length_object.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_decimal.h"
 #include "vec/data_types/data_type_fixed_length_object.h"
-#include "vec/data_types/data_type_number.h"
 #include "vec/io/io_helper.h"
 
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+template <typename T>
+class ColumnDecimal;
+template <typename T>
+class DataTypeNumber;
+template <typename>
+class ColumnVector;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 template <typename T>
diff --git a/be/src/vec/aggregate_functions/aggregate_function_avg_weighted.h b/be/src/vec/aggregate_functions/aggregate_function_avg_weighted.h
index cc14f1e1b3..c99d5df8d6 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_avg_weighted.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_avg_weighted.h
@@ -17,13 +17,34 @@
 
 #pragma once
 
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
 #include <cmath>
+#include <memory>
+#include <type_traits>
 
 #include "runtime/decimalv2_value.h"
+#include "util/binary_cast.hpp"
 #include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column_vector.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/io/io_helper.h"
 
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+class IColumn;
+template <typename T>
+class ColumnDecimal;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 template <typename T>
diff --git a/be/src/vec/aggregate_functions/aggregate_function_bit.h b/be/src/vec/aggregate_functions/aggregate_function_bit.h
index 0714ca6f34..9538633721 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_bit.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_bit.h
@@ -20,13 +20,27 @@
 
 #pragma once
 
-#include <vector>
+#include <stddef.h>
+
+#include <memory>
 
 #include "vec/aggregate_functions/aggregate_function.h"
-#include "vec/columns/column_vector.h"
-#include "vec/data_types/data_type_number.h"
+#include "vec/core/types.h"
 #include "vec/io/io_helper.h"
 
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+class IColumn;
+template <typename T>
+class DataTypeNumber;
+template <typename>
+class ColumnVector;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 template <typename T>
diff --git a/be/src/vec/aggregate_functions/aggregate_function_bitmap.cpp b/be/src/vec/aggregate_functions/aggregate_function_bitmap.cpp
index e896ab2996..35e4636dfa 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_bitmap.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_bitmap.cpp
@@ -19,6 +19,8 @@
 
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
 #include "vec/aggregate_functions/helpers.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_bitmap.h b/be/src/vec/aggregate_functions/aggregate_function_bitmap.h
index c7c537d2e3..98d8bb69f4 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_bitmap.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_bitmap.h
@@ -17,17 +17,34 @@
 
 #pragma once
 
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <string>
 #include <vector>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "util/bitmap_value.h"
 #include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column_complex.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
 #include "vec/common/assert_cast.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_bitmap.h"
-#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
-#include "vec/io/io_helper.h"
+
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+class IColumn;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_collect.cpp b/be/src/vec/aggregate_functions/aggregate_function_collect.cpp
index 957ce4f3fa..d6aab14411 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_collect.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_collect.cpp
@@ -17,6 +17,11 @@
 
 #include "vec/aggregate_functions/aggregate_function_collect.h"
 
+#include <fmt/format.h>
+
+#include <boost/iterator/iterator_facade.hpp>
+#include <type_traits>
+
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
 #include "vec/aggregate_functions/helpers.h"
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_collect.h b/be/src/vec/aggregate_functions/aggregate_function_collect.h
index f0ba7fbe77..206109e8b8 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_collect.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_collect.h
@@ -17,19 +17,43 @@
 
 #pragma once
 
-#include <algorithm>
+#include <assert.h>
+#include <glog/logging.h>
+#include <string.h>
+
+#include <limits>
+#include <memory>
+#include <new>
+#include <string>
 
-#include "common/status.h"
 #include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/aggregate_functions/key_holder_helpers.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
-#include "vec/common/aggregation_common.h"
+#include "vec/columns/column_decimal.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
 #include "vec/common/hash_table/hash_set.h"
+#include "vec/common/hash_table/hash_table_key_holder.h"
 #include "vec/common/pod_array_fwd.h"
+#include "vec/common/string_buffer.hpp"
 #include "vec/common/string_ref.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_array.h"
-#include "vec/data_types/data_type_string.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/io/io_helper.h"
+#include "vec/io/var_int.h"
+
+namespace doris {
+namespace vectorized {
+class Arena;
+} // namespace vectorized
+} // namespace doris
+template <typename, typename>
+struct DefaultHash;
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_combinator.h b/be/src/vec/aggregate_functions/aggregate_function_combinator.h
index cc62e38aea..9f67df02db 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_combinator.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_combinator.h
@@ -20,11 +20,11 @@
 
 #pragma once
 
-#include <vec/aggregate_functions/aggregate_function.h>
-#include <vec/data_types/data_type.h>
-
 #include <memory>
 
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/data_types/data_type.h"
+
 namespace doris::vectorized {
 
 /** Aggregate function combinator allows to take one aggregate function
diff --git a/be/src/vec/aggregate_functions/aggregate_function_count.cpp b/be/src/vec/aggregate_functions/aggregate_function_count.cpp
index 2fe3692d9e..8c54714b04 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_count.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_count.cpp
@@ -20,6 +20,8 @@
 
 #include "vec/aggregate_functions/aggregate_function_count.h"
 
+#include <string>
+
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
 #include "vec/aggregate_functions/factory_helpers.h"
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_count.h b/be/src/vec/aggregate_functions/aggregate_function_count.h
index 63808ac107..389e0f1df4 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_count.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_count.h
@@ -20,14 +20,31 @@
 
 #pragma once
 
-#include <array>
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <vector>
 
-#include "common/logging.h"
 #include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
 #include "vec/common/assert_cast.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_number.h"
-#include "vec/io/io_helper.h"
+#include "vec/io/var_int.h"
+
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_distinct.cpp b/be/src/vec/aggregate_functions/aggregate_function_distinct.cpp
index 4d57c689b7..5b2269a27d 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_distinct.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_distinct.cpp
@@ -20,9 +20,12 @@
 
 #include "vec/aggregate_functions/aggregate_function_distinct.h"
 
+#include <ostream>
+
 #include "vec/aggregate_functions/aggregate_function_combinator.h"
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
 #include "vec/aggregate_functions/helpers.h"
+#include "vec/data_types/data_type_nullable.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_distinct.h b/be/src/vec/aggregate_functions/aggregate_function_distinct.h
index 3961e32447..f21a2c0acc 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_distinct.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_distinct.h
@@ -20,15 +20,39 @@
 
 #pragma once
 
+#include <assert.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <memory>
+#include <new>
+#include <string>
+#include <vector>
+
 #include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/aggregate_functions/key_holder_helpers.h"
-#include "vec/common/aggregation_common.h"
+#include "vec/columns/column.h"
 #include "vec/common/assert_cast.h"
-#include "vec/common/field_visitors.h"
 #include "vec/common/hash_table/hash_set.h"
-#include "vec/common/hash_table/hash_table.h"
-#include "vec/common/sip_hash.h"
+#include "vec/common/hash_table/hash_table_key_holder.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/io/io_helper.h"
+#include "vec/io/var_int.h"
+
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+template <typename>
+class ColumnVector;
+} // namespace vectorized
+} // namespace doris
+template <typename, typename>
+struct DefaultHash;
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_group_concat.cpp b/be/src/vec/aggregate_functions/aggregate_function_group_concat.cpp
index 92a840949f..c43fb44333 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_group_concat.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_group_concat.cpp
@@ -17,6 +17,9 @@
 
 #include "vec/aggregate_functions/aggregate_function_group_concat.h"
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+
 #include "vec/aggregate_functions/helpers.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/aggregate_functions/aggregate_function_group_concat.h b/be/src/vec/aggregate_functions/aggregate_function_group_concat.h
index c4316fb2c7..6ddc600452 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_group_concat.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_group_concat.h
@@ -17,13 +17,28 @@
 
 #pragma once
 
-#include "common/status.h"
+#include <string.h>
+
+#include <memory>
+#include <string>
+
 #include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
+#include "vec/columns/column_string.h"
 #include "vec/common/string_ref.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_string.h"
 #include "vec/io/io_helper.h"
 
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+class IColumn;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 struct AggregateFunctionGroupConcatData {
diff --git a/be/src/vec/aggregate_functions/aggregate_function_histogram.cpp b/be/src/vec/aggregate_functions/aggregate_function_histogram.cpp
index b396862074..5b06af2839 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_histogram.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_histogram.cpp
@@ -17,8 +17,15 @@
 
 #include "vec/aggregate_functions/aggregate_function_histogram.h"
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+
+#include <algorithm>
+
 #include "vec/aggregate_functions/helpers.h"
 #include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_histogram.h b/be/src/vec/aggregate_functions/aggregate_function_histogram.h
index 392d02693f..4c3f6548f1 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_histogram.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_histogram.h
@@ -17,10 +17,40 @@
 
 #pragma once
 
+#include <rapidjson/stringbuffer.h>
+#include <stddef.h>
+
+#include <iterator>
+#include <map>
+#include <memory>
+#include <string>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
 #include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_decimal.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type_string.h"
+#include "vec/io/io_helper.h"
 #include "vec/utils/histogram_helpers.hpp"
 
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+template <typename>
+class ColumnVector;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 template <typename T>
diff --git a/be/src/vec/aggregate_functions/aggregate_function_hll_union_agg.cpp b/be/src/vec/aggregate_functions/aggregate_function_hll_union_agg.cpp
index ec3488e4a1..4f2232d00d 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_hll_union_agg.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_hll_union_agg.cpp
@@ -17,6 +17,8 @@
 
 #include "vec/aggregate_functions/aggregate_function_hll_union_agg.h"
 
+#include <algorithm>
+
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
 #include "vec/aggregate_functions/helpers.h"
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_hll_union_agg.h b/be/src/vec/aggregate_functions/aggregate_function_hll_union_agg.h
index 3e8ee3cb29..7db9c06875 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_hll_union_agg.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_hll_union_agg.h
@@ -17,17 +17,35 @@
 
 #pragma once
 
-#include <type_traits>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <string>
 
 #include "olap/hll.h"
+#include "util/slice.h"
 #include "vec/aggregate_functions/aggregate_function.h"
-#include "vec/columns/column_string.h"
-#include "vec/columns/column_vector.h"
+#include "vec/columns/column_complex.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_hll.h"
 #include "vec/data_types/data_type_number.h"
-#include "vec/data_types/data_type_string.h"
 #include "vec/io/io_helper.h"
 
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+class IColumn;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 struct AggregateFunctionHLLData {
diff --git a/be/src/vec/aggregate_functions/aggregate_function_max_by.cpp b/be/src/vec/aggregate_functions/aggregate_function_max_by.cpp
index b2c44689bc..35b4a4400d 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_max_by.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_max_by.cpp
@@ -15,6 +15,8 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <memory>
+
 #include "vec/aggregate_functions/aggregate_function_min_max_by.h"
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_min_by.cpp b/be/src/vec/aggregate_functions/aggregate_function_min_by.cpp
index a860bea088..6e1d96ddf5 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_min_by.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_min_by.cpp
@@ -15,6 +15,8 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <memory>
+
 #include "vec/aggregate_functions/aggregate_function_min_max_by.h"
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_min_max.cpp b/be/src/vec/aggregate_functions/aggregate_function_min_max.cpp
index d09beb1d49..27afac7789 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_min_max.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_min_max.cpp
@@ -23,6 +23,7 @@
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
 #include "vec/aggregate_functions/factory_helpers.h"
 #include "vec/aggregate_functions/helpers.h"
+#include "vec/data_types/data_type_nullable.h"
 
 namespace doris::vectorized {
 /// min, max, any
diff --git a/be/src/vec/aggregate_functions/aggregate_function_min_max.h b/be/src/vec/aggregate_functions/aggregate_function_min_max.h
index 1afd8e75f6..4d8a2ae598 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_min_max.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_min_max.h
@@ -20,15 +20,37 @@
 
 #pragma once
 
+#include <fmt/format.h>
+#include <string.h>
+
+#include <memory>
+#include <vector>
+
 #include "common/logging.h"
 #include "vec/aggregate_functions/aggregate_function.h"
-#include "vec/columns/column_decimal.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_fixed_length_object.h"
-#include "vec/columns/column_vector.h"
+#include "vec/columns/column_string.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/bit_helpers.h"
+#include "vec/common/string_buffer.hpp"
+#include "vec/common/string_ref.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_fixed_length_object.h"
+#include "vec/data_types/data_type_string.h"
 #include "vec/io/io_helper.h"
 
+namespace doris {
+namespace vectorized {
+class Arena;
+template <typename T>
+class ColumnDecimal;
+template <typename>
+class ColumnVector;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 /// For numeric values.
diff --git a/be/src/vec/aggregate_functions/aggregate_function_orthogonal_bitmap.cpp b/be/src/vec/aggregate_functions/aggregate_function_orthogonal_bitmap.cpp
index 3f182a88c3..97269ced37 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_orthogonal_bitmap.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_orthogonal_bitmap.cpp
@@ -17,9 +17,17 @@
 
 #include "vec/aggregate_functions/aggregate_function_orthogonal_bitmap.h"
 
+#include <map>
+#include <ostream>
+
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
 #include "vec/aggregate_functions/helpers.h"
-#include "vec/data_types/data_type_string.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
+
+namespace doris {
+struct StringRef;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_orthogonal_bitmap.h b/be/src/vec/aggregate_functions/aggregate_function_orthogonal_bitmap.h
index 16f16eaf98..737088de60 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_orthogonal_bitmap.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_orthogonal_bitmap.h
@@ -17,20 +17,38 @@
 
 #pragma once
 
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <string>
+#include <string_view>
+#include <type_traits>
+
 #include "util/bitmap_expr_calculation.h"
 #include "util/bitmap_intersect.h"
 #include "util/bitmap_value.h"
 #include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column_complex.h"
-#include "vec/columns/column_nullable.h"
 #include "vec/columns/column_vector.h"
-#include "vec/common/assert_cast.h"
 #include "vec/core/types.h"
 #include "vec/data_types/data_type_bitmap.h"
-#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/io/io_helper.h"
 
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+class ColumnString;
+class IColumn;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 template <typename T>
diff --git a/be/src/vec/aggregate_functions/aggregate_function_percentile_approx.cpp b/be/src/vec/aggregate_functions/aggregate_function_percentile_approx.cpp
index 1c160e7289..5c54fdf7da 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_percentile_approx.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_percentile_approx.cpp
@@ -18,7 +18,6 @@
 #include "vec/aggregate_functions/aggregate_function_percentile_approx.h"
 
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
-#include "vec/aggregate_functions/factory_helpers.h"
 #include "vec/aggregate_functions/helpers.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/aggregate_functions/aggregate_function_percentile_approx.h b/be/src/vec/aggregate_functions/aggregate_function_percentile_approx.h
index dc3e232850..b94b922840 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_percentile_approx.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_percentile_approx.h
@@ -17,16 +17,42 @@
 
 #pragma once
 
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <cmath>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <vector>
+
 #include "util/counts.h"
 #include "util/tdigest.h"
 #include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
 #include "vec/columns/column_nullable.h"
 #include "vec/columns/column_vector.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_array.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/io/io_helper.h"
 
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 struct PercentileApproxState {
diff --git a/be/src/vec/aggregate_functions/aggregate_function_quantile_state.cpp b/be/src/vec/aggregate_functions/aggregate_function_quantile_state.cpp
index 4c8ec27296..10be5a6357 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_quantile_state.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_quantile_state.cpp
@@ -17,7 +17,10 @@
 
 #include "vec/aggregate_functions/aggregate_function_quantile_state.h"
 
+#include <algorithm>
+
 #include "vec/aggregate_functions//aggregate_function_simple_factory.h"
+#include "vec/data_types/data_type.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_quantile_state.h b/be/src/vec/aggregate_functions/aggregate_function_quantile_state.h
index 6b07f79648..00f93abb41 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_quantile_state.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_quantile_state.h
@@ -17,15 +17,30 @@
 
 #pragma once
 
+#include <stddef.h>
+
+#include <memory>
+#include <string>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "util/quantile_state.h"
 #include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column_complex.h"
 #include "vec/columns/column_nullable.h"
 #include "vec/common/assert_cast.h"
-#include "vec/data_types/data_type_nullable.h"
-#include "vec/data_types/data_type_number.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_quantilestate.h"
 
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+class IColumn;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 struct AggregateFunctionQuantileStateUnionOp {
diff --git a/be/src/vec/aggregate_functions/aggregate_function_reader.cpp b/be/src/vec/aggregate_functions/aggregate_function_reader.cpp
index 034eac8313..6f27a5dd12 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_reader.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_reader.cpp
@@ -17,6 +17,18 @@
 
 #include "vec/aggregate_functions/aggregate_function_reader.h"
 
+#include <algorithm>
+#include <string>
+
+#include "vec/aggregate_functions/aggregate_function_bitmap.h"
+#include "vec/aggregate_functions/aggregate_function_hll_union_agg.h"
+#include "vec/aggregate_functions/aggregate_function_min_max.h"
+#include "vec/aggregate_functions/aggregate_function_quantile_state.h"
+#include "vec/aggregate_functions/aggregate_function_reader_first_last.h"
+#include "vec/aggregate_functions/aggregate_function_simple_factory.h"
+#include "vec/aggregate_functions/aggregate_function_sum.h"
+#include "vec/aggregate_functions/helpers.h"
+
 namespace doris::vectorized {
 
 // auto spread at nullable condition, null value do not participate aggregate
diff --git a/be/src/vec/aggregate_functions/aggregate_function_reader.h b/be/src/vec/aggregate_functions/aggregate_function_reader.h
index a062a7f496..ef35ea16fd 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_reader.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_reader.h
@@ -17,13 +17,11 @@
 
 #pragma once
 
-#include "vec/aggregate_functions/aggregate_function_bitmap.h"
-#include "vec/aggregate_functions/aggregate_function_hll_union_agg.h"
-#include "vec/aggregate_functions/aggregate_function_min_max.h"
-#include "vec/aggregate_functions/aggregate_function_quantile_state.h"
-#include "vec/aggregate_functions/aggregate_function_reader_first_last.h"
-#include "vec/aggregate_functions/aggregate_function_simple_factory.h"
-#include "vec/aggregate_functions/aggregate_function_sum.h"
+namespace doris {
+namespace vectorized {
+class AggregateFunctionSimpleFactory;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_retention.h b/be/src/vec/aggregate_functions/aggregate_function_retention.h
index 5a01e66810..8eb8db218a 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_retention.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_retention.h
@@ -21,13 +21,35 @@
 
 #pragma once
 
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+
+#include "gutil/integral_types.h"
 #include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_array.h"
-#include "vec/data_types/data_type_decimal.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/data_types/data_type_number.h"
 #include "vec/io/var_int.h"
 
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 struct RetentionState {
     static constexpr size_t MAX_EVENTS = 32;
diff --git a/be/src/vec/aggregate_functions/aggregate_function_rpc.h b/be/src/vec/aggregate_functions/aggregate_function_rpc.h
index eabfdf2fbc..40d314d2b2 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_rpc.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_rpc.h
@@ -17,11 +17,13 @@
 
 #pragma once
 
+#include <gen_cpp/Exprs_types.h>
+#include <gen_cpp/function_service.pb.h>
+
 #include <cstdint>
 #include <memory>
 
 #include "common/status.h"
-#include "gen_cpp/Exprs_types.h"
 #include "json2pb/json_to_pb.h"
 #include "json2pb/pb_to_json.h"
 #include "runtime/exec_env.h"
@@ -389,4 +391,4 @@ private:
     DataTypePtr _return_type;
 };
 
-} // namespace doris::vectorized
\ No newline at end of file
+} // namespace doris::vectorized
diff --git a/be/src/vec/aggregate_functions/aggregate_function_sequence_match.cpp b/be/src/vec/aggregate_functions/aggregate_function_sequence_match.cpp
index 4f157fa107..bc5b3cbb77 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_sequence_match.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_sequence_match.cpp
@@ -17,9 +17,23 @@
 
 #include "vec/aggregate_functions/aggregate_function_sequence_match.h"
 
+#include <boost/iterator/iterator_facade.hpp>
+
 #include "common/logging.h"
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
 #include "vec/aggregate_functions/helpers.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
+
+namespace doris {
+namespace vectorized {
+class VecDateTimeValue;
+struct DateTimeV2ValueType;
+struct DateV2ValueType;
+template <typename T>
+class DateV2Value;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_sequence_match.h b/be/src/vec/aggregate_functions/aggregate_function_sequence_match.h
index 4fe5b075f1..3e7e0b65e9 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_sequence_match.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_sequence_match.h
@@ -21,16 +21,43 @@
 
 #pragma once
 
+#include <string.h>
+
+#include <algorithm>
 #include <bitset>
+#include <boost/iterator/iterator_facade.hpp>
+#include <cstdint>
+#include <functional>
+#include <iterator>
+#include <memory>
+#include <ostream>
 #include <stack>
+#include <string>
+#include <tuple>
+#include <utility>
+#include <vector>
 
 #include "common/logging.h"
+#include "util/binary_cast.hpp"
 #include "vec/aggregate_functions/aggregate_function.h"
-#include "vec/columns/column_array.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
-#include "vec/data_types/data_type_decimal.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type_number.h"
 #include "vec/io/io_helper.h"
-#include "vec/io/var_int.h"
+
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+class IColumn;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_simple_factory.cpp b/be/src/vec/aggregate_functions/aggregate_function_simple_factory.cpp
index 5649438cdd..8c0ae92c07 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_simple_factory.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_simple_factory.cpp
@@ -20,12 +20,12 @@
 
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
 
+#include <mutex>
+
 #include "vec/aggregate_functions/aggregate_function_reader.h"
 
 namespace doris::vectorized {
 
-class AggregateFunctionSimpleFactory;
-
 void register_aggregate_function_combinator_sort(AggregateFunctionSimpleFactory& factory);
 void register_aggregate_function_combinator_distinct(AggregateFunctionSimpleFactory& factory);
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_simple_factory.h b/be/src/vec/aggregate_functions/aggregate_function_simple_factory.h
index 4ebc804d5d..bff49e9d9e 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_simple_factory.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_simple_factory.h
@@ -21,13 +21,13 @@
 #pragma once
 
 #include <functional>
-#include <iostream>
 #include <memory>
-#include <mutex>
+#include <string>
 #include <unordered_map>
+#include <utility>
+#include <vector>
 
 #include "vec/aggregate_functions/aggregate_function.h"
-#include "vec/core/field.h"
 #include "vec/data_types/data_type.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/aggregate_functions/aggregate_function_sort.cpp b/be/src/vec/aggregate_functions/aggregate_function_sort.cpp
index 8d14621709..c80d36aec5 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_sort.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_sort.cpp
@@ -17,6 +17,8 @@
 
 #include "vec/aggregate_functions/aggregate_function_sort.h"
 
+#include <algorithm>
+
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/aggregate_functions/aggregate_function_sort.h b/be/src/vec/aggregate_functions/aggregate_function_sort.h
index 4abdb9c980..c9b192e030 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_sort.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_sort.h
@@ -17,16 +17,37 @@
 
 #pragma once
 
+#include <fmt/format.h>
+#include <gen_cpp/data.pb.h>
+#include <gen_cpp/segment_v2.pb.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <memory>
+#include <new>
+#include <ostream>
 #include <string>
 #include <utility>
+#include <vector>
 
 #include "runtime/runtime_state.h"
 #include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/core/sort_block.h"
 #include "vec/core/sort_description.h"
+#include "vec/core/types.h"
 #include "vec/io/io_helper.h"
 
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 struct AggregateFunctionSortData {
diff --git a/be/src/vec/aggregate_functions/aggregate_function_stddev.cpp b/be/src/vec/aggregate_functions/aggregate_function_stddev.cpp
index 7e96c2e5a4..2c4fa002d5 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_stddev.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_stddev.cpp
@@ -17,9 +17,16 @@
 
 #include "vec/aggregate_functions/aggregate_function_stddev.h"
 
+#include <fmt/format.h>
+
+#include <string>
+
 #include "common/logging.h"
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
 #include "vec/aggregate_functions/helpers.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
+
 namespace doris::vectorized {
 
 template <template <typename, bool> class AggregateFunctionTemplate,
diff --git a/be/src/vec/aggregate_functions/aggregate_function_stddev.h b/be/src/vec/aggregate_functions/aggregate_function_stddev.h
index 625f41888b..2b38badd7b 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_stddev.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_stddev.h
@@ -17,14 +17,39 @@
 
 #pragma once
 
-#include "common/status.h"
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <cmath>
+#include <memory>
+#include <type_traits>
+
+#include "olap/olap_common.h"
+#include "runtime/decimalv2_value.h"
 #include "vec/aggregate_functions/aggregate_function.h"
-#include "vec/columns/columns_number.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
 #include "vec/common/assert_cast.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_decimal.h"
-#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/io/io_helper.h"
+
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+template <typename T>
+class ColumnDecimal;
+template <typename>
+class ColumnVector;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 template <typename T, bool is_stddev>
diff --git a/be/src/vec/aggregate_functions/aggregate_function_sum.h b/be/src/vec/aggregate_functions/aggregate_function_sum.h
index 3710f4d915..4c4dd9de26 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_sum.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_sum.h
@@ -20,12 +20,35 @@
 
 #pragma once
 
+#include <stddef.h>
+
+#include <memory>
+#include <type_traits>
+#include <vector>
+
 #include "vec/aggregate_functions/aggregate_function.h"
-#include "vec/columns/column_vector.h"
+#include "vec/columns/column.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_decimal.h"
-#include "vec/data_types/data_type_number.h"
 #include "vec/io/io_helper.h"
 
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+template <typename T>
+class ColumnDecimal;
+template <typename T>
+class DataTypeNumber;
+template <typename>
+class ColumnVector;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 template <typename T>
diff --git a/be/src/vec/aggregate_functions/aggregate_function_topn.cpp b/be/src/vec/aggregate_functions/aggregate_function_topn.cpp
index 8fa8200cb6..20513ef189 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_topn.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_topn.cpp
@@ -15,10 +15,15 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <vec/aggregate_functions/aggregate_function_topn.h>
+#include "vec/aggregate_functions/aggregate_function_topn.h"
+
+#include <fmt/format.h>
+#include <glog/logging.h>
 
 #include "vec/aggregate_functions/helpers.h"
 #include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+
 namespace doris::vectorized {
 
 AggregateFunctionPtr create_aggregate_function_topn(const std::string& name,
diff --git a/be/src/vec/aggregate_functions/aggregate_function_topn.h b/be/src/vec/aggregate_functions/aggregate_function_topn.h
index 7bfacf7977..b9ecedab19 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_topn.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_topn.h
@@ -17,23 +17,48 @@
 
 #pragma once
 
+#include <rapidjson/encodings.h>
 #include <rapidjson/stringbuffer.h>
 #include <rapidjson/writer.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <algorithm>
+#include <functional>
+#include <iterator>
+#include <memory>
+#include <string>
 #include <type_traits>
-#include <unordered_map>
+#include <utility>
+#include <vector>
 
 #include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
+#include "vec/columns/column_nullable.h"
 #include "vec/columns/column_string.h"
 #include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
 #include "vec/common/hash_table/phmap_fwd_decl.h"
+#include "vec/common/string_ref.h"
 #include "vec/core/types.h"
 #include "vec/data_types/data_type_array.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_string.h"
 #include "vec/io/io_helper.h"
 
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+template <typename T>
+class ColumnDecimal;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 // space-saving algorithm
diff --git a/be/src/vec/aggregate_functions/aggregate_function_uniq.cpp b/be/src/vec/aggregate_functions/aggregate_function_uniq.cpp
index 1016d8593b..a528162b28 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_uniq.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_uniq.cpp
@@ -20,11 +20,13 @@
 
 #include "vec/aggregate_functions/aggregate_function_uniq.h"
 
-#include "common/logging.h"
+#include <string>
+
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
-#include "vec/aggregate_functions/factory_helpers.h"
 #include "vec/aggregate_functions/helpers.h"
-#include "vec/data_types/data_type_string.h"
+#include "vec/common/hash_table/hash.h" // IWYU pragma: keep
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_uniq.h b/be/src/vec/aggregate_functions/aggregate_function_uniq.h
index 592cb5a925..0028fa7a38 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_uniq.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_uniq.h
@@ -20,18 +20,41 @@
 
 #pragma once
 
-#include <type_traits>
+#include <stddef.h>
 
-#include "gutil/hash/city.h"
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <type_traits>
+#include <vector>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "vec/aggregate_functions/aggregate_function.h"
-#include "vec/columns/column_decimal.h"
-#include "vec/common/aggregation_common.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
 #include "vec/common/assert_cast.h"
-#include "vec/common/bit_cast.h"
-#include "vec/common/hash_table/hash_set.h"
 #include "vec/common/hash_table/phmap_fwd_decl.h"
-#include "vec/common/typeid_cast.h"
+#include "vec/common/sip_hash.h"
+#include "vec/common/string_ref.h"
+#include "vec/common/uint128.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_number.h"
+#include "vec/io/io_helper.h"
+#include "vec/io/var_int.h"
+
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+template <typename T>
+class ColumnDecimal;
+} // namespace vectorized
+} // namespace doris
+template <typename T>
+struct HashCRC32;
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_window.cpp b/be/src/vec/aggregate_functions/aggregate_function_window.cpp
index a9bd6547f9..f18668690d 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_window.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_window.cpp
@@ -20,9 +20,13 @@
 
 #include "vec/aggregate_functions/aggregate_function_window.h"
 
-#include "common/logging.h"
+#include <string>
+#include <variant>
+
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
 #include "vec/aggregate_functions/helpers.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/utils/template_helpers.hpp"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/aggregate_functions/aggregate_function_window.h b/be/src/vec/aggregate_functions/aggregate_function_window.h
index 83a4586ab2..83348112c7 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_window.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_window.h
@@ -20,16 +20,34 @@
 
 #pragma once
 
-#include "factory_helpers.h"
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <ostream>
+
+#include "gutil/integral_types.h"
 #include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/aggregate_functions/aggregate_function_reader_first_last.h"
-#include "vec/aggregate_functions/helpers.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
 #include "vec/columns/column_vector.h"
-#include "vec/data_types/data_type_decimal.h"
-#include "vec/data_types/data_type_nullable.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_number.h"
-#include "vec/data_types/data_type_string.h"
-#include "vec/io/io_helper.h"
+
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/aggregate_functions/aggregate_function_window_funnel.cpp b/be/src/vec/aggregate_functions/aggregate_function_window_funnel.cpp
index 66106f54b4..6d9ca27fe1 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_window_funnel.cpp
+++ b/be/src/vec/aggregate_functions/aggregate_function_window_funnel.cpp
@@ -17,8 +17,14 @@
 
 #include "vec/aggregate_functions/aggregate_function_window_funnel.h"
 
+#include <algorithm>
+#include <ostream>
+#include <string>
+
+#include "common/logging.h"
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
 #include "vec/aggregate_functions/helpers.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_nullable.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/aggregate_functions/aggregate_function_window_funnel.h b/be/src/vec/aggregate_functions/aggregate_function_window_funnel.h
index 0b86e8f23b..e44f0204a5 100644
--- a/be/src/vec/aggregate_functions/aggregate_function_window_funnel.h
+++ b/be/src/vec/aggregate_functions/aggregate_function_window_funnel.h
@@ -21,11 +21,35 @@
 
 #pragma once
 
-#include "common/logging.h"
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <iterator>
+#include <memory>
+#include <optional>
+#include <utility>
+#include <vector>
+
+#include "util/binary_cast.hpp"
 #include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
-#include "vec/data_types/data_type_decimal.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type_number.h"
 #include "vec/io/var_int.h"
+#include "vec/runtime/vdatetime_value.h"
+
+namespace doris {
+namespace vectorized {
+class Arena;
+class BufferReadable;
+class BufferWritable;
+class IColumn;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/columns/column.cpp b/be/src/vec/columns/column.cpp
index 6f8b2c4fee..9156428b89 100644
--- a/be/src/vec/columns/column.cpp
+++ b/be/src/vec/columns/column.cpp
@@ -20,12 +20,11 @@
 
 #include "vec/columns/column.h"
 
-#include <sstream>
-
+#include "util/simd/bits.h"
 #include "vec/columns/column_const.h"
 #include "vec/columns/column_nullable.h"
-#include "vec/core/field.h"
 #include "vec/core/sort_block.h"
+#include "vec/data_types/data_type.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/columns/column.h b/be/src/vec/columns/column.h
index c7e9e6f1e1..e30c35ba31 100644
--- a/be/src/vec/columns/column.h
+++ b/be/src/vec/columns/column.h
@@ -20,6 +20,21 @@
 
 #pragma once
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stdint.h>
+#include <sys/types.h>
+
+#include <algorithm>
+#include <functional>
+#include <ostream>
+#include <string>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
+#include "gutil/integral_types.h"
 #include "olap/olap_common.h"
 #include "runtime/define_primitive_type.h"
 #include "vec/common/cow.h"
@@ -59,9 +74,8 @@ class SipHash;
 namespace doris::vectorized {
 
 class Arena;
-class Field;
-
 class ColumnSorter;
+
 using EqualFlags = std::vector<uint8_t>;
 using EqualRange = std::pair<int, int>;
 
diff --git a/be/src/vec/columns/column_array.cpp b/be/src/vec/columns/column_array.cpp
index 8908800736..4215bd36bd 100644
--- a/be/src/vec/columns/column_array.cpp
+++ b/be/src/vec/columns/column_array.cpp
@@ -20,12 +20,28 @@
 
 #include "vec/columns/column_array.h"
 
+#include <assert.h>
+#include <string.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <limits>
+#include <memory>
+#include <vector>
+
 #include "vec/columns/column_const.h"
 #include "vec/columns/column_nullable.h"
 #include "vec/columns/column_string.h"
 #include "vec/columns/columns_common.h"
 #include "vec/columns/columns_number.h"
+#include "vec/common/arena.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/memcpy_small.h"
+#include "vec/common/typeid_cast.h"
+#include "vec/common/unaligned.h"
+#include "vec/data_types/data_type.h"
+
+class SipHash;
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/columns/column_array.h b/be/src/vec/columns/column_array.h
index 2d8c9fd5d1..4f08c269fb 100644
--- a/be/src/vec/columns/column_array.h
+++ b/be/src/vec/columns/column_array.h
@@ -20,13 +20,37 @@
 
 #pragma once
 
+#include <glog/logging.h>
+#include <stdint.h>
+#include <sys/types.h>
+
+#include <functional>
+#include <ostream>
+#include <string>
+#include <type_traits>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/status.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_impl.h"
 #include "vec/columns/column_vector.h"
-#include "vec/common/arena.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/cow.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/field.h"
 #include "vec/core/types.h"
 
+class SipHash;
+
+namespace doris {
+namespace vectorized {
+class Arena;
+} // namespace vectorized
+} // namespace doris
+
 //TODO: use marcos below to decouple array function calls
 #define ALL_COLUMNS_NUMBER                                                                       \
     ColumnUInt8, ColumnInt8, ColumnInt16, ColumnInt32, ColumnInt64, ColumnInt128, ColumnFloat32, \
diff --git a/be/src/vec/columns/column_const.cpp b/be/src/vec/columns/column_const.cpp
index 0beec35355..13124dbd23 100644
--- a/be/src/vec/columns/column_const.cpp
+++ b/be/src/vec/columns/column_const.cpp
@@ -20,16 +20,19 @@
 
 #include "vec/columns/column_const.h"
 
+#include <fmt/format.h>
+
 #include <algorithm>
 #include <cstddef>
 #include <utility>
 
-#include "gutil/port.h"
 #include "runtime/raw_value.h"
+#include "util/hash_util.hpp"
 #include "vec/columns/columns_common.h"
-#include "vec/common/pod_array.h"
 #include "vec/common/sip_hash.h"
 #include "vec/common/typeid_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/columns/column_const.h b/be/src/vec/columns/column_const.h
index 3ab11d6768..bb17f7eb04 100644
--- a/be/src/vec/columns/column_const.h
+++ b/be/src/vec/columns/column_const.h
@@ -20,17 +20,39 @@
 
 #pragma once
 
-#include <cstddef>
-#include <initializer_list>
-#include <type_traits>
+#include <glog/logging.h>
+#include <stdint.h>
+#include <sys/types.h>
 
+#include <cstddef>
+#include <functional>
+#include <initializer_list>
+#include <ostream>
+#include <string>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+#include "runtime/define_primitive_type.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_nullable.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/cow.h"
+#include "vec/common/string_ref.h"
 #include "vec/common/typeid_cast.h"
-#include "vec/core/block.h"
 #include "vec/core/column_numbers.h"
 #include "vec/core/field.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+
+class SipHash;
+
+namespace doris {
+namespace vectorized {
+class Arena;
+class Block;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/columns/column_decimal.cpp b/be/src/vec/columns/column_decimal.cpp
index ab375d1d0d..880c446842 100644
--- a/be/src/vec/columns/column_decimal.cpp
+++ b/be/src/vec/columns/column_decimal.cpp
@@ -20,15 +20,24 @@
 
 #include "vec/columns/column_decimal.h"
 
-#include "common/config.h"
+#include <fmt/format.h>
+
+#include <limits>
+#include <ostream>
+#include <string>
+
+#include "olap/decimal12.h"
+#include "runtime/decimalv2_value.h"
+#include "util/hash_util.hpp"
 #include "util/simd/bits.h"
-#include "vec/columns/column_impl.h"
 #include "vec/columns/columns_common.h"
 #include "vec/common/arena.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/int_exp.h"
 #include "vec/common/sip_hash.h"
 #include "vec/common/unaligned.h"
 #include "vec/core/sort_block.h"
+#include "vec/data_types/data_type.h"
 
 template <typename T>
 bool decimal_less(T x, T y, doris::vectorized::UInt32 x_scale, doris::vectorized::UInt32 y_scale);
diff --git a/be/src/vec/columns/column_decimal.h b/be/src/vec/columns/column_decimal.h
index fe00b98b65..0096789c12 100644
--- a/be/src/vec/columns/column_decimal.h
+++ b/be/src/vec/columns/column_decimal.h
@@ -20,16 +20,37 @@
 
 #pragma once
 
-#include <cmath>
-#include <type_traits>
+#include <glog/logging.h>
+#include <stdint.h>
+#include <string.h>
+#include <sys/types.h>
 
-#include "olap/decimal12.h"
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <vector>
+
+#include "gutil/integral_types.h"
+#include "runtime/define_primitive_type.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_impl.h"
 #include "vec/columns/column_vector_helper.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/cow.h"
+#include "vec/common/pod_array.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/common/string_ref.h"
 #include "vec/common/typeid_cast.h"
 #include "vec/core/field.h"
+#include "vec/core/types.h"
+
+class SipHash;
+
+namespace doris {
+namespace vectorized {
+class Arena;
+class ColumnSorter;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/columns/column_map.cpp b/be/src/vec/columns/column_map.cpp
index ff3fc808fb..4f06ee4dda 100644
--- a/be/src/vec/columns/column_map.cpp
+++ b/be/src/vec/columns/column_map.cpp
@@ -20,6 +20,20 @@
 
 #include "vec/columns/column_map.h"
 
+#include <string.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <limits>
+#include <memory>
+#include <vector>
+
+#include "vec/common/arena.h"
+#include "vec/common/typeid_cast.h"
+#include "vec/common/unaligned.h"
+
+class SipHash;
+
 namespace doris::vectorized {
 
 /** A column of map values.
diff --git a/be/src/vec/columns/column_map.h b/be/src/vec/columns/column_map.h
index 668240ec30..6cb83adf25 100644
--- a/be/src/vec/columns/column_map.h
+++ b/be/src/vec/columns/column_map.h
@@ -20,13 +20,37 @@
 
 #pragma once
 
+#include <glog/logging.h>
+#include <stdint.h>
+#include <sys/types.h>
+
+#include <functional>
+#include <ostream>
+#include <string>
+#include <type_traits>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/status.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
 #include "vec/columns/column_impl.h"
-#include "vec/common/arena.h"
+#include "vec/columns/column_vector.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/cow.h"
+#include "vec/common/string_ref.h"
 #include "vec/core/field.h"
 #include "vec/core/types.h"
 
+class SipHash;
+
+namespace doris {
+namespace vectorized {
+class Arena;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 /** A column of map values.
diff --git a/be/src/vec/columns/column_nullable.cpp b/be/src/vec/columns/column_nullable.cpp
index 959ae55472..42cda1d18d 100644
--- a/be/src/vec/columns/column_nullable.cpp
+++ b/be/src/vec/columns/column_nullable.cpp
@@ -20,13 +20,20 @@
 
 #include "vec/columns/column_nullable.h"
 
+#include <string.h>
+
+#include <algorithm>
+
+#include "util/hash_util.hpp"
 #include "util/simd/bits.h"
 #include "vec/columns/column_const.h"
 #include "vec/common/arena.h"
 #include "vec/common/assert_cast.h"
 #include "vec/common/nan_utils.h"
+#include "vec/common/sip_hash.h"
 #include "vec/common/typeid_cast.h"
 #include "vec/core/sort_block.h"
+#include "vec/data_types/data_type.h"
 #include "vec/utils/util.hpp"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/columns/column_nullable.h b/be/src/vec/columns/column_nullable.h
index b815121e03..7d4001f6fe 100644
--- a/be/src/vec/columns/column_nullable.h
+++ b/be/src/vec/columns/column_nullable.h
@@ -20,15 +20,42 @@
 
 #pragma once
 
-#include "util/sse_util.hpp"
+#include <glog/logging.h>
+#include <stdint.h>
+#include <sys/types.h>
+
+#include <functional>
+#include <ostream>
+#include <string>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/status.h"
+#include "olap/olap_common.h"
+#include "runtime/define_primitive_type.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_impl.h"
 #include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/cow.h"
+#include "vec/common/string_ref.h"
 #include "vec/common/typeid_cast.h"
+#include "vec/core/field.h"
 #include "vec/core/types.h"
 
+class SipHash;
+
+namespace doris {
+namespace vectorized {
+class Arena;
+class ColumnSorter;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 using NullMap = ColumnUInt8::Container;
diff --git a/be/src/vec/columns/column_object.cpp b/be/src/vec/columns/column_object.cpp
index 1af9f3f630..7da8e82f39 100644
--- a/be/src/vec/columns/column_object.cpp
+++ b/be/src/vec/columns/column_object.cpp
@@ -18,22 +18,40 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Columns/ColumnObject.cpp
 // and modified by Doris
 
-#include <parallel_hashmap/phmap.h>
-#include <vec/columns/column_array.h>
-#include <vec/columns/column_nullable.h>
-#include <vec/columns/column_object.h>
-#include <vec/columns/columns_number.h>
-#include <vec/common/field_visitors.h>
-#include <vec/common/hash_table/hash_set.h>
-#include <vec/common/pod_array_fwd.h>
-#include <vec/common/schema_util.h>
-#include <vec/core/field.h>
-#include <vec/data_types/convert_field_to_type.h>
-#include <vec/data_types/data_type_array.h>
-#include <vec/data_types/data_type_nothing.h>
-#include <vec/data_types/get_least_supertype.h>
+#include "vec/columns/column_object.h"
 
-#include <vec/data_types/data_type_factory.hpp>
+#include <assert.h>
+#include <fmt/format.h>
+#include <parallel_hashmap/phmap.h>
+
+#include <functional>
+#include <limits>
+#include <map>
+#include <optional>
+
+#include "vec/columns/column_array.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/field_visitors.h"
+#include "vec/common/schema_util.h"
+#include "vec/core/field.h"
+#include "vec/data_types/convert_field_to_type.h"
+#include "vec/data_types/data_type_array.h"
+#include "vec/data_types/data_type_factory.hpp"
+#include "vec/data_types/data_type_nothing.h"
+#include "vec/data_types/get_least_supertype.h"
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/logging.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_vector.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/typeid_cast.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_decimal.h"
+#include "vec/data_types/data_type_nullable.h"
 
 namespace doris::vectorized {
 namespace {
diff --git a/be/src/vec/columns/column_object.h b/be/src/vec/columns/column_object.h
index af5cb5cc9a..b4b868c798 100644
--- a/be/src/vec/columns/column_object.h
+++ b/be/src/vec/columns/column_object.h
@@ -19,16 +19,36 @@
 // and modified by Doris
 
 #pragma once
-#include <vec/columns/column.h>
-#include <vec/columns/subcolumn_tree.h>
-#include <vec/common/hash_table/hash_map.h>
-#include <vec/common/pod_array.h>
-#include <vec/core/field.h>
-#include <vec/core/names.h>
-#include <vec/data_types/data_type.h>
-#include <vec/json/json_parser.h>
+#include <glog/logging.h>
+#include <sys/types.h>
+
+#include <algorithm>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <string_view>
+#include <unordered_set>
+#include <utility>
+#include <vector>
 
 #include "common/status.h"
+#include "vec/columns/column.h"
+#include "vec/columns/subcolumn_tree.h"
+#include "vec/common/cow.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/json/path_in_data.h"
+
+class SipHash;
+
+namespace doris {
+namespace vectorized {
+class Arena;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 /// Info that represents a scalar or array field in a decomposed view.
diff --git a/be/src/vec/columns/column_string.cpp b/be/src/vec/columns/column_string.cpp
index c6317e7453..66fec57787 100644
--- a/be/src/vec/columns/column_string.cpp
+++ b/be/src/vec/columns/column_string.cpp
@@ -20,13 +20,18 @@
 
 #include "vec/columns/column_string.h"
 
-#include "vec/columns/column_impl.h"
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <ostream>
+
+#include "util/simd/bits.h"
 #include "vec/columns/columns_common.h"
 #include "vec/common/arena.h"
 #include "vec/common/assert_cast.h"
 #include "vec/common/memcmp_small.h"
 #include "vec/common/unaligned.h"
 #include "vec/core/sort_block.h"
+#include "vec/data_types/data_type.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/columns/column_string.h b/be/src/vec/columns/column_string.h
index fc9fda41d4..c747ca2847 100644
--- a/be/src/vec/columns/column_string.h
+++ b/be/src/vec/columns/column_string.h
@@ -20,18 +20,40 @@
 
 #pragma once
 
+#include <glog/logging.h>
+#include <stdint.h>
+#include <sys/types.h>
+
 #include <cassert>
 #include <cstring>
+#include <typeinfo>
+#include <vector>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/exception.h"
+#include "common/status.h"
+#include "gutil/integral_types.h"
+#include "runtime/define_primitive_type.h"
+#include "util/hash_util.hpp"
 #include "vec/columns/column.h"
 #include "vec/columns/column_impl.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/cow.h"
 #include "vec/common/memcmp_small.h"
 #include "vec/common/memcpy_small.h"
-#include "vec/common/pod_array.h"
+#include "vec/common/pod_array_fwd.h"
 #include "vec/common/sip_hash.h"
+#include "vec/common/string_ref.h"
 #include "vec/core/field.h"
+#include "vec/core/types.h"
+
+namespace doris {
+namespace vectorized {
+class Arena;
+class ColumnSorter;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
@@ -73,7 +95,6 @@ private:
 
     template <bool positive>
     struct less;
-
     template <bool positive>
     struct lessWithCollation;
 
diff --git a/be/src/vec/columns/column_struct.cpp b/be/src/vec/columns/column_struct.cpp
index f477c20e57..2a5a505fb8 100644
--- a/be/src/vec/columns/column_struct.cpp
+++ b/be/src/vec/columns/column_struct.cpp
@@ -20,6 +20,18 @@
 
 #include "vec/columns/column_struct.h"
 
+#include <functional>
+
+#include "vec/common/assert_cast.h"
+#include "vec/common/typeid_cast.h"
+
+class SipHash;
+namespace doris {
+namespace vectorized {
+class Arena;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 std::string ColumnStruct::get_name() const {
diff --git a/be/src/vec/columns/column_struct.h b/be/src/vec/columns/column_struct.h
index 88a95767f6..dfa8bd6f5d 100644
--- a/be/src/vec/columns/column_struct.h
+++ b/be/src/vec/columns/column_struct.h
@@ -20,15 +20,33 @@
 
 #pragma once
 
+#include <glog/logging.h>
+#include <stdint.h>
+#include <sys/types.h>
+
+#include <algorithm>
+#include <ostream>
+#include <string>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_impl.h"
-#include "vec/columns/column_vector.h"
-#include "vec/common/arena.h"
-#include "vec/common/assert_cast.h"
-#include "vec/common/typeid_cast.h"
+#include "vec/common/cow.h"
+#include "vec/common/string_ref.h"
 #include "vec/core/field.h"
 #include "vec/core/types.h"
 
+class SipHash;
+
+namespace doris {
+namespace vectorized {
+class Arena;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 /** Column, that is just group of few another columns.
diff --git a/be/src/vec/columns/column_vector.cpp b/be/src/vec/columns/column_vector.cpp
index 45c8b94ddb..388b436bc5 100644
--- a/be/src/vec/columns/column_vector.cpp
+++ b/be/src/vec/columns/column_vector.cpp
@@ -20,23 +20,26 @@
 
 #include "vec/columns/column_vector.h"
 
+#include <fmt/format.h>
 #include <pdqsort.h>
-#include <vec/common/radix_sort.h>
 
-#include <cmath>
-#include <cstring>
+#include <limits>
+#include <ostream>
+#include <string>
 
+#include "util/hash_util.hpp"
 #include "util/simd/bits.h"
-#include "util/stack_util.h"
 #include "vec/columns/column_impl.h"
 #include "vec/columns/columns_common.h"
 #include "vec/common/arena.h"
 #include "vec/common/assert_cast.h"
 #include "vec/common/bit_cast.h"
 #include "vec/common/nan_utils.h"
+#include "vec/common/radix_sort.h"
 #include "vec/common/sip_hash.h"
 #include "vec/common/unaligned.h"
 #include "vec/core/sort_block.h"
+#include "vec/data_types/data_type.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/columns/column_vector.h b/be/src/vec/columns/column_vector.h
index 791893f1c4..c30796792e 100644
--- a/be/src/vec/columns/column_vector.h
+++ b/be/src/vec/columns/column_vector.h
@@ -20,16 +20,47 @@
 
 #pragma once
 
-#include <cmath>
-#include <type_traits>
+#include <glog/logging.h>
+#include <stdint.h>
+#include <string.h>
+#include <sys/types.h>
 
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <cmath>
+#include <initializer_list>
+#include <string>
+#include <type_traits>
+#include <typeinfo>
+#include <vector>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/status.h"
+#include "gutil/integral_types.h"
 #include "olap/uint24.h"
+#include "runtime/define_primitive_type.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_impl.h"
 #include "vec/columns/column_vector_helper.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/cow.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/common/string_ref.h"
+#include "vec/common/uint128.h"
 #include "vec/common/unaligned.h"
 #include "vec/core/field.h"
+#include "vec/core/types.h"
+#include "vec/runtime/vdatetime_value.h"
+
+class SipHash;
+
+namespace doris {
+namespace vectorized {
+class Arena;
+class ColumnSorter;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/columns/columns_common.cpp b/be/src/vec/columns/columns_common.cpp
index 4a4200a002..4888db3415 100644
--- a/be/src/vec/columns/columns_common.cpp
+++ b/be/src/vec/columns/columns_common.cpp
@@ -20,11 +20,21 @@
 
 #include "vec/columns/columns_common.h"
 
+#include <string.h>
+
+#include <boost/iterator/iterator_facade.hpp>
+
 #include "util/simd/bits.h"
 #include "util/sse_util.hpp"
 #include "vec/columns/column.h"
-#include "vec/columns/column_array.h"
-#include "vec/columns/column_vector.h"
+#include "vec/columns/column_array.h" // IWYU pragma: keep
+
+namespace doris {
+namespace vectorized {
+template <typename T>
+class ColumnVector;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/columns/columns_common.h b/be/src/vec/columns/columns_common.h
index 6102cef86b..bd5985f9ec 100644
--- a/be/src/vec/columns/columns_common.h
+++ b/be/src/vec/columns/columns_common.h
@@ -20,7 +20,20 @@
 
 #pragma once
 
+#include <glog/logging.h>
+#include <stdint.h>
+#include <sys/types.h>
+
+#include <ostream>
+#include <string>
+#include <vector>
+
+#include "common/exception.h"
+#include "common/status.h"
 #include "vec/columns/column.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/common/typeid_cast.h"
+#include "vec/core/types.h"
 
 /// Common helper methods for implementation of different columns.
 
diff --git a/be/src/vec/columns/subcolumn_tree.h b/be/src/vec/columns/subcolumn_tree.h
index f3de45e054..0a12593870 100644
--- a/be/src/vec/columns/subcolumn_tree.h
+++ b/be/src/vec/columns/subcolumn_tree.h
@@ -19,10 +19,10 @@
 // and modified by Doris
 
 #pragma once
-#include <vec/columns/column.h>
-#include <vec/common/hash_table/hash_map.h>
-#include <vec/data_types/data_type.h>
-#include <vec/json/path_in_data.h>
+#include "vec/columns/column.h"
+#include "vec/common/hash_table/hash_map.h"
+#include "vec/data_types/data_type.h"
+#include "vec/json/path_in_data.h"
 namespace doris::vectorized {
 /// Tree that represents paths in document
 /// with additional data in nodes.
diff --git a/be/src/vec/common/allocator.cpp b/be/src/vec/common/allocator.cpp
index edfe657c6d..e6b163734b 100644
--- a/be/src/vec/common/allocator.cpp
+++ b/be/src/vec/common/allocator.cpp
@@ -17,7 +17,18 @@
 
 #include "vec/common/allocator.h"
 
+#include <glog/logging.h>
+
+#include <atomic>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <memory>
+#include <new>
+#include <thread>
+
 // Allocator is used by too many files. For compilation speed, put dependencies in `.cpp` as much as possible.
+#include "runtime/memory/mem_tracker_limiter.h"
+#include "runtime/memory/thread_mem_tracker_mgr.h"
 #include "runtime/thread_context.h"
 #include "util/mem_info.h"
 
diff --git a/be/src/vec/common/allocator.h b/be/src/vec/common/allocator.h
index 7ae7ff7082..ea2a2b1dbd 100644
--- a/be/src/vec/common/allocator.h
+++ b/be/src/vec/common/allocator.h
@@ -23,12 +23,14 @@
 // TODO: Readable
 
 #include <fmt/format.h>
+#include <stdint.h>
 #include <string.h>
 
 #include "common/config.h"
 #include "common/status.h"
 #include "runtime/memory/chunk.h"
 #include "runtime/memory/chunk_allocator.h"
+#include "util/sse_util.hpp"
 
 #ifdef NDEBUG
 #define ALLOCATOR_ASLR 0
@@ -37,7 +39,6 @@
 #endif
 
 #if !defined(__APPLE__) && !defined(__FreeBSD__)
-#include <malloc.h>
 #else
 #define _DARWIN_C_SOURCE
 #endif
@@ -46,14 +47,15 @@
 
 #include <algorithm>
 #include <cstdlib>
+#include <string>
 
-#include "common/compiler_util.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #ifdef THREAD_SANITIZER
 /// Thread sanitizer does not intercept mremap. The usage of mremap will lead to false positives.
 #define DISABLE_MREMAP 1
 #endif
 #include "common/exception.h"
-#include "vec/common/allocator_fwd.h"
 #include "vec/common/mremap.h"
 
 /// Required for older Darwin builds, that lack definition of MAP_ANONYMOUS
diff --git a/be/src/vec/common/arena.h b/be/src/vec/common/arena.h
index 1fd52dca10..d98f4eeaa7 100644
--- a/be/src/vec/common/arena.h
+++ b/be/src/vec/common/arena.h
@@ -31,6 +31,7 @@
 #endif
 #include "gutil/dynamic_annotations.h"
 #include "vec/common/allocator.h"
+#include "vec/common/allocator_fwd.h"
 #include "vec/common/memcpy_small.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/common/hash_table/hash_map.h b/be/src/vec/common/hash_table/hash_map.h
index 813950533c..347371fbd9 100644
--- a/be/src/vec/common/hash_table/hash_map.h
+++ b/be/src/vec/common/hash_table/hash_map.h
@@ -247,4 +247,4 @@ using HashMapWithStackMemory = HashMapTable<
         Key, HashMapCellWithSavedHash<Key, Mapped, Hash>, Hash,
         HashTableGrower<initial_size_degree>,
         HashTableAllocatorWithStackMemory<(1ULL << initial_size_degree) *
-                                          sizeof(HashMapCellWithSavedHash<Key, Mapped, Hash>)>>;
\ No newline at end of file
+                                          sizeof(HashMapCellWithSavedHash<Key, Mapped, Hash>)>>;
diff --git a/be/src/vec/common/hash_table/hash_table_allocator.h b/be/src/vec/common/hash_table/hash_table_allocator.h
index 77a58ce14b..2159bf474a 100644
--- a/be/src/vec/common/hash_table/hash_table_allocator.h
+++ b/be/src/vec/common/hash_table/hash_table_allocator.h
@@ -21,6 +21,7 @@
 #pragma once
 
 #include "vec/common/allocator.h"
+#include "vec/common/allocator_fwd.h"
 
 /**
   * We are going to use the entire memory we allocated when resizing a hash
diff --git a/be/src/vec/common/hash_table/phmap_fwd_decl.h b/be/src/vec/common/hash_table/phmap_fwd_decl.h
index d37eb898e1..bee55fc51d 100644
--- a/be/src/vec/common/hash_table/phmap_fwd_decl.h
+++ b/be/src/vec/common/hash_table/phmap_fwd_decl.h
@@ -17,7 +17,7 @@
 
 #pragma once
 
-#include <parallel_hashmap/phmap.h>
+#include <parallel_hashmap/phmap.h> // IWYU pragma: export
 
 #include "vec/common/allocator.h"
 
diff --git a/be/src/vec/common/hex.h b/be/src/vec/common/hex.h
index b84aabad21..114639a842 100644
--- a/be/src/vec/common/hex.h
+++ b/be/src/vec/common/hex.h
@@ -19,7 +19,6 @@
 // and modified by Doris
 
 #pragma once
-#include <cstddef>
 #include <cstring>
 #include <string>
 
diff --git a/be/src/vec/common/mremap.cpp b/be/src/vec/common/mremap.cpp
index 32152717a7..d7cd1b8d0c 100644
--- a/be/src/vec/common/mremap.cpp
+++ b/be/src/vec/common/mremap.cpp
@@ -20,9 +20,9 @@
 
 #include "vec/common/mremap.h"
 
-#include <errno.h>
+// IWYU pragma: no_include <bthread/errno.h>
+#include <errno.h> // IWYU pragma: keep
 
-#include <cstddef>
 #include <cstdlib>
 #include <cstring>
 
diff --git a/be/src/vec/common/mremap.h b/be/src/vec/common/mremap.h
index eeeee20e96..48fb62f8a2 100644
--- a/be/src/vec/common/mremap.h
+++ b/be/src/vec/common/mremap.h
@@ -20,9 +20,6 @@
 
 #pragma once
 
-#include <sys/types.h>
-
-#include <cstddef>
 #if !defined(_MSC_VER)
 #include <sys/mman.h>
 #endif
diff --git a/be/src/vec/common/pod_array.h b/be/src/vec/common/pod_array.h
index 8179338548..d01d354c7c 100644
--- a/be/src/vec/common/pod_array.h
+++ b/be/src/vec/common/pod_array.h
@@ -21,19 +21,23 @@
 #pragma once
 
 #include <common/compiler_util.h>
+#include <stdint.h>
 #include <string.h>
+#include <sys/types.h>
 
 #include <algorithm>
-#include <boost/iterator_adaptors.hpp>
-#include <boost/noncopyable.hpp>
+#include <boost/core/noncopyable.hpp>
+#include <boost/iterator/iterator_adaptor.hpp>
 #include <cassert>
 #include <cstddef>
-#include <memory>
+#include <initializer_list>
+#include <utility>
 
-#include "vec/common/allocator.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "vec/common/allocator.h" // IWYU pragma: keep
 #include "vec/common/bit_helpers.h"
 #include "vec/common/memcpy_small.h"
-#include "vec/common/strong_typedef.h"
 
 #ifndef NDEBUG
 #include <sys/mman.h>
diff --git a/be/src/vec/common/schema_util.cpp b/be/src/vec/common/schema_util.cpp
index 4afca9097a..7bcfacb1a0 100644
--- a/be/src/vec/common/schema_util.cpp
+++ b/be/src/vec/common/schema_util.cpp
@@ -15,27 +15,47 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <vec/columns/column_array.h>
-#include <vec/columns/column_object.h>
-#include <vec/common/schema_util.h>
-#include <vec/core/field.h>
-#include <vec/data_types/data_type_array.h>
-#include <vec/functions/simple_function_factory.h>
+#include "vec/common/schema_util.h"
 
-#include <vec/data_types/data_type_factory.hpp>
+#include <assert.h>
+#include <fmt/format.h>
+#include <gen_cpp/FrontendService.h>
+#include <gen_cpp/FrontendService_types.h>
+#include <gen_cpp/HeartbeatService_types.h>
+#include <gen_cpp/MasterService_types.h>
+#include <gen_cpp/Status_types.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+
+#include <algorithm>
+#include <memory>
+#include <ostream>
+#include <unordered_set>
+#include <utility>
 #include <vector>
 
-#include "gen_cpp/FrontendService.h"
-#include "gen_cpp/HeartbeatService_types.h"
-#include "olap/rowset/rowset_writer_context.h"
+#include "common/config.h"
+#include "olap/olap_common.h"
 #include "runtime/client_cache.h"
 #include "runtime/exec_env.h"
 #include "util/thrift_rpc_helper.h"
 #include "vec/columns/column.h"
+#include "vec/columns/column_array.h"
+#include "vec/columns/column_object.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/typeid_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/field.h"
+#include "vec/core/names.h"
 #include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_array.h"
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/functions/function.h"
+#include "vec/functions/simple_function_factory.h"
+#include "vec/json/path_in_data.h"
 
 namespace doris::vectorized::schema_util {
 
diff --git a/be/src/vec/common/schema_util.h b/be/src/vec/common/schema_util.h
index 6ecc9b308f..f0d86afbb4 100644
--- a/be/src/vec/common/schema_util.h
+++ b/be/src/vec/common/schema_util.h
@@ -17,20 +17,32 @@
 
 #pragma once
 
+#include <gen_cpp/Descriptors_types.h>
 #include <parallel_hashmap/phmap.h>
-#include <vec/columns/column_object.h>
-#include <vec/common/field_visitors.h>
-#include <vec/core/block.h>
-#include <vec/core/column_with_type_and_name.h>
-#include <vec/data_types/data_type_number.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <map>
+#include <mutex>
+#include <string>
+
+#include "common/status.h"
 #include "olap/tablet_schema.h"
-#include "runtime/descriptors.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/field.h"
 #include "vec/data_types/data_type.h"
 
 namespace doris {
 class LocalSchemaChangeRecorder;
-}
+enum class FieldType;
+
+namespace vectorized {
+class Block;
+class IColumn;
+struct ColumnWithTypeAndName;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized::schema_util {
 /// Returns number of dimensions in Array type. 0 if type is not array.
diff --git a/be/src/vec/common/sip_hash.h b/be/src/vec/common/sip_hash.h
index e9afdca10a..b9416ef34a 100644
--- a/be/src/vec/common/sip_hash.h
+++ b/be/src/vec/common/sip_hash.h
@@ -36,7 +36,8 @@
 #include <string>
 #include <type_traits>
 
-#include "common/compiler_util.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "vec/common/unaligned.h"
 #include "vec/core/types.h"
 
diff --git a/be/src/vec/common/sort/heap_sorter.cpp b/be/src/vec/common/sort/heap_sorter.cpp
index 406176ff51..3e417b8504 100644
--- a/be/src/vec/common/sort/heap_sorter.cpp
+++ b/be/src/vec/common/sort/heap_sorter.cpp
@@ -17,7 +17,25 @@
 
 #include "vec/common/sort/heap_sorter.h"
 
+#include <glog/logging.h>
+
+#include <algorithm>
+
+#include "runtime/thread_context.h"
 #include "util/defer_op.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/common/sort/vsort_exec_exprs.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/sort_description.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/exprs/vexpr_context.h"
+
+namespace doris {
+class ObjectPool;
+class RowDescriptor;
+class RuntimeState;
+} // namespace doris
 
 namespace doris::vectorized {
 HeapSorter::HeapSorter(VSortExecExprs& vsort_exec_exprs, int limit, int64_t offset,
diff --git a/be/src/vec/common/sort/heap_sorter.h b/be/src/vec/common/sort/heap_sorter.h
index 9b3f27ebf2..eba924cb9f 100644
--- a/be/src/vec/common/sort/heap_sorter.h
+++ b/be/src/vec/common/sort/heap_sorter.h
@@ -16,9 +16,30 @@
 // under the License.
 
 #pragma once
-#include <queue>
+#include <gen_cpp/Metrics_types.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <memory>
+#include <queue>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
+#include "util/runtime_profile.h"
 #include "vec/common/sort/sorter.h"
+#include "vec/core/block.h"
+#include "vec/core/field.h"
+#include "vec/core/sort_cursor.h"
+
+namespace doris {
+class ObjectPool;
+class RowDescriptor;
+class RuntimeState;
+namespace vectorized {
+class VSortExecExprs;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/common/sort/sorter.cpp b/be/src/vec/common/sort/sorter.cpp
index 2e630fb44c..982aafc1a8 100644
--- a/be/src/vec/common/sort/sorter.cpp
+++ b/be/src/vec/common/sort/sorter.cpp
@@ -17,10 +17,32 @@
 
 #include "vec/common/sort/sorter.h"
 
+#include <glog/logging.h>
+
+#include <algorithm>
+#include <cstddef>
+#include <functional>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/object_pool.h"
 #include "runtime/block_spill_manager.h"
+#include "runtime/exec_env.h"
 #include "runtime/thread_context.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
 #include "vec/core/block_spill_reader.h"
 #include "vec/core/block_spill_writer.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/sort_block.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/exprs/vexpr_context.h"
+
+namespace doris {
+class RowDescriptor;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/common/sort/sorter.h b/be/src/vec/common/sort/sorter.h
index 880e70f726..3cf45cd537 100644
--- a/be/src/vec/common/sort/sorter.h
+++ b/be/src/vec/common/sort/sorter.h
@@ -16,18 +16,32 @@
 // under the License.
 
 #pragma once
-#include <queue>
+#include <gen_cpp/Metrics_types.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <deque>
+#include <memory>
+#include <queue>
+#include <vector>
 
-#include "common/consts.h"
 #include "common/status.h"
+#include "runtime/runtime_state.h"
+#include "util/runtime_profile.h"
 #include "vec/common/sort/vsort_exec_exprs.h"
 #include "vec/core/block.h"
 #include "vec/core/block_spill_reader.h"
-#include "vec/core/sort_block.h"
+#include "vec/core/field.h"
 #include "vec/core/sort_cursor.h"
+#include "vec/core/sort_description.h"
 #include "vec/runtime/vsorted_run_merger.h"
 #include "vec/utils/util.hpp"
 
+namespace doris {
+class ObjectPool;
+class RowDescriptor;
+} // namespace doris
+
 namespace doris::vectorized {
 
 // TODO: now we only use merge sort
diff --git a/be/src/vec/common/sort/topn_sorter.cpp b/be/src/vec/common/sort/topn_sorter.cpp
index ec242c3ac2..21c42eff0d 100644
--- a/be/src/vec/common/sort/topn_sorter.cpp
+++ b/be/src/vec/common/sort/topn_sorter.cpp
@@ -17,6 +17,26 @@
 
 #include "vec/common/sort/topn_sorter.h"
 
+#include <glog/logging.h>
+
+#include <algorithm>
+#include <queue>
+
+#include "common/object_pool.h"
+#include "vec/core/block.h"
+#include "vec/core/sort_cursor.h"
+#include "vec/utils/util.hpp"
+
+namespace doris {
+class RowDescriptor;
+class RuntimeProfile;
+class RuntimeState;
+
+namespace vectorized {
+class VSortExecExprs;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 TopNSorter::TopNSorter(VSortExecExprs& vsort_exec_exprs, int limit, int64_t offset,
diff --git a/be/src/vec/common/sort/topn_sorter.h b/be/src/vec/common/sort/topn_sorter.h
index fd432b6fab..a286eedc05 100644
--- a/be/src/vec/common/sort/topn_sorter.h
+++ b/be/src/vec/common/sort/topn_sorter.h
@@ -16,10 +16,27 @@
 // under the License.
 
 #pragma once
-#include <queue>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <memory>
+#include <vector>
+
+#include "common/status.h"
 #include "vec/common/sort/sorter.h"
 
+namespace doris {
+class ObjectPool;
+class RowDescriptor;
+class RuntimeProfile;
+class RuntimeState;
+
+namespace vectorized {
+class Block;
+class VSortExecExprs;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 class TopNSorter final : public Sorter {
diff --git a/be/src/vec/common/sort/vsort_exec_exprs.cpp b/be/src/vec/common/sort/vsort_exec_exprs.cpp
index 373b6b4c34..536914770d 100644
--- a/be/src/vec/common/sort/vsort_exec_exprs.cpp
+++ b/be/src/vec/common/sort/vsort_exec_exprs.cpp
@@ -17,6 +17,21 @@
 
 #include "vec/common/sort/vsort_exec_exprs.h"
 
+#include <gen_cpp/Exprs_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <stddef.h>
+
+#include "vec/exprs/vexpr.h"
+
+namespace doris {
+class ObjectPool;
+class RowDescriptor;
+class RuntimeState;
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 Status VSortExecExprs::init(const TSortInfo& sort_info, ObjectPool* pool) {
diff --git a/be/src/vec/common/sort/vsort_exec_exprs.h b/be/src/vec/common/sort/vsort_exec_exprs.h
index c030c86fb2..14ce9ffbcb 100644
--- a/be/src/vec/common/sort/vsort_exec_exprs.h
+++ b/be/src/vec/common/sort/vsort_exec_exprs.h
@@ -17,12 +17,17 @@
 
 #pragma once
 
-#include "runtime/runtime_state.h"
-#include "vec/exprs/vexpr.h"
+#include <vector>
+
+#include "common/status.h"
 
 namespace doris {
 
-class MemTracker;
+class ObjectPool;
+class RowDescriptor;
+class RuntimeState;
+class TExpr;
+class TSortInfo;
 
 // Helper class to Prepare() , Open() and Close() the ordering expressions used to perform
 // comparisons in a sort. Used by TopNNode, SortNode.  When two
@@ -31,6 +36,7 @@ class MemTracker;
 // If _materialize_tuple is true, SortExecExprs also stores the slot expressions used to
 // materialize the sort tuples.
 namespace vectorized {
+class VExprContext;
 
 class VSortExecExprs {
 public:
diff --git a/be/src/vec/common/string_ref.cpp b/be/src/vec/common/string_ref.cpp
index d292b8160f..9b36293d6f 100644
--- a/be/src/vec/common/string_ref.cpp
+++ b/be/src/vec/common/string_ref.cpp
@@ -20,7 +20,8 @@
 
 #include "string_ref.h"
 
-#include "common/compiler_util.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 
 namespace doris {
 
diff --git a/be/src/vec/common/string_ref.h b/be/src/vec/common/string_ref.h
index 212024fb06..abc745b1db 100644
--- a/be/src/vec/common/string_ref.h
+++ b/be/src/vec/common/string_ref.h
@@ -20,8 +20,13 @@
 
 #pragma once
 
+// IWYU pragma: no_include <crc32intrin.h>
+#include <glog/logging.h>
+#include <stdint.h>
+
+#include <algorithm>
 #include <climits>
-#include <functional>
+#include <cstring>
 #include <ostream>
 #include <string>
 #include <string_view>
@@ -29,7 +34,7 @@
 
 #include "gutil/hash/city.h"
 #include "gutil/hash/hash128to64.h"
-#include "util/cpu_info.h"
+#include "gutil/int128.h"
 #include "util/hash_util.hpp"
 #include "util/slice.h"
 #include "util/sse_util.hpp"
diff --git a/be/src/vec/common/string_utils/string_utils.h b/be/src/vec/common/string_utils/string_utils.h
index f8c64013e3..327bd2dc43 100644
--- a/be/src/vec/common/string_utils/string_utils.h
+++ b/be/src/vec/common/string_utils/string_utils.h
@@ -23,7 +23,6 @@
 #include <cstddef>
 #include <cstring>
 #include <string>
-#include <type_traits>
 
 namespace detail {
 bool starts_with(const std::string& s, const char* prefix, size_t prefix_size);
diff --git a/be/src/vec/core/block.cpp b/be/src/vec/core/block.cpp
index dc7cef55c1..638e643d1e 100644
--- a/be/src/vec/core/block.cpp
+++ b/be/src/vec/core/block.cpp
@@ -20,30 +20,47 @@
 
 #include "vec/core/block.h"
 
+#include <assert.h>
 #include <fmt/format.h>
-#include <glog/logging.h>
+#include <gen_cpp/data.pb.h>
 #include <snappy.h>
+#include <sys/types.h>
+
+#include <algorithm>
+#include <iomanip>
+#include <iterator>
+#include <limits>
 
 #include "agent/be_exec_version_manager.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/config.h"
+#include "common/logging.h"
 #include "common/status.h"
 #include "runtime/descriptors.h"
 #include "runtime/thread_context.h"
-#include "udf/udf.h"
 #include "util/block_compression.h"
 #include "util/faststring.h"
+#include "util/runtime_profile.h"
 #include "util/simd/bits.h"
+#include "util/slice.h"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column.h"
-#include "vec/columns/column_array.h"
 #include "vec/columns/column_const.h"
 #include "vec/columns/column_nullable.h"
 #include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
 #include "vec/common/assert_cast.h"
-#include "vec/common/schema_util.h"
-#include "vec/common/string_ref.h"
-#include "vec/common/typeid_cast.h"
 #include "vec/data_types/data_type_factory.hpp"
 
+class SipHash;
+
+namespace doris {
+namespace segment_v2 {
+enum CompressionTypePB : int;
+} // namespace segment_v2
+} // namespace doris
+
 namespace doris::vectorized {
 
 Block::Block(std::initializer_list<ColumnWithTypeAndName> il) : data {il} {
diff --git a/be/src/vec/core/block.h b/be/src/vec/core/block.h
index dba2e62ea3..26d78566f1 100644
--- a/be/src/vec/core/block.h
+++ b/be/src/vec/core/block.h
@@ -20,28 +20,41 @@
 
 #pragma once
 
+#include <glog/logging.h>
 #include <parallel_hashmap/phmap.h>
+#include <stddef.h>
+#include <stdint.h>
 
 #include <initializer_list>
 #include <list>
+#include <memory>
+#include <ostream>
 #include <set>
+#include <string>
 #include <utility>
 #include <vector>
 
-#include "gen_cpp/data.pb.h"
-#include "runtime/descriptors.h"
+#include "common/status.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_nullable.h"
 #include "vec/core/column_with_type_and_name.h"
 #include "vec/core/columns_with_type_and_name.h"
 #include "vec/core/names.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_nullable.h"
 
+class SipHash;
+
 namespace doris {
 
-class RowDescriptor;
-class Status;
 class TupleDescriptor;
+class PBlock;
+class SlotDescriptor;
+
+namespace segment_v2 {
+enum CompressionTypePB : int;
+} // namespace segment_v2
 
 namespace vectorized {
 
@@ -52,6 +65,7 @@ namespace vectorized {
   * Allows to insert, remove columns in arbitrary position, to change order of columns.
   */
 class MutableBlock;
+
 class Block {
 private:
     using Container = ColumnsWithTypeAndName;
diff --git a/be/src/vec/core/block_spill_reader.cpp b/be/src/vec/core/block_spill_reader.cpp
index be14bc4d55..a1f6e29abe 100644
--- a/be/src/vec/core/block_spill_reader.cpp
+++ b/be/src/vec/core/block_spill_reader.cpp
@@ -17,13 +17,26 @@
 
 #include "vec/core/block_spill_reader.h"
 
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/data.pb.h>
+#include <glog/logging.h>
+#include <unistd.h>
+
+#include <algorithm>
+
 #include "io/file_factory.h"
-#include "io/fs/file_system.h"
+#include "io/fs/file_reader.h"
 #include "io/fs/local_file_system.h"
-#include "olap/iterators.h"
 #include "runtime/block_spill_manager.h"
+#include "runtime/exec_env.h"
+#include "util/slice.h"
+#include "vec/core/block.h"
 
 namespace doris {
+namespace io {
+class FileSystem;
+} // namespace io
+
 namespace vectorized {
 void BlockSpillReader::_init_profile() {
     read_time_ = ADD_TIMER(profile_, "ReadTime");
diff --git a/be/src/vec/core/block_spill_reader.h b/be/src/vec/core/block_spill_reader.h
index 5b46541f84..a82024ca14 100644
--- a/be/src/vec/core/block_spill_reader.h
+++ b/be/src/vec/core/block_spill_reader.h
@@ -17,12 +17,21 @@
 
 #pragma once
 
-#include "io/fs/file_reader.h"
+#include <stddef.h>
+#include <stdint.h>
+
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "common/status.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "util/runtime_profile.h"
-#include "vec/core/block.h"
 
 namespace doris {
 namespace vectorized {
+class Block;
+
 // Read data spilled to local file.
 class BlockSpillReader {
 public:
diff --git a/be/src/vec/core/block_spill_writer.cpp b/be/src/vec/core/block_spill_writer.cpp
index a501253c63..8694be20d3 100644
--- a/be/src/vec/core/block_spill_writer.cpp
+++ b/be/src/vec/core/block_spill_writer.cpp
@@ -17,9 +17,20 @@
 
 #include "vec/core/block_spill_writer.h"
 
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/data.pb.h>
+#include <gen_cpp/segment_v2.pb.h>
+#include <unistd.h>
+
+#include <algorithm>
+
 #include "agent/be_exec_version_manager.h"
 #include "io/file_factory.h"
-#include "runtime/runtime_state.h"
+#include "runtime/exec_env.h"
+#include "runtime/thread_context.h"
+#include "vec/columns/column.h"
+#include "vec/core/column_with_type_and_name.h"
 
 namespace doris {
 namespace vectorized {
diff --git a/be/src/vec/core/block_spill_writer.h b/be/src/vec/core/block_spill_writer.h
index dad8e836a7..d3cf0b8479 100644
--- a/be/src/vec/core/block_spill_writer.h
+++ b/be/src/vec/core/block_spill_writer.h
@@ -17,9 +17,17 @@
 
 #pragma once
 
-#include "io/fs/local_file_writer.h"
+#include <stddef.h>
+#include <stdint.h>
+
+#include <memory>
+#include <string>
+
+#include "common/status.h"
+#include "io/fs/file_writer.h"
 #include "util/runtime_profile.h"
 #include "vec/core/block.h"
+
 namespace doris {
 namespace vectorized {
 
diff --git a/be/src/vec/core/column_with_type_and_name.cpp b/be/src/vec/core/column_with_type_and_name.cpp
index 38f22db882..9ac2bbe6e4 100644
--- a/be/src/vec/core/column_with_type_and_name.cpp
+++ b/be/src/vec/core/column_with_type_and_name.cpp
@@ -18,10 +18,18 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Core/ColumnWithTypeAndName.cpp
 // and modified by Doris
 
-#include <ostream>
-#include <sstream>
+#include "vec/core/column_with_type_and_name.h"
 
-#include "vec/core/columns_with_type_and_name.h"
+#include <gen_cpp/data.pb.h>
+#include <stddef.h>
+
+#include <memory>
+#include <sstream>
+#include <string>
+
+#include "vec/columns/column.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/core/column_with_type_and_name.h b/be/src/vec/core/column_with_type_and_name.h
index 8bffdfa04c..caf68f4626 100644
--- a/be/src/vec/core/column_with_type_and_name.h
+++ b/be/src/vec/core/column_with_type_and_name.h
@@ -20,9 +20,19 @@
 
 #pragma once
 
-#include "vec/columns/column.h"
+#include <stddef.h>
+
+#include <iosfwd>
+#include <memory>
+#include <string>
+
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
 
+namespace doris {
+class PColumnMeta;
+} // namespace doris
+
 namespace doris::vectorized {
 
 // class WriteBuffer;
diff --git a/be/src/vec/core/field.cpp b/be/src/vec/core/field.cpp
index ca0e783064..aeb55c1d79 100644
--- a/be/src/vec/core/field.cpp
+++ b/be/src/vec/core/field.cpp
@@ -20,8 +20,18 @@
 
 #include "vec/core/field.h"
 
+#include "vec/core/accurate_comparison.h"
 #include "vec/core/decimal_comparison.h"
+#include "vec/data_types/data_type_decimal.h"
 #include "vec/io/io_helper.h"
+#include "vec/io/var_int.h"
+
+namespace doris {
+namespace vectorized {
+class BufferReadable;
+class BufferWritable;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/core/field.h b/be/src/vec/core/field.h
index b8b3f20700..4c4896ce9a 100644
--- a/be/src/vec/core/field.h
+++ b/be/src/vec/core/field.h
@@ -20,17 +20,38 @@
 
 #pragma once
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stdint.h>
+#include <string.h>
+
 #include <algorithm>
 #include <cassert>
 #include <functional>
+#include <map>
+#include <new>
+#include <ostream>
+#include <string>
+#include <string_view>
 #include <type_traits>
+#include <utility>
 #include <vector>
 
-#include "vec/common/int_exp.h"
-#include "vec/common/strong_typedef.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "vec/common/uint128.h"
 #include "vec/core/types.h"
 
+namespace doris {
+namespace vectorized {
+template <typename T>
+struct TypeId;
+template <typename T>
+struct TypeName;
+} // namespace vectorized
+struct PackedInt128;
+} // namespace doris
+
 namespace doris::vectorized {
 
 template <typename T>
@@ -70,6 +91,7 @@ struct AvgNearestFieldTypeTrait<Int64> {
 };
 
 class Field;
+
 using FieldVector = std::vector<Field>;
 
 /// Array and Tuple use the same storage type -- FieldVector, but we declare
@@ -1214,7 +1236,4 @@ std::enable_if_t<!std::is_same_v<std::decay_t<T>, Field>, Field&> Field::operato
     return *this;
 }
 
-class ReadBuffer;
-class WriteBuffer;
-
 } // namespace doris::vectorized
diff --git a/be/src/vec/core/materialize_block.cpp b/be/src/vec/core/materialize_block.cpp
index 18e6e46d8e..294800202d 100644
--- a/be/src/vec/core/materialize_block.cpp
+++ b/be/src/vec/core/materialize_block.cpp
@@ -19,6 +19,12 @@
 // and modified by Doris
 
 #include "vec/core/materialize_block.h"
+
+#include <stddef.h>
+
+#include "vec/columns/column.h"
+#include "vec/core/column_with_type_and_name.h"
+
 namespace doris::vectorized {
 
 Block materialize_block(const Block& block) {
diff --git a/be/src/vec/core/sort_block.cpp b/be/src/vec/core/sort_block.cpp
index bf7ca56d95..e31b18a9fc 100644
--- a/be/src/vec/core/sort_block.cpp
+++ b/be/src/vec/core/sort_block.cpp
@@ -20,8 +20,7 @@
 
 #include "vec/core/sort_block.h"
 
-#include "vec/columns/column_string.h"
-#include "vec/common/typeid_cast.h"
+#include "vec/core/column_with_type_and_name.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/core/sort_block.h b/be/src/vec/core/sort_block.h
index 9d4b9ca2b7..6b24c160dc 100644
--- a/be/src/vec/core/sort_block.h
+++ b/be/src/vec/core/sort_block.h
@@ -19,11 +19,36 @@
 // and modified by Doris
 
 #pragma once
+#include <glog/logging.h>
 #include <pdqsort.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <utility>
+#include <vector>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "util/simd/bits.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/common/memcmp_small.h"
+#include "vec/common/string_ref.h"
 #include "vec/core/block.h"
 #include "vec/core/sort_description.h"
+#include "vec/core/types.h"
+
+namespace doris {
+namespace vectorized {
+template <typename T>
+class ColumnDecimal;
+template <typename>
+class ColumnVector;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/convert_field_to_type.cpp b/be/src/vec/data_types/convert_field_to_type.cpp
index c516702678..bc7a821e64 100644
--- a/be/src/vec/data_types/convert_field_to_type.cpp
+++ b/be/src/vec/data_types/convert_field_to_type.cpp
@@ -18,19 +18,34 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Interpreters/convert_field_to_type.cpp
 // and modified by Doris
 
-#include <vec/common/field_visitors.h>
-#include <vec/common/typeid_cast.h>
-#include <vec/core/accurate_comparison.h>
-#include <vec/data_types/convert_field_to_type.h>
-#include <vec/data_types/data_type_array.h>
-#include <vec/data_types/data_type_date.h>
-#include <vec/data_types/data_type_date_time.h>
-#include <vec/data_types/data_type_decimal.h>
-#include <vec/data_types/data_type_nullable.h>
-#include <vec/data_types/data_type_number.h>
-#include <vec/data_types/data_type_object.h>
-#include <vec/data_types/data_type_string.h>
-// #include <vec/data_types/data_type_tuple.h>
+#include "vec/data_types/convert_field_to_type.h"
+
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <memory>
+#include <ostream>
+#include <string>
+#include <type_traits>
+#include <vector>
+
+#include "vec/common/field_visitors.h"
+#include "vec/common/typeid_cast.h"
+#include "vec/core/accurate_comparison.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_array.h"
+#include "vec/data_types/data_type_nullable.h"
+
+namespace doris {
+namespace vectorized {
+struct UInt128;
+} // namespace vectorized
+} // namespace doris
+
+// #include "vec/data_types/data_type_tuple.h"
 namespace doris::vectorized {
 /** Checking for a `Field from` of `From` type falls to a range of values of type `To`.
   * `From` and `To` - numeric types. They can be floating-point types.
diff --git a/be/src/vec/data_types/convert_field_to_type.h b/be/src/vec/data_types/convert_field_to_type.h
index 1f5567a139..8d57994c31 100644
--- a/be/src/vec/data_types/convert_field_to_type.h
+++ b/be/src/vec/data_types/convert_field_to_type.h
@@ -19,7 +19,9 @@
 // and modified by Doris
 
 #pragma once
-#include <vec/core/field.h>
+#include "common/status.h"
+#include "vec/core/field.h"
+
 namespace doris::vectorized {
 
 class IDataType;
diff --git a/be/src/vec/data_types/data_type.cpp b/be/src/vec/data_types/data_type.cpp
index aee0630585..458a204576 100644
--- a/be/src/vec/data_types/data_type.cpp
+++ b/be/src/vec/data_types/data_type.cpp
@@ -21,10 +21,23 @@
 #include "vec/data_types/data_type.h"
 
 #include <fmt/format.h>
+#include <gen_cpp/data.pb.h>
+#include <gen_cpp/types.pb.h>
+
+#include <algorithm>
+#include <utility>
 
 #include "common/logging.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
+#include "vec/core/field.h"
+
+namespace doris {
+namespace vectorized {
+class BufferWritable;
+class ReadBuffer;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type.h b/be/src/vec/data_types/data_type.h
index d05560bd5c..9af716c5ca 100644
--- a/be/src/vec/data_types/data_type.h
+++ b/be/src/vec/data_types/data_type.h
@@ -20,26 +20,32 @@
 
 #pragma once
 
-#include <boost/noncopyable.hpp>
-#include <memory>
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+#include <stdint.h>
 
-#include "gen_cpp/data.pb.h"
+#include <boost/core/noncopyable.hpp>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "common/status.h"
 #include "runtime/define_primitive_type.h"
 #include "vec/common/cow.h"
-#include "vec/common/string_buffer.hpp"
 #include "vec/core/types.h"
 #include "vec/data_types/serde/data_type_serde.h"
-#include "vec/io/reader_buffer.h"
 
 namespace doris {
-class PBlock;
-class PColumn;
+class PColumnMeta;
+enum PGenericType_TypeId : int;
 
 namespace vectorized {
 
 class IDataType;
-
 class IColumn;
+class BufferWritable;
+class ReadBuffer;
+
 using ColumnPtr = COW<IColumn>::Ptr;
 using MutableColumnPtr = COW<IColumn>::MutablePtr;
 
diff --git a/be/src/vec/data_types/data_type_array.cpp b/be/src/vec/data_types/data_type_array.cpp
index d691594d55..f79f660a51 100644
--- a/be/src/vec/data_types/data_type_array.cpp
+++ b/be/src/vec/data_types/data_type_array.cpp
@@ -20,14 +20,26 @@
 
 #include "vec/data_types/data_type_array.h"
 
-#include "gen_cpp/data.pb.h"
-#include "util/stack_util.h"
+#include <ctype.h>
+#include <gen_cpp/data.pb.h>
+#include <glog/logging.h>
+#include <string.h>
+
+#include <typeinfo>
+#include <utility>
+
+#include "runtime/decimalv2_value.h"
+#include "util/types.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
 #include "vec/columns/column_const.h"
 #include "vec/columns/column_nullable.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/string_buffer.hpp"
+#include "vec/common/string_ref.h"
+#include "vec/common/typeid_cast.h"
 #include "vec/data_types/data_type_nullable.h"
+#include "vec/io/reader_buffer.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_array.h b/be/src/vec/data_types/data_type_array.h
index 47a23320c3..9409102ee1 100644
--- a/be/src/vec/data_types/data_type_array.h
+++ b/be/src/vec/data_types/data_type_array.h
@@ -21,9 +21,29 @@
 #pragma once
 
 #include <gen_cpp/Types_types.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <memory>
+#include <string>
+
+#include "common/status.h"
+#include "runtime/define_primitive_type.h"
 #include "serde/data_type_array_serde.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
+#include "vec/data_types/serde/data_type_serde.h"
+
+namespace doris {
+class PColumnMeta;
+
+namespace vectorized {
+class BufferWritable;
+class IColumn;
+class ReadBuffer;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_bitmap.cpp b/be/src/vec/data_types/data_type_bitmap.cpp
index 7e15dadc8e..b7b5ba1aa1 100644
--- a/be/src/vec/data_types/data_type_bitmap.cpp
+++ b/be/src/vec/data_types/data_type_bitmap.cpp
@@ -17,11 +17,15 @@
 
 #include "vec/data_types/data_type_bitmap.h"
 
+#include <utility>
+
 #include "util/bitmap_value.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_complex.h"
 #include "vec/columns/column_const.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/string_buffer.hpp"
+#include "vec/common/string_ref.h"
 #include "vec/io/io_helper.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/data_types/data_type_bitmap.h b/be/src/vec/data_types/data_type_bitmap.h
index 8946c2fec6..082b7ff2d3 100644
--- a/be/src/vec/data_types/data_type_bitmap.h
+++ b/be/src/vec/data_types/data_type_bitmap.h
@@ -17,13 +17,31 @@
 
 #pragma once
 #include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <memory>
+#include <ostream>
+#include <string>
+#include <typeinfo>
+
+#include "runtime/define_primitive_type.h"
 #include "serde/data_type_bitmap_serde.h"
 #include "util/bitmap_value.h"
-#include "vec/columns/column.h"
 #include "vec/columns/column_complex.h"
+#include "vec/core/field.h"
 #include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
+#include "vec/data_types/serde/data_type_serde.h"
+
+namespace doris {
+namespace vectorized {
+class BufferReadable;
+class BufferWritable;
+class IColumn;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 class DataTypeBitMap : public IDataType {
diff --git a/be/src/vec/data_types/data_type_date.cpp b/be/src/vec/data_types/data_type_date.cpp
index 4c57985253..219cb351e2 100644
--- a/be/src/vec/data_types/data_type_date.cpp
+++ b/be/src/vec/data_types/data_type_date.cpp
@@ -20,13 +20,22 @@
 
 #include "vec/data_types/data_type_date.h"
 
-#include "runtime/datetime_value.h"
+#include <typeinfo>
+#include <utility>
+
 #include "util/binary_cast.hpp"
+#include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/string_buffer.hpp"
 #include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/io/io_helper.h"
+#include "vec/io/reader_buffer.h"
 #include "vec/runtime/vdatetime_value.h"
+
 namespace doris::vectorized {
 bool DataTypeDate::equals(const IDataType& rhs) const {
     return typeid(rhs) == typeid(*this);
diff --git a/be/src/vec/data_types/data_type_date.h b/be/src/vec/data_types/data_type_date.h
index c22373863e..6648932d5f 100644
--- a/be/src/vec/data_types/data_type_date.h
+++ b/be/src/vec/data_types/data_type_date.h
@@ -20,8 +20,27 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <string>
+
+#include "common/status.h"
+#include "runtime/define_primitive_type.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_number_base.h"
 
+namespace doris {
+namespace vectorized {
+class BufferWritable;
+class ReadBuffer;
+class IColumn;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 class DataTypeDate final : public DataTypeNumberBase<Int64> {
diff --git a/be/src/vec/data_types/data_type_date_time.cpp b/be/src/vec/data_types/data_type_date_time.cpp
index 8fe6c351ee..04da44a0a6 100644
--- a/be/src/vec/data_types/data_type_date_time.cpp
+++ b/be/src/vec/data_types/data_type_date_time.cpp
@@ -20,13 +20,22 @@
 
 #include "vec/data_types/data_type_date_time.h"
 
-#include "runtime/datetime_value.h"
+#include <typeinfo>
+#include <utility>
+
 #include "util/binary_cast.hpp"
+#include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/string_buffer.hpp"
 #include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/io/io_helper.h"
+#include "vec/io/reader_buffer.h"
 #include "vec/runtime/vdatetime_value.h"
+
 namespace doris::vectorized {
 
 bool DataTypeDateTime::equals(const IDataType& rhs) const {
diff --git a/be/src/vec/data_types/data_type_date_time.h b/be/src/vec/data_types/data_type_date_time.h
index a82bc6dbde..c4056f3612 100644
--- a/be/src/vec/data_types/data_type_date_time.h
+++ b/be/src/vec/data_types/data_type_date_time.h
@@ -20,12 +20,28 @@
 
 #pragma once
 
-#include "runtime/define_primitive_type.h"
-#include "vec/data_types/data_type_date.h"
-#include "vec/data_types/data_type_number_base.h"
-#include "vec/data_types/data_type_time_v2.h"
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
 
-class DateLUTImpl;
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <string>
+
+#include "common/status.h"
+#include "runtime/define_primitive_type.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_number_base.h"
+
+namespace doris {
+namespace vectorized {
+class BufferWritable;
+class ReadBuffer;
+class IColumn;
+class DataTypeDate;
+class DataTypeDateV2;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_decimal.cpp b/be/src/vec/data_types/data_type_decimal.cpp
index 482b32f81a..a588c1d0ae 100644
--- a/be/src/vec/data_types/data_type_decimal.cpp
+++ b/be/src/vec/data_types/data_type_decimal.cpp
@@ -21,16 +21,22 @@
 #include "vec/data_types/data_type_decimal.h"
 
 #include <fmt/format.h>
+#include <gen_cpp/data.pb.h>
+#include <string.h>
 
 #include <utility>
 
-#include "gen_cpp/data.pb.h"
+#include "runtime/decimalv2_value.h"
+#include "util/string_parser.hpp"
 #include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
+#include "vec/columns/columns_number.h"
 #include "vec/common/assert_cast.h"
 #include "vec/common/int_exp.h"
+#include "vec/common/string_buffer.hpp"
 #include "vec/common/typeid_cast.h"
 #include "vec/io/io_helper.h"
+#include "vec/io/reader_buffer.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_decimal.h b/be/src/vec/data_types/data_type_decimal.h
index d4f9690c69..a3672b05df 100644
--- a/be/src/vec/data_types/data_type_decimal.h
+++ b/be/src/vec/data_types/data_type_decimal.h
@@ -19,18 +19,48 @@
 // and modified by Doris
 
 #pragma once
+#include <fmt/format.h>
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
 #include <cmath>
+#include <limits>
+#include <memory>
+#include <ostream>
+#include <string>
 #include <type_traits>
 
-#include "common/config.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/logging.h"
+#include "common/status.h"
+#include "olap/olap_common.h"
 #include "runtime/define_primitive_type.h"
 #include "serde/data_type_decimal_serde.h"
+#include "util/binary_cast.hpp"
 #include "vec/columns/column_decimal.h"
 #include "vec/common/arithmetic_overflow.h"
 #include "vec/common/typeid_cast.h"
+#include "vec/core/field.h"
 #include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
-#include "vec/data_types/data_type_number.h"
+#include "vec/data_types/data_type_number.h" // IWYU pragma: keep
+#include "vec/data_types/serde/data_type_serde.h"
+
+namespace doris {
+class DecimalV2Value;
+class PColumnMeta;
+
+namespace vectorized {
+class BufferWritable;
+class IColumn;
+class ReadBuffer;
+template <typename T>
+struct TypeId;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_factory.cpp b/be/src/vec/data_types/data_type_factory.cpp
index a14776fdbe..c1ca9245c7 100644
--- a/be/src/vec/data_types/data_type_factory.cpp
+++ b/be/src/vec/data_types/data_type_factory.cpp
@@ -20,12 +20,42 @@
 
 #include "vec/data_types/data_type_factory.hpp"
 
+#include <arrow/type.h>
+#include <fmt/format.h>
+#include <gen_cpp/data.pb.h>
 #include <gen_cpp/types.pb.h>
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <ostream>
+
+#include "common/consts.h"
 #include "data_type_time.h"
+#include "olap/field.h"
+#include "olap/olap_common.h"
 #include "runtime/define_primitive_type.h"
+#include "vec/common/uint128.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type_array.h"
+#include "vec/data_types/data_type_bitmap.h"
+#include "vec/data_types/data_type_date.h"
+#include "vec/data_types/data_type_date_time.h"
+#include "vec/data_types/data_type_decimal.h"
+#include "vec/data_types/data_type_fixed_length_object.h"
 #include "vec/data_types/data_type_hll.h"
+#include "vec/data_types/data_type_jsonb.h"
+#include "vec/data_types/data_type_map.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/data_types/data_type_number.h"
 #include "vec/data_types/data_type_object.h"
+#include "vec/data_types/data_type_quantilestate.h"
+#include "vec/data_types/data_type_string.h"
+#include "vec/data_types/data_type_struct.h"
+#include "vec/data_types/data_type_time_v2.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_factory.hpp b/be/src/vec/data_types/data_type_factory.hpp
index 155beac08c..6bef55938f 100644
--- a/be/src/vec/data_types/data_type_factory.hpp
+++ b/be/src/vec/data_types/data_type_factory.hpp
@@ -19,32 +19,29 @@
 // and modified by Doris
 
 #pragma once
-#include <mutex>
-#include <string>
-#include <type_traits>
+#include <gen_cpp/Types_types.h>
+
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
 
-#include "arrow/type.h"
-#include "common/consts.h"
-#include "gen_cpp/data.pb.h"
-#include "olap/field.h"
 #include "olap/tablet_schema.h"
 #include "runtime/types.h"
-#include "vec/data_types/data_type.h"
-#include "vec/data_types/data_type_array.h"
-#include "vec/data_types/data_type_bitmap.h"
-#include "vec/data_types/data_type_date.h"
-#include "vec/data_types/data_type_date_time.h"
-#include "vec/data_types/data_type_decimal.h"
-#include "vec/data_types/data_type_fixed_length_object.h"
-#include "vec/data_types/data_type_hll.h"
-#include "vec/data_types/data_type_jsonb.h"
-#include "vec/data_types/data_type_map.h"
-#include "vec/data_types/data_type_nothing.h"
-#include "vec/data_types/data_type_nullable.h"
-#include "vec/data_types/data_type_number.h"
-#include "vec/data_types/data_type_quantilestate.h"
-#include "vec/data_types/data_type_string.h"
-#include "vec/data_types/data_type_struct.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+
+namespace arrow {
+class DataType;
+} // namespace arrow
+namespace doris {
+class Field;
+class PColumnMeta;
+enum class FieldType;
+
+namespace vectorized {
+enum class TypeIndex;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_fixed_length_object.cpp b/be/src/vec/data_types/data_type_fixed_length_object.cpp
index b8413fedb9..0dcde08fec 100644
--- a/be/src/vec/data_types/data_type_fixed_length_object.cpp
+++ b/be/src/vec/data_types/data_type_fixed_length_object.cpp
@@ -17,6 +17,14 @@
 
 #include "vec/data_types/data_type_fixed_length_object.h"
 
+#include <glog/logging.h>
+#include <string.h>
+
+#include <ostream>
+
+#include "vec/columns/column.h"
+#include "vec/common/assert_cast.h"
+
 namespace doris::vectorized {
 
 char* DataTypeFixedLengthObject::serialize(const IColumn& column, char* buf,
diff --git a/be/src/vec/data_types/data_type_fixed_length_object.h b/be/src/vec/data_types/data_type_fixed_length_object.h
index 65fc01f137..92c502df61 100644
--- a/be/src/vec/data_types/data_type_fixed_length_object.h
+++ b/be/src/vec/data_types/data_type_fixed_length_object.h
@@ -17,10 +17,26 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <memory>
+#include <typeinfo>
+
+#include "runtime/define_primitive_type.h"
 #include "serde/data_type_fixedlengthobject_serde.h"
 #include "vec/columns/column_fixed_length_object.h"
+#include "vec/core/field.h"
 #include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
+#include "vec/data_types/serde/data_type_serde.h"
+
+namespace doris {
+namespace vectorized {
+class IColumn;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_hll.cpp b/be/src/vec/data_types/data_type_hll.cpp
index cae56e5b25..ee8b3a381b 100644
--- a/be/src/vec/data_types/data_type_hll.cpp
+++ b/be/src/vec/data_types/data_type_hll.cpp
@@ -17,10 +17,16 @@
 
 #include "vec/data_types/data_type_hll.h"
 
+#include <string.h>
+
+#include <utility>
+
+#include "util/slice.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_complex.h"
 #include "vec/columns/column_const.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/string_buffer.hpp"
 #include "vec/io/io_helper.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/data_types/data_type_hll.h b/be/src/vec/data_types/data_type_hll.h
index 8d7d77a79d..2a597a7a18 100644
--- a/be/src/vec/data_types/data_type_hll.h
+++ b/be/src/vec/data_types/data_type_hll.h
@@ -16,12 +16,32 @@
 // under the License.
 
 #pragma once
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <memory>
+#include <ostream>
+#include <string>
+#include <typeinfo>
+
 #include "olap/hll.h"
+#include "runtime/define_primitive_type.h"
 #include "serde/data_type_hll_serde.h"
-#include "vec/columns/column.h"
 #include "vec/columns/column_complex.h"
+#include "vec/core/field.h"
 #include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
+#include "vec/data_types/serde/data_type_serde.h"
+
+namespace doris {
+namespace vectorized {
+class BufferReadable;
+class BufferWritable;
+class IColumn;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 class DataTypeHLL : public IDataType {
diff --git a/be/src/vec/data_types/data_type_jsonb.cpp b/be/src/vec/data_types/data_type_jsonb.cpp
index 08196d9668..06708bb5d1 100644
--- a/be/src/vec/data_types/data_type_jsonb.cpp
+++ b/be/src/vec/data_types/data_type_jsonb.cpp
@@ -17,12 +17,21 @@
 
 #include "data_type_jsonb.h"
 
-#include "gen_cpp/data.pb.h"
-#include "util/sse_util.hpp"
+#include <typeinfo>
+#include <utility>
+
+#include "util/jsonb_utils.h"
 #include "vec/columns/column_const.h"
 #include "vec/common/assert_cast.h"
-#include "vec/core/field.h"
-#include "vec/io/io_helper.h"
+#include "vec/common/string_buffer.hpp"
+#include "vec/common/string_ref.h"
+#include "vec/io/reader_buffer.h"
+
+namespace doris {
+namespace vectorized {
+class IColumn;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_jsonb.h b/be/src/vec/data_types/data_type_jsonb.h
index c6460a2f36..defe1a9879 100644
--- a/be/src/vec/data_types/data_type_jsonb.h
+++ b/be/src/vec/data_types/data_type_jsonb.h
@@ -17,12 +17,31 @@
 
 #pragma once
 
-#include <ostream>
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <memory>
+#include <string>
+
+#include "common/status.h"
+#include "runtime/define_primitive_type.h"
 #include "runtime/jsonb_value.h"
 #include "vec/columns/column_string.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_string.h"
+#include "vec/data_types/serde/data_type_serde.h"
+#include "vec/data_types/serde/data_type_string_serde.h"
+
+namespace doris {
+namespace vectorized {
+class BufferWritable;
+class IColumn;
+class ReadBuffer;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 class DataTypeJsonb final : public IDataType {
diff --git a/be/src/vec/data_types/data_type_map.cpp b/be/src/vec/data_types/data_type_map.cpp
index 75664c300c..0752ad813e 100644
--- a/be/src/vec/data_types/data_type_map.cpp
+++ b/be/src/vec/data_types/data_type_map.cpp
@@ -17,7 +17,23 @@
 
 #include "data_type_map.h"
 
+#include <ctype.h>
+#include <gen_cpp/data.pb.h>
+#include <glog/logging.h>
+#include <string.h>
+
 #include <string>
+#include <typeinfo>
+
+#include "vec/columns/column.h"
+#include "vec/columns/column_array.h"
+#include "vec/columns/column_map.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/string_buffer.hpp"
+#include "vec/common/string_ref.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/io/reader_buffer.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_map.h b/be/src/vec/data_types/data_type_map.h
index db8dadff8f..e01ac90805 100644
--- a/be/src/vec/data_types/data_type_map.h
+++ b/be/src/vec/data_types/data_type_map.h
@@ -20,15 +20,30 @@
 
 #pragma once
 
-#include "gen_cpp/data.pb.h"
-#include "serde/data_type_map_serde.h"
-#include "util/stack_util.h"
-#include "vec/columns/column_array.h"
-#include "vec/columns/column_map.h"
-#include "vec/columns/column_nullable.h"
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <memory>
+#include <string>
+
+#include "common/status.h"
+#include "runtime/define_primitive_type.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
-#include "vec/data_types/data_type_array.h"
-#include "vec/data_types/data_type_nullable.h"
+#include "vec/data_types/serde/data_type_map_serde.h"
+#include "vec/data_types/serde/data_type_serde.h"
+
+namespace doris {
+class PColumnMeta;
+
+namespace vectorized {
+class BufferWritable;
+class IColumn;
+class ReadBuffer;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 /** Map data type.
diff --git a/be/src/vec/data_types/data_type_nothing.cpp b/be/src/vec/data_types/data_type_nothing.cpp
index ce2f619754..09f3421507 100644
--- a/be/src/vec/data_types/data_type_nothing.cpp
+++ b/be/src/vec/data_types/data_type_nothing.cpp
@@ -20,6 +20,8 @@
 
 #include "vec/data_types/data_type_nothing.h"
 
+#include <typeinfo>
+
 #include "vec/columns/column_nothing.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/data_types/data_type_nothing.h b/be/src/vec/data_types/data_type_nothing.h
index de27f62810..450aa2802f 100644
--- a/be/src/vec/data_types/data_type_nothing.h
+++ b/be/src/vec/data_types/data_type_nothing.h
@@ -20,8 +20,25 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <ostream>
+#include <string>
+
+#include "runtime/define_primitive_type.h"
 #include "vec/core/field.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
+#include "vec/data_types/serde/data_type_serde.h"
+
+namespace doris {
+namespace vectorized {
+class IColumn;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_nullable.cpp b/be/src/vec/data_types/data_type_nullable.cpp
index 7fe364a2d9..40b8b9661f 100644
--- a/be/src/vec/data_types/data_type_nullable.cpp
+++ b/be/src/vec/data_types/data_type_nullable.cpp
@@ -20,16 +20,24 @@
 
 #include "vec/data_types/data_type_nullable.h"
 
-#include <gen_cpp/Opcodes_types.h>
+#include <fmt/format.h>
+#include <gen_cpp/data.pb.h>
+#include <glog/logging.h>
+#include <string.h>
+
+#include <utility>
+#include <vector>
 
-#include "gen_cpp/data.pb.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/columns/columns_number.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/string_buffer.hpp"
 #include "vec/common/typeid_cast.h"
 #include "vec/core/field.h"
 #include "vec/data_types/data_type_nothing.h"
+#include "vec/io/reader_buffer.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_nullable.h b/be/src/vec/data_types/data_type_nullable.h
index 57d24c10b0..80948bdb9d 100644
--- a/be/src/vec/data_types/data_type_nullable.h
+++ b/be/src/vec/data_types/data_type_nullable.h
@@ -20,8 +20,30 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <memory>
+#include <string>
+
+#include "common/status.h"
+#include "runtime/define_primitive_type.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
 #include "vec/data_types/serde/data_type_nullable_serde.h"
+#include "vec/data_types/serde/data_type_serde.h"
+
+namespace doris {
+class PColumnMeta;
+
+namespace vectorized {
+class BufferWritable;
+class IColumn;
+class ReadBuffer;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_number_base.cpp b/be/src/vec/data_types/data_type_number_base.cpp
index 004b7fff8d..5994a7aa4c 100644
--- a/be/src/vec/data_types/data_type_number_base.cpp
+++ b/be/src/vec/data_types/data_type_number_base.cpp
@@ -20,7 +20,13 @@
 
 #include "vec/data_types/data_type_number_base.h"
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <string.h>
+
+#include <limits>
 #include <type_traits>
+#include <utility>
 
 #include "gutil/strings/numbers.h"
 #include "util/mysql_global.h"
@@ -28,7 +34,9 @@
 #include "vec/columns/column_const.h"
 #include "vec/columns/column_vector.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/string_buffer.hpp"
 #include "vec/io/io_helper.h"
+#include "vec/io/reader_buffer.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_number_base.h b/be/src/vec/data_types/data_type_number_base.h
index 4e9b3b6a37..b7f4f6bef0 100644
--- a/be/src/vec/data_types/data_type_number_base.h
+++ b/be/src/vec/data_types/data_type_number_base.h
@@ -20,12 +20,35 @@
 
 #pragma once
 
-#include <cstdint>
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
 
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <cstdint>
+#include <memory>
+#include <string>
+#include <type_traits>
+
+#include "common/status.h"
+#include "runtime/define_primitive_type.h"
 #include "serde/data_type_number_serde.h"
 #include "vec/columns/column_vector.h"
+#include "vec/common/uint128.h"
+#include "vec/core/field.h"
 #include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
+#include "vec/data_types/serde/data_type_serde.h"
+
+namespace doris {
+namespace vectorized {
+class BufferWritable;
+class IColumn;
+class ReadBuffer;
+template <typename T>
+struct TypeId;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_object.cpp b/be/src/vec/data_types/data_type_object.cpp
index 2a2b85ebf9..cd30800da6 100644
--- a/be/src/vec/data_types/data_type_object.cpp
+++ b/be/src/vec/data_types/data_type_object.cpp
@@ -18,12 +18,28 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/DataTypes/DataTypeObject.cpp
 // and modified by Doris
 
+#include "vec/data_types/data_type_object.h"
+
+#include <gen_cpp/data.pb.h>
+#include <string.h>
 #include <util/string_util.h>
-#include <vec/data_types/data_type_object.h>
-#include <vec/io/io_helper.h>
 
 #include <cassert>
-#include <vec/data_types/data_type_factory.hpp>
+#include <memory>
+#include <utility>
+#include <vector>
+
+#include "vec/columns/column_object.h"
+#include "vec/common/typeid_cast.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_factory.hpp"
+#include "vec/json/path_in_data.h"
+
+namespace doris {
+namespace vectorized {
+class IColumn;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_object.h b/be/src/vec/data_types/data_type_object.h
index 202a1fd79f..f68c64196f 100644
--- a/be/src/vec/data_types/data_type_object.h
+++ b/be/src/vec/data_types/data_type_object.h
@@ -19,11 +19,29 @@
 // and modified by Doris
 
 #pragma once
-#include <vec/columns/column_object.h>
-#include <vec/core/field.h>
-#include <vec/data_types/data_type.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <memory>
+#include <ostream>
+#include <string>
+
+#include "runtime/define_primitive_type.h"
 #include "serde/data_type_object_serde.h"
+#include "vec/columns/column_object.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/serde/data_type_serde.h"
+
+namespace doris {
+namespace vectorized {
+class IColumn;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 class DataTypeObject : public IDataType {
diff --git a/be/src/vec/data_types/data_type_quantilestate.cpp b/be/src/vec/data_types/data_type_quantilestate.cpp
index 59f7665d67..a6a37e1f07 100644
--- a/be/src/vec/data_types/data_type_quantilestate.cpp
+++ b/be/src/vec/data_types/data_type_quantilestate.cpp
@@ -17,8 +17,12 @@
 
 #include "vec/data_types/data_type_quantilestate.h"
 
+#include "util/slice.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_complex.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/string_buffer.hpp"
+#include "vec/common/string_ref.h"
 #include "vec/io/io_helper.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/data_types/data_type_quantilestate.h b/be/src/vec/data_types/data_type_quantilestate.h
index 9659e65822..71968b7f88 100644
--- a/be/src/vec/data_types/data_type_quantilestate.h
+++ b/be/src/vec/data_types/data_type_quantilestate.h
@@ -17,12 +17,32 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <memory>
+#include <ostream>
+#include <string>
+#include <typeinfo>
+
+#include "runtime/define_primitive_type.h"
 #include "serde/data_type_quantilestate_serde.h"
 #include "util/quantile_state.h"
-#include "vec/columns/column.h"
 #include "vec/columns/column_complex.h"
+#include "vec/core/field.h"
 #include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
+#include "vec/data_types/serde/data_type_serde.h"
+
+namespace doris {
+namespace vectorized {
+class BufferReadable;
+class BufferWritable;
+class IColumn;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 template <typename T>
@@ -91,4 +111,4 @@ public:
 };
 using DataTypeQuantileStateDouble = DataTypeQuantileState<double>;
 
-} // namespace doris::vectorized
\ No newline at end of file
+} // namespace doris::vectorized
diff --git a/be/src/vec/data_types/data_type_string.cpp b/be/src/vec/data_types/data_type_string.cpp
index 077d45d0b6..bf23705bb7 100644
--- a/be/src/vec/data_types/data_type_string.cpp
+++ b/be/src/vec/data_types/data_type_string.cpp
@@ -20,14 +20,19 @@
 
 #include "vec/data_types/data_type_string.h"
 
-#include <string_view>
+#include <string.h>
+
+#include <typeinfo>
+#include <utility>
 
-#include "util/sse_util.hpp"
 #include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
 #include "vec/columns/column_string.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/string_buffer.hpp"
+#include "vec/common/string_ref.h"
 #include "vec/core/field.h"
+#include "vec/io/reader_buffer.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_string.h b/be/src/vec/data_types/data_type_string.h
index ea400b1438..b7cf804897 100644
--- a/be/src/vec/data_types/data_type_string.h
+++ b/be/src/vec/data_types/data_type_string.h
@@ -20,10 +20,29 @@
 
 #pragma once
 
-#include <ostream>
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <memory>
+#include <string>
+
+#include "common/status.h"
+#include "runtime/define_primitive_type.h"
 #include "serde/data_type_string_serde.h"
+#include "vec/columns/column_string.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
+#include "vec/data_types/serde/data_type_serde.h"
+
+namespace doris {
+namespace vectorized {
+class BufferWritable;
+class IColumn;
+class ReadBuffer;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_struct.cpp b/be/src/vec/data_types/data_type_struct.cpp
index f0dfd34c8f..663f7af949 100644
--- a/be/src/vec/data_types/data_type_struct.cpp
+++ b/be/src/vec/data_types/data_type_struct.cpp
@@ -20,8 +20,28 @@
 
 #include "vec/data_types/data_type_struct.h"
 
+#include <ctype.h>
+#include <fmt/format.h>
+#include <gen_cpp/data.pb.h>
+#include <glog/logging.h>
+#include <string.h>
+
+#include <algorithm>
+#include <memory>
+#include <ostream>
+#include <typeinfo>
+#include <unordered_set>
+#include <utility>
+#include <vector>
+
+#include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
+#include "vec/columns/column_nullable.h"
 #include "vec/columns/column_struct.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/string_buffer.hpp"
+#include "vec/common/string_ref.h"
+#include "vec/io/reader_buffer.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_struct.h b/be/src/vec/data_types/data_type_struct.h
index 65f2131ac7..8b09a7cc88 100644
--- a/be/src/vec/data_types/data_type_struct.h
+++ b/be/src/vec/data_types/data_type_struct.h
@@ -20,17 +20,32 @@
 
 #pragma once
 
-#include <exception>
-#include <optional>
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+#include <stdint.h>
 
-#include "gen_cpp/data.pb.h"
-#include "serde/data_type_struct_serde.h"
-#include "util/stack_util.h"
-#include "vec/columns/column_array.h"
-#include "vec/columns/column_nullable.h"
-#include "vec/columns/column_struct.h"
+#include <memory>
+#include <optional>
+#include <string>
+#include <vector>
+
+#include "common/status.h"
+#include "runtime/define_primitive_type.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
-#include "vec/data_types/data_type_nullable.h"
+#include "vec/data_types/serde/data_type_serde.h"
+#include "vec/data_types/serde/data_type_struct_serde.h"
+
+namespace doris {
+class PColumnMeta;
+
+namespace vectorized {
+class BufferWritable;
+class IColumn;
+class ReadBuffer;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_time.cpp b/be/src/vec/data_types/data_type_time.cpp
index 38b83ebf60..7881124f3e 100644
--- a/be/src/vec/data_types/data_type_time.cpp
+++ b/be/src/vec/data_types/data_type_time.cpp
@@ -20,11 +20,21 @@
 
 #include "vec/data_types/data_type_time.h"
 
+#include <typeinfo>
+#include <utility>
+
 #include "util/date_func.h"
-#include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/string_buffer.hpp"
+
+namespace doris {
+namespace vectorized {
+class IColumn;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_time.h b/be/src/vec/data_types/data_type_time.h
index 64c06c073d..5f508f56d4 100644
--- a/be/src/vec/data_types/data_type_time.h
+++ b/be/src/vec/data_types/data_type_time.h
@@ -20,8 +20,29 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <string>
+
+#include "runtime/define_primitive_type.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/data_types/data_type_number_base.h"
+#include "vec/data_types/serde/data_type_number_serde.h"
+#include "vec/data_types/serde/data_type_serde.h"
+
+namespace doris {
+namespace vectorized {
+class BufferWritable;
+class IColumn;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/data_type_time_v2.cpp b/be/src/vec/data_types/data_type_time_v2.cpp
index 789815ca29..8baf8a6776 100644
--- a/be/src/vec/data_types/data_type_time_v2.cpp
+++ b/be/src/vec/data_types/data_type_time_v2.cpp
@@ -17,15 +17,31 @@
 
 #include "vec/data_types/data_type_time_v2.h"
 
+#include <gen_cpp/data.pb.h>
+
+#include <memory>
+#include <ostream>
 #include <string>
+#include <typeinfo>
+#include <utility>
 
 #include "util/binary_cast.hpp"
 #include "vec/columns/column_const.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/string_buffer.hpp"
 #include "vec/core/types.h"
 #include "vec/io/io_helper.h"
+#include "vec/io/reader_buffer.h"
 #include "vec/runtime/vdatetime_value.h"
 
+namespace doris {
+namespace vectorized {
+class IColumn;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 bool DataTypeDateV2::equals(const IDataType& rhs) const {
     return typeid(rhs) == typeid(*this);
diff --git a/be/src/vec/data_types/data_type_time_v2.h b/be/src/vec/data_types/data_type_time_v2.h
index a60547b562..8872eeafc0 100644
--- a/be/src/vec/data_types/data_type_time_v2.h
+++ b/be/src/vec/data_types/data_type_time_v2.h
@@ -17,8 +17,35 @@
 
 #pragma once
 
+#include <fmt/format.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <string>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/status.h"
 #include "runtime/define_primitive_type.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_number_base.h"
+#include "vec/data_types/serde/data_type_number_serde.h"
+#include "vec/data_types/serde/data_type_serde.h"
+
+namespace doris {
+class PColumnMeta;
+
+namespace vectorized {
+class BufferWritable;
+class ReadBuffer;
+class IColumn;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/get_least_supertype.cpp b/be/src/vec/data_types/get_least_supertype.cpp
index 5c135932a9..3d08a77443 100644
--- a/be/src/vec/data_types/get_least_supertype.cpp
+++ b/be/src/vec/data_types/get_least_supertype.cpp
@@ -20,9 +20,20 @@
 
 #include "vec/data_types/get_least_supertype.h"
 
-#include <unordered_set>
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <vector>
 
 #include "vec/common/typeid_cast.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_date_time.h"
 #include "vec/data_types/data_type_decimal.h"
 #include "vec/data_types/data_type_nothing.h"
diff --git a/be/src/vec/data_types/get_least_supertype.h b/be/src/vec/data_types/get_least_supertype.h
index e0ab943803..ad52f553df 100644
--- a/be/src/vec/data_types/get_least_supertype.h
+++ b/be/src/vec/data_types/get_least_supertype.h
@@ -21,9 +21,15 @@
 #pragma once
 
 #include <parallel_hashmap/phmap.h>
-#include <vec/data_types/data_type.h>
 
 #include "common/status.h"
+#include "vec/data_types/data_type.h"
+
+namespace doris {
+namespace vectorized {
+enum class TypeIndex;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/nested_utils.cpp b/be/src/vec/data_types/nested_utils.cpp
index 8a957828b8..22c9dfc6df 100644
--- a/be/src/vec/data_types/nested_utils.cpp
+++ b/be/src/vec/data_types/nested_utils.cpp
@@ -20,9 +20,7 @@
 
 #include "vec/data_types/nested_utils.h"
 
-#include "vec/columns/column_const.h"
 #include "vec/common/string_utils/string_utils.h"
-#include "vec/common/typeid_cast.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/data_types/nested_utils.h b/be/src/vec/data_types/nested_utils.h
index 8c2684e7eb..081e31d272 100644
--- a/be/src/vec/data_types/nested_utils.h
+++ b/be/src/vec/data_types/nested_utils.h
@@ -20,6 +20,9 @@
 
 #pragma once
 
+#include <string>
+#include <utility>
+
 #include "vec/core/block.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/data_types/serde/data_type_array_serde.h b/be/src/vec/data_types/serde/data_type_array_serde.h
index d7b4e2fc62..755b74391e 100644
--- a/be/src/vec/data_types/serde/data_type_array_serde.h
+++ b/be/src/vec/data_types/serde/data_type_array_serde.h
@@ -15,10 +15,20 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#pragma once
+
+#include <glog/logging.h>
+
+#include <ostream>
+
+#include "common/status.h"
 #include "data_type_serde.h"
+
 namespace doris {
+class PValues;
 
 namespace vectorized {
+class IColumn;
 
 class DataTypeArraySerDe : public DataTypeSerDe {
 public:
diff --git a/be/src/vec/data_types/serde/data_type_bitmap_serde.cpp b/be/src/vec/data_types/serde/data_type_bitmap_serde.cpp
index da0f2ed3ea..e09b863240 100644
--- a/be/src/vec/data_types/serde/data_type_bitmap_serde.cpp
+++ b/be/src/vec/data_types/serde/data_type_bitmap_serde.cpp
@@ -17,12 +17,19 @@
 
 #include "data_type_bitmap_serde.h"
 
+#include <gen_cpp/types.pb.h>
+
+#include <string>
+
+#include "util/bitmap_value.h"
 #include "vec/columns/column_complex.h"
-#include "vec/data_types/data_type.h"
+#include "vec/common/assert_cast.h"
 
 namespace doris {
 
 namespace vectorized {
+class IColumn;
+
 Status DataTypeBitMapSerDe::write_column_to_pb(const IColumn& column, PValues& result, int start,
                                                int end) const {
     auto ptype = result.mutable_type();
diff --git a/be/src/vec/data_types/serde/data_type_bitmap_serde.h b/be/src/vec/data_types/serde/data_type_bitmap_serde.h
index d37575a16c..f59519b7f9 100644
--- a/be/src/vec/data_types/serde/data_type_bitmap_serde.h
+++ b/be/src/vec/data_types/serde/data_type_bitmap_serde.h
@@ -15,10 +15,16 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#pragma once
+
+#include "common/status.h"
 #include "data_type_serde.h"
+
 namespace doris {
+class PValues;
 
 namespace vectorized {
+class IColumn;
 
 class DataTypeBitMapSerDe : public DataTypeSerDe {
 public:
diff --git a/be/src/vec/data_types/serde/data_type_decimal_serde.h b/be/src/vec/data_types/serde/data_type_decimal_serde.h
index 4cd5ee484f..a269546eb8 100644
--- a/be/src/vec/data_types/serde/data_type_decimal_serde.h
+++ b/be/src/vec/data_types/serde/data_type_decimal_serde.h
@@ -15,12 +15,23 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#pragma once
+
+#include <gen_cpp/types.pb.h>
+#include <stddef.h>
+
+#include "common/status.h"
 #include "data_type_serde.h"
-#include "vec/columns/column_decimal.h"
+#include "olap/olap_common.h"
+#include "vec/columns/column.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/types.h"
 
 namespace doris {
 
 namespace vectorized {
+template <typename T>
+class ColumnDecimal;
 
 template <typename T>
 class DataTypeDecimalSerDe : public DataTypeSerDe {
diff --git a/be/src/vec/data_types/serde/data_type_fixedlengthobject_serde.h b/be/src/vec/data_types/serde/data_type_fixedlengthobject_serde.h
index 61244184ca..563ef98523 100644
--- a/be/src/vec/data_types/serde/data_type_fixedlengthobject_serde.h
+++ b/be/src/vec/data_types/serde/data_type_fixedlengthobject_serde.h
@@ -15,10 +15,20 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#pragma once
+
+#include <glog/logging.h>
+
+#include <ostream>
+
+#include "common/status.h"
 #include "data_type_serde.h"
+
 namespace doris {
+class PValues;
 
 namespace vectorized {
+class IColumn;
 
 class DataTypeFixedLengthObjectSerDe : public DataTypeSerDe {
 public:
diff --git a/be/src/vec/data_types/serde/data_type_hll_serde.cpp b/be/src/vec/data_types/serde/data_type_hll_serde.cpp
index 880818b830..96dbf23b25 100644
--- a/be/src/vec/data_types/serde/data_type_hll_serde.cpp
+++ b/be/src/vec/data_types/serde/data_type_hll_serde.cpp
@@ -17,11 +17,21 @@
 
 #include "data_type_hll_serde.h"
 
+#include <gen_cpp/types.pb.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <string>
+
+#include "olap/hll.h"
+#include "util/slice.h"
 #include "vec/columns/column_complex.h"
+#include "vec/common/assert_cast.h"
 
 namespace doris {
 
 namespace vectorized {
+class IColumn;
 
 Status DataTypeHLLSerDe::write_column_to_pb(const IColumn& column, PValues& result, int start,
                                             int end) const {
diff --git a/be/src/vec/data_types/serde/data_type_hll_serde.h b/be/src/vec/data_types/serde/data_type_hll_serde.h
index 6a84e394a2..538c35ed0a 100644
--- a/be/src/vec/data_types/serde/data_type_hll_serde.h
+++ b/be/src/vec/data_types/serde/data_type_hll_serde.h
@@ -15,10 +15,16 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#pragma once
+
+#include "common/status.h"
 #include "data_type_serde.h"
+
 namespace doris {
+class PValues;
 
 namespace vectorized {
+class IColumn;
 
 class DataTypeHLLSerDe : public DataTypeSerDe {
 public:
diff --git a/be/src/vec/data_types/serde/data_type_map_serde.h b/be/src/vec/data_types/serde/data_type_map_serde.h
index 2b105fec61..ac72465dc1 100644
--- a/be/src/vec/data_types/serde/data_type_map_serde.h
+++ b/be/src/vec/data_types/serde/data_type_map_serde.h
@@ -15,10 +15,20 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#pragma once
+
+#include <glog/logging.h>
+
+#include <ostream>
+
+#include "common/status.h"
 #include "data_type_serde.h"
+
 namespace doris {
+class PValues;
 
 namespace vectorized {
+class IColumn;
 
 class DataTypeMapSerDe : public DataTypeSerDe {
 public:
diff --git a/be/src/vec/data_types/serde/data_type_nullable_serde.cpp b/be/src/vec/data_types/serde/data_type_nullable_serde.cpp
index e43cf818f2..aefb4f7575 100644
--- a/be/src/vec/data_types/serde/data_type_nullable_serde.cpp
+++ b/be/src/vec/data_types/serde/data_type_nullable_serde.cpp
@@ -17,7 +17,18 @@
 
 #include "data_type_nullable_serde.h"
 
+#include <gen_cpp/types.pb.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+
+#include "vec/columns/column.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/data_types/serde/data_type_serde.h"
 
 namespace doris {
 
diff --git a/be/src/vec/data_types/serde/data_type_nullable_serde.h b/be/src/vec/data_types/serde/data_type_nullable_serde.h
index 0c9666fbf8..14aeabeb8a 100644
--- a/be/src/vec/data_types/serde/data_type_nullable_serde.h
+++ b/be/src/vec/data_types/serde/data_type_nullable_serde.h
@@ -15,10 +15,16 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#pragma once
+
+#include "common/status.h"
 #include "data_type_serde.h"
+
 namespace doris {
+class PValues;
 
 namespace vectorized {
+class IColumn;
 
 class DataTypeNullableSerDe : public DataTypeSerDe {
 public:
diff --git a/be/src/vec/data_types/serde/data_type_number_serde.h b/be/src/vec/data_types/serde/data_type_number_serde.h
index c5d63fa1c5..bd76fd0de8 100644
--- a/be/src/vec/data_types/serde/data_type_number_serde.h
+++ b/be/src/vec/data_types/serde/data_type_number_serde.h
@@ -15,8 +15,18 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#pragma once
+
+#include <gen_cpp/types.pb.h>
+#include <stddef.h>
+
+#include "common/status.h"
 #include "data_type_serde.h"
+#include "olap/olap_common.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_vector.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/types.h"
 
 namespace doris {
 
@@ -158,4 +168,4 @@ Status DataTypeNumberSerDe<T>::write_column_to_pb(const IColumn& column, PValues
 }
 
 } // namespace vectorized
-} // namespace doris
\ No newline at end of file
+} // namespace doris
diff --git a/be/src/vec/data_types/serde/data_type_object_serde.h b/be/src/vec/data_types/serde/data_type_object_serde.h
index bf2f77712a..579176f405 100644
--- a/be/src/vec/data_types/serde/data_type_object_serde.h
+++ b/be/src/vec/data_types/serde/data_type_object_serde.h
@@ -15,10 +15,20 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#pragma once
+
+#include <glog/logging.h>
+
+#include <ostream>
+
+#include "common/status.h"
 #include "data_type_serde.h"
+
 namespace doris {
+class PValues;
 
 namespace vectorized {
+class IColumn;
 
 class DataTypeObjectSerDe : public DataTypeSerDe {
 public:
diff --git a/be/src/vec/data_types/serde/data_type_quantilestate_serde.h b/be/src/vec/data_types/serde/data_type_quantilestate_serde.h
index 8aa70a142a..9dc41652e4 100644
--- a/be/src/vec/data_types/serde/data_type_quantilestate_serde.h
+++ b/be/src/vec/data_types/serde/data_type_quantilestate_serde.h
@@ -15,7 +15,16 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#pragma once
+
+#include <gen_cpp/types.pb.h>
+#include <stddef.h>
+
+#include "common/status.h"
 #include "data_type_serde.h"
+#include "vec/columns/column.h"
+#include "vec/common/string_ref.h"
+
 namespace doris {
 
 namespace vectorized {
diff --git a/be/src/vec/data_types/serde/data_type_serde.h b/be/src/vec/data_types/serde/data_type_serde.h
index a9ade5befa..0dc101f9a5 100644
--- a/be/src/vec/data_types/serde/data_type_serde.h
+++ b/be/src/vec/data_types/serde/data_type_serde.h
@@ -16,15 +16,17 @@
 // under the License.
 
 #pragma once
+
 #include <memory>
-#include <string>
+#include <vector>
 
 #include "common/status.h"
-#include "vec/columns/column.h"
 
 namespace doris {
+class PValues;
 
 namespace vectorized {
+class IColumn;
 // Deserialize means read from different file format or memory format,
 // for example read from arrow, read from parquet.
 // Serialize means write the column cell or the total column into another
@@ -65,4 +67,4 @@ using DataTypeSerDeSPtr = std::shared_ptr<DataTypeSerDe>;
 using DataTypeSerDeSPtrs = std::vector<DataTypeSerDeSPtr>;
 
 } // namespace vectorized
-} // namespace doris
\ No newline at end of file
+} // namespace doris
diff --git a/be/src/vec/data_types/serde/data_type_string_serde.cpp b/be/src/vec/data_types/serde/data_type_string_serde.cpp
index 83e96e1bed..112fcc6bed 100644
--- a/be/src/vec/data_types/serde/data_type_string_serde.cpp
+++ b/be/src/vec/data_types/serde/data_type_string_serde.cpp
@@ -17,7 +17,12 @@
 
 #include "data_type_string_serde.h"
 
+#include <gen_cpp/types.pb.h>
+#include <stddef.h>
+
+#include "vec/columns/column.h"
 #include "vec/columns/column_string.h"
+#include "vec/common/string_ref.h"
 
 namespace doris {
 namespace vectorized {
diff --git a/be/src/vec/data_types/serde/data_type_string_serde.h b/be/src/vec/data_types/serde/data_type_string_serde.h
index 1ddb1c113c..4ede2ef761 100644
--- a/be/src/vec/data_types/serde/data_type_string_serde.h
+++ b/be/src/vec/data_types/serde/data_type_string_serde.h
@@ -15,10 +15,16 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#pragma once
+
+#include "common/status.h"
 #include "data_type_serde.h"
+
 namespace doris {
+class PValues;
 
 namespace vectorized {
+class IColumn;
 
 class DataTypeStringSerDe : public DataTypeSerDe {
 public:
diff --git a/be/src/vec/data_types/serde/data_type_struct_serde.h b/be/src/vec/data_types/serde/data_type_struct_serde.h
index c2907ae583..e895513bd9 100644
--- a/be/src/vec/data_types/serde/data_type_struct_serde.h
+++ b/be/src/vec/data_types/serde/data_type_struct_serde.h
@@ -15,10 +15,20 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#pragma once
+
+#include <glog/logging.h>
+
+#include <ostream>
+
+#include "common/status.h"
 #include "data_type_serde.h"
+
 namespace doris {
+class PValues;
 
 namespace vectorized {
+class IColumn;
 
 class DataTypeStructSerDe : public DataTypeSerDe {
 public:
diff --git a/be/src/vec/exec/data_gen_functions/vdata_gen_function_inf.h b/be/src/vec/exec/data_gen_functions/vdata_gen_function_inf.h
index 86e453dbb5..c8fc78180e 100644
--- a/be/src/vec/exec/data_gen_functions/vdata_gen_function_inf.h
+++ b/be/src/vec/exec/data_gen_functions/vdata_gen_function_inf.h
@@ -27,6 +27,7 @@ namespace doris {
 
 class RuntimeState;
 class Status;
+class TScanRangeParams;
 
 namespace vectorized {
 
diff --git a/be/src/vec/exec/data_gen_functions/vnumbers_tvf.cpp b/be/src/vec/exec/data_gen_functions/vnumbers_tvf.cpp
index b6324851cd..278e974332 100644
--- a/be/src/vec/exec/data_gen_functions/vnumbers_tvf.cpp
+++ b/be/src/vec/exec/data_gen_functions/vnumbers_tvf.cpp
@@ -17,13 +17,19 @@
 
 #include "vec/exec/data_gen_functions/vnumbers_tvf.h"
 
-#include <sstream>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <glog/logging.h>
 
-#include "exec/exec_node.h"
-#include "gen_cpp/PlanNodes_types.h"
+#include <utility>
+
+#include "common/status.h"
+#include "runtime/descriptors.h"
 #include "runtime/runtime_state.h"
-#include "util/runtime_profile.h"
-#include "vec/common/string_ref.h"
+#include "vec/columns/column.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/data_types/data_type.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/data_gen_functions/vnumbers_tvf.h b/be/src/vec/exec/data_gen_functions/vnumbers_tvf.h
index 382263f33f..fde777fce9 100644
--- a/be/src/vec/exec/data_gen_functions/vnumbers_tvf.h
+++ b/be/src/vec/exec/data_gen_functions/vnumbers_tvf.h
@@ -17,19 +17,22 @@
 
 #pragma once
 
-#include <memory>
+#include <stdint.h>
 
-#include "runtime/descriptors.h"
+#include <vector>
+
+#include "common/global_types.h"
 #include "vec/exec/data_gen_functions/vdata_gen_function_inf.h"
 
 namespace doris {
 
-class TextConverter;
 class TupleDescriptor;
 class RuntimeState;
 class Status;
+class TScanRangeParams;
 
 namespace vectorized {
+class Block;
 
 class VNumbersTVF : public VDataGenFunctionInf {
 public:
diff --git a/be/src/vec/exec/format/csv/csv_reader.cpp b/be/src/vec/exec/format/csv/csv_reader.cpp
index ee15846cde..b292ce98ee 100644
--- a/be/src/vec/exec/format/csv/csv_reader.cpp
+++ b/be/src/vec/exec/format/csv/csv_reader.cpp
@@ -17,27 +17,55 @@
 
 #include "csv_reader.h"
 
+#include <fmt/format.h>
 #include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Types_types.h>
 #include <gen_cpp/internal_service.pb.h>
+#include <glog/logging.h>
 
+#include <algorithm>
+#include <map>
+#include <new>
+#include <ostream>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/consts.h"
 #include "common/status.h"
 #include "exec/decompressor.h"
+#include "exec/line_reader.h"
 #include "exec/text_converter.h"
 #include "exec/text_converter.hpp"
 #include "io/file_factory.h"
 #include "io/fs/broker_file_reader.h"
 #include "io/fs/buffered_reader.h"
+#include "io/fs/file_reader.h"
 #include "io/fs/s3_file_reader.h"
-#include "olap/iterators.h"
-#include "olap/olap_common.h"
+#include "runtime/descriptors.h"
+#include "runtime/runtime_state.h"
+#include "runtime/types.h"
 #include "util/string_util.h"
 #include "util/utf8_check.h"
+#include "vec/common/typeid_cast.h"
 #include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/exec/format/file_reader/new_plain_binary_line_reader.h"
 #include "vec/exec/format/file_reader/new_plain_text_line_reader.h"
 #include "vec/exec/scan/vscanner.h"
 
+namespace doris {
+class RuntimeProfile;
+namespace vectorized {
+class IColumn;
+} // namespace vectorized
+
+namespace io {
+class IOContext;
+enum class FileCachePolicy : uint8_t;
+} // namespace io
+} // namespace doris
+
 namespace doris::vectorized {
 
 const static Slice _s_null_slice = Slice("\\N");
diff --git a/be/src/vec/exec/format/csv/csv_reader.h b/be/src/vec/exec/format/csv/csv_reader.h
index f2c2009777..dd41f17d4b 100644
--- a/be/src/vec/exec/format/csv/csv_reader.h
+++ b/be/src/vec/exec/format/csv/csv_reader.h
@@ -17,9 +17,21 @@
 
 #pragma once
 
+#include <gen_cpp/PlanNodes_types.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+#include "common/status.h"
 #include "io/file_factory.h"
-#include "io/fs/file_reader.h"
-#include "olap/iterators.h"
+#include "io/fs/file_reader_writer_fwd.h"
+#include "util/slice.h"
+#include "vec/data_types/data_type.h"
 #include "vec/exec/format/generic_reader.h"
 
 namespace doris {
@@ -28,10 +40,20 @@ class LineReader;
 class TextConverter;
 class Decompressor;
 class SlotDescriptor;
+class RuntimeProfile;
+class RuntimeState;
+
+namespace io {
+class FileSystem;
+class IOContext;
+} // namespace io
+struct TypeDescriptor;
 
 namespace vectorized {
 
 struct ScannerCounter;
+class Block;
+
 class CsvReader : public GenericReader {
 public:
     CsvReader(RuntimeState* state, RuntimeProfile* profile, ScannerCounter* counter,
diff --git a/be/src/vec/exec/format/file_reader/new_plain_binary_line_reader.cpp b/be/src/vec/exec/format/file_reader/new_plain_binary_line_reader.cpp
index 0e6c34a368..0a1e45c99d 100644
--- a/be/src/vec/exec/format/file_reader/new_plain_binary_line_reader.cpp
+++ b/be/src/vec/exec/format/file_reader/new_plain_binary_line_reader.cpp
@@ -17,12 +17,15 @@
 
 #include "new_plain_binary_line_reader.h"
 
-#include <gen_cpp/Types_types.h>
+#include <gen_cpp/internal_service.pb.h>
 
 #include "io/fs/file_reader.h"
 #include "io/fs/stream_load_pipe.h"
 
 namespace doris {
+namespace io {
+class IOContext;
+} // namespace io
 
 NewPlainBinaryLineReader::NewPlainBinaryLineReader(io::FileReaderSPtr file_reader)
         : _file_reader(file_reader) {}
diff --git a/be/src/vec/exec/format/file_reader/new_plain_binary_line_reader.h b/be/src/vec/exec/format/file_reader/new_plain_binary_line_reader.h
index a96af1c158..96ea001bea 100644
--- a/be/src/vec/exec/format/file_reader/new_plain_binary_line_reader.h
+++ b/be/src/vec/exec/format/file_reader/new_plain_binary_line_reader.h
@@ -17,13 +17,21 @@
 
 #pragma once
 
-#include <gen_cpp/Types_types.h>
-#include <gen_cpp/internal_service.pb.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <memory>
+
+#include "common/status.h"
 #include "exec/line_reader.h"
-#include "io/fs/file_reader.h"
+#include "io/fs/file_reader_writer_fwd.h"
 
 namespace doris {
+class PDataRow;
+
+namespace io {
+class IOContext;
+} // namespace io
 
 // only used for FORMAT_PROTO type, which used for insert
 // transaction(begin/insert into/commit)
diff --git a/be/src/vec/exec/format/file_reader/new_plain_text_line_reader.cpp b/be/src/vec/exec/format/file_reader/new_plain_text_line_reader.cpp
index 3984e68f33..b8fd10c021 100644
--- a/be/src/vec/exec/format/file_reader/new_plain_text_line_reader.cpp
+++ b/be/src/vec/exec/format/file_reader/new_plain_text_line_reader.cpp
@@ -17,10 +17,18 @@
 
 #include "vec/exec/format/file_reader/new_plain_text_line_reader.h"
 
+#include <gen_cpp/Metrics_types.h>
+#include <glog/logging.h>
+#include <string.h>
+
+#include <algorithm>
+#include <memory>
+#include <ostream>
+
 #include "common/status.h"
 #include "exec/decompressor.h"
 #include "io/fs/file_reader.h"
-#include "olap/iterators.h"
+#include "util/slice.h"
 
 // INPUT_CHUNK must
 //  larger than 15B for correct lz4 file decompressing
diff --git a/be/src/vec/exec/format/file_reader/new_plain_text_line_reader.h b/be/src/vec/exec/format/file_reader/new_plain_text_line_reader.h
index 5d5eb74dfb..ec6ab1310d 100644
--- a/be/src/vec/exec/format/file_reader/new_plain_text_line_reader.h
+++ b/be/src/vec/exec/format/file_reader/new_plain_text_line_reader.h
@@ -17,8 +17,13 @@
 
 #pragma once
 
+#include <stddef.h>
+#include <stdint.h>
+
+#include <string>
+
 #include "exec/line_reader.h"
-#include "io/fs/file_reader.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "util/runtime_profile.h"
 
 namespace doris {
diff --git a/be/src/vec/exec/format/json/new_json_reader.cpp b/be/src/vec/exec/format/json/new_json_reader.cpp
index a5241aef3d..41bede7c4f 100644
--- a/be/src/vec/exec/format/json/new_json_reader.cpp
+++ b/be/src/vec/exec/format/json/new_json_reader.cpp
@@ -17,31 +17,66 @@
 
 #include "vec/exec/format/json/new_json_reader.h"
 
-#include <simdjson/error.h>
+#include <fmt/format.h>
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <inttypes.h>
+#include <rapidjson/error/en.h>
+#include <rapidjson/reader.h>
+#include <rapidjson/stringbuffer.h>
+#include <rapidjson/writer.h>
+#include <simdjson/simdjson.h> // IWYU pragma: keep
+#include <stdio.h>
+#include <string.h>
 
-#include "common/compiler_util.h"
+#include <algorithm>
+#include <map>
+#include <ostream>
+#include <string_view>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "exprs/json_functions.h"
 #include "io/file_factory.h"
 #include "io/fs/broker_file_reader.h"
 #include "io/fs/buffered_reader.h"
 #include "io/fs/s3_file_reader.h"
 #include "io/fs/stream_load_pipe.h"
-#include "olap/iterators.h"
 #include "runtime/descriptors.h"
 #include "runtime/runtime_state.h"
 #include "util/defer_op.h"
-#include "util/string_util.h"
 #include "vec/core/block.h"
 #include "vec/exec/format/file_reader/new_plain_text_line_reader.h"
 #include "vec/exec/scan/vscanner.h"
 #include "vec/json/simd_json_parser.h"
-
 // dynamic table
+#include "common/config.h"
+#include "io/fs/file_reader.h"
+#include "runtime/define_primitive_type.h"
+#include "runtime/types.h"
+#include "util/slice.h"
+#include "util/uid_util.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
 #include "vec/columns/column_object.h"
+#include "vec/columns/column_string.h"
+#include "vec/common/assert_cast.h"
 #include "vec/common/schema_util.h"
+#include "vec/common/typeid_cast.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/json/json_parser.h"
 #include "vec/json/parse2column.h"
 
+namespace doris {
+namespace io {
+class IOContext;
+enum class FileCachePolicy : uint8_t;
+} // namespace io
+} // namespace doris
+
 namespace doris::vectorized {
 using namespace ErrorCode;
 
diff --git a/be/src/vec/exec/format/json/new_json_reader.h b/be/src/vec/exec/format/json/new_json_reader.h
index 5d9b35a8c2..170c1248b1 100644
--- a/be/src/vec/exec/format/json/new_json_reader.h
+++ b/be/src/vec/exec/format/json/new_json_reader.h
@@ -17,30 +17,60 @@
 
 #pragma once
 
+#include <rapidjson/allocators.h>
 #include <rapidjson/document.h>
-#include <rapidjson/error/en.h>
-#include <rapidjson/stringbuffer.h>
-#include <rapidjson/writer.h>
+#include <rapidjson/encodings.h>
+#include <rapidjson/rapidjson.h>
+#include <simdjson/common_defs.h>
+#include <simdjson/simdjson.h> // IWYU pragma: keep
+#include <stddef.h>
+#include <stdint.h>
 
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+#include "common/status.h"
+#include "exec/line_reader.h"
+#include "exprs/json_functions.h"
 #include "io/file_factory.h"
-#include "io/fs/file_reader.h"
-#include "olap/iterators.h"
+#include "io/fs/file_reader_writer_fwd.h"
+#include "util/runtime_profile.h"
 #include "vec/common/hash_table/hash_map.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/types.h"
 #include "vec/exec/format/generic_reader.h"
+#include "vec/json/json_parser.h"
+#include "vec/json/simd_json_parser.h"
+
+namespace simdjson {
+namespace fallback {
+namespace ondemand {
+class object;
+} // namespace ondemand
+} // namespace fallback
+} // namespace simdjson
 
 namespace doris {
 
-class FileReader;
-struct JsonPath;
-class LineReader;
 class SlotDescriptor;
+class RuntimeState;
+class TFileRangeDesc;
+class TFileScanRangeParams;
+
+namespace io {
+class FileSystem;
+class IOContext;
+} // namespace io
+struct TypeDescriptor;
 
 namespace vectorized {
 
 struct ScannerCounter;
-template <typename ParserImpl>
-class JSONDataParser;
-class SimdJSONParser;
+class Block;
+class IColumn;
 
 class NewJsonReader : public GenericReader {
 public:
diff --git a/be/src/vec/exec/format/orc/vorc_reader.cpp b/be/src/vec/exec/format/orc/vorc_reader.cpp
index ada1284f21..6644121b00 100644
--- a/be/src/vec/exec/format/orc/vorc_reader.cpp
+++ b/be/src/vec/exec/format/orc/vorc_reader.cpp
@@ -17,21 +17,64 @@
 
 #include "vorc_reader.h"
 
+#include <cctz/civil_time_detail.h>
+#include <ctype.h>
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+
+#include <algorithm>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <exception>
+#include <iterator>
+#include <map>
+#include <ostream>
 #include <tuple>
+#include <variant>
 
 #include "cctz/civil_time.h"
 #include "cctz/time_zone.h"
+#include "common/exception.h"
+#include "exec/olap_utils.h"
+#include "gutil/casts.h"
 #include "gutil/strings/substitute.h"
 #include "io/fs/file_reader.h"
-#include "olap/iterators.h"
+#include "orc/Exceptions.hh"
+#include "orc/Int128.hh"
+#include "orc/MemoryPool.hh"
+#include "orc/OrcFile.hh"
+#include "orc/sargs/Literal.hh"
+#include "orc/sargs/SearchArgument.hh"
+#include "runtime/decimalv2_value.h"
+#include "runtime/define_primitive_type.h"
+#include "runtime/primitive_type.h"
 #include "util/slice.h"
+#include "util/timezone_utils.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
 #include "vec/columns/column_map.h"
+#include "vec/columns/column_nullable.h"
 #include "vec/columns/column_struct.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_array.h"
 #include "vec/data_types/data_type_map.h"
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_struct.h"
+#include "vec/runtime/vdatetime_value.h"
+
+namespace doris {
+class RuntimeState;
+
+namespace io {
+class IOContext;
+enum class FileCachePolicy : uint8_t;
+} // namespace io
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/format/orc/vorc_reader.h b/be/src/vec/exec/format/orc/vorc_reader.h
index 6e520dce15..b2d9e22cbf 100644
--- a/be/src/vec/exec/format/orc/vorc_reader.h
+++ b/be/src/vec/exec/format/orc/vorc_reader.h
@@ -17,21 +17,70 @@
 
 #pragma once
 
+#include <cctz/time_zone.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <list>
+#include <memory>
 #include <orc/OrcFile.hh>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
+#include <vector>
 
 #include "common/config.h"
+#include "common/status.h"
 #include "exec/olap_common.h"
 #include "io/file_factory.h"
 #include "io/fs/file_reader.h"
+#include "io/fs/file_reader_writer_fwd.h"
+#include "olap/olap_common.h"
+#include "orc/Reader.hh"
+#include "orc/Type.hh"
+#include "orc/Vector.hh"
+#include "runtime/types.h"
+#include "util/runtime_profile.h"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column_array.h"
-#include "vec/core/block.h"
-#include "vec/data_types/data_type_decimal.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/exec/format/format_common.h"
 #include "vec/exec/format/generic_reader.h"
 
+namespace doris {
+class RuntimeState;
+class TFileRangeDesc;
+class TFileScanRangeParams;
+
+namespace io {
+class FileSystem;
+class IOContext;
+} // namespace io
+namespace vectorized {
+class Block;
+class VecDateTimeValue;
+struct DateTimeV2ValueType;
+template <typename T>
+class ColumnVector;
+template <typename T>
+class DataTypeDecimal;
+template <typename T>
+class DateV2Value;
+template <typename T>
+struct Decimal;
+} // namespace vectorized
+} // namespace doris
+namespace orc {
+template <class T>
+class DataBuffer;
+} // namespace orc
+
 namespace doris::vectorized {
 
 class ORCFileInputStream;
+
 class OrcReader : public GenericReader {
 public:
     struct Statistics {
diff --git a/be/src/vec/exec/format/parquet/bool_plain_decoder.cpp b/be/src/vec/exec/format/parquet/bool_plain_decoder.cpp
index ff05c559a5..cc41ad05ef 100644
--- a/be/src/vec/exec/format/parquet/bool_plain_decoder.cpp
+++ b/be/src/vec/exec/format/parquet/bool_plain_decoder.cpp
@@ -17,6 +17,15 @@
 
 #include "vec/exec/format/parquet/bool_plain_decoder.h"
 
+#include <glog/logging.h>
+
+#include <algorithm>
+
+#include "util/bit_util.h"
+#include "vec/columns/column_vector.h"
+#include "vec/core/types.h"
+#include "vec/exec/format/parquet/parquet_common.h"
+
 namespace doris::vectorized {
 Status BoolPlainDecoder::skip_values(size_t num_values) {
     int skip_cached = std::min(num_unpacked_values_ - unpacked_value_idx_, (int)num_values);
diff --git a/be/src/vec/exec/format/parquet/bool_plain_decoder.h b/be/src/vec/exec/format/parquet/bool_plain_decoder.h
index 08e54ecd60..6b6543ef4d 100644
--- a/be/src/vec/exec/format/parquet/bool_plain_decoder.h
+++ b/be/src/vec/exec/format/parquet/bool_plain_decoder.h
@@ -17,9 +17,24 @@
 
 #pragma once
 
+#include <stddef.h>
+#include <stdint.h>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/status.h"
+#include "util/bit_stream_utils.h"
 #include "util/bit_stream_utils.inline.h"
+#include "util/slice.h"
+#include "vec/data_types/data_type.h"
 #include "vec/exec/format/parquet/decoder.h"
 
+namespace doris {
+namespace vectorized {
+class ColumnSelectVector;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 /// Decoder bit-packed boolean-encoded values.
 /// Implementation from https://github.com/apache/impala/blob/master/be/src/exec/parquet/parquet-bool-decoder.h
diff --git a/be/src/vec/exec/format/parquet/bool_rle_decoder.cpp b/be/src/vec/exec/format/parquet/bool_rle_decoder.cpp
index 46d403e6f7..7e52a9ef62 100644
--- a/be/src/vec/exec/format/parquet/bool_rle_decoder.cpp
+++ b/be/src/vec/exec/format/parquet/bool_rle_decoder.cpp
@@ -17,7 +17,17 @@
 
 #include "vec/exec/format/parquet/bool_rle_decoder.h"
 
-#include "util/bit_util.h"
+#include <glog/logging.h>
+
+#include <algorithm>
+#include <ostream>
+#include <string>
+
+#include "util/coding.h"
+#include "util/slice.h"
+#include "vec/columns/column_vector.h"
+#include "vec/core/types.h"
+#include "vec/exec/format/parquet/parquet_common.h"
 
 namespace doris::vectorized {
 void BoolRLEDecoder::set_data(Slice* slice) {
diff --git a/be/src/vec/exec/format/parquet/bool_rle_decoder.h b/be/src/vec/exec/format/parquet/bool_rle_decoder.h
index 6b0c94825f..18180d1481 100644
--- a/be/src/vec/exec/format/parquet/bool_rle_decoder.h
+++ b/be/src/vec/exec/format/parquet/bool_rle_decoder.h
@@ -17,10 +17,23 @@
 
 #pragma once
 
+#include <stddef.h>
+#include <stdint.h>
+
+#include <vector>
+
+#include "common/status.h"
 #include "util/rle_encoding.h"
-#include "vec/exec/format/parquet/bool_plain_decoder.h"
+#include "vec/data_types/data_type.h"
 #include "vec/exec/format/parquet/decoder.h"
 
+namespace doris {
+namespace vectorized {
+class ColumnSelectVector;
+} // namespace vectorized
+struct Slice;
+} // namespace doris
+
 namespace doris::vectorized {
 class BoolRLEDecoder final : public Decoder {
 public:
diff --git a/be/src/vec/exec/format/parquet/byte_array_dict_decoder.cpp b/be/src/vec/exec/format/parquet/byte_array_dict_decoder.cpp
index 6962b2a47b..40aaf8c66b 100644
--- a/be/src/vec/exec/format/parquet/byte_array_dict_decoder.cpp
+++ b/be/src/vec/exec/format/parquet/byte_array_dict_decoder.cpp
@@ -17,8 +17,14 @@
 
 #include "vec/exec/format/parquet/byte_array_dict_decoder.h"
 
+#include <utility>
+
 #include "util/coding.h"
+#include "util/rle_encoding.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_dictionary.h"
+#include "vec/columns/column_string.h"
+#include "vec/common/assert_cast.h"
 #include "vec/data_types/data_type_nullable.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/exec/format/parquet/byte_array_dict_decoder.h b/be/src/vec/exec/format/parquet/byte_array_dict_decoder.h
index 8492937e83..474df7e45d 100644
--- a/be/src/vec/exec/format/parquet/byte_array_dict_decoder.h
+++ b/be/src/vec/exec/format/parquet/byte_array_dict_decoder.h
@@ -17,11 +17,30 @@
 
 #pragma once
 
-#include <cstdint>
+#include <string.h>
 
+#include <cstdint>
+#include <memory>
+#include <unordered_map>
+#include <vector>
+
+#include "common/status.h"
 #include "gutil/endian.h"
-#include "util/rle_encoding.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/exec/format/format_common.h"
 #include "vec/exec/format/parquet/decoder.h"
+#include "vec/exec/format/parquet/parquet_common.h"
+
+namespace doris {
+namespace vectorized {
+class ColumnString;
+template <typename T>
+class ColumnDecimal;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/format/parquet/byte_array_plain_decoder.cpp b/be/src/vec/exec/format/parquet/byte_array_plain_decoder.cpp
index 833472665b..5fad50ac01 100644
--- a/be/src/vec/exec/format/parquet/byte_array_plain_decoder.cpp
+++ b/be/src/vec/exec/format/parquet/byte_array_plain_decoder.cpp
@@ -17,6 +17,11 @@
 
 #include "vec/exec/format/parquet/byte_array_plain_decoder.h"
 
+#include <algorithm>
+#include <vector>
+
+#include "vec/columns/column.h"
+#include "vec/common/string_ref.h"
 #include "vec/data_types/data_type_nullable.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/exec/format/parquet/byte_array_plain_decoder.h b/be/src/vec/exec/format/parquet/byte_array_plain_decoder.h
index cd35ceaa6f..f741eaa257 100644
--- a/be/src/vec/exec/format/parquet/byte_array_plain_decoder.h
+++ b/be/src/vec/exec/format/parquet/byte_array_plain_decoder.h
@@ -17,10 +17,29 @@
 
 #pragma once
 
-#include <cstdint>
+#include <string.h>
 
+#include <cstdint>
+#include <memory>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/status.h"
+#include "gutil/endian.h"
 #include "util/coding.h"
+#include "util/slice.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/exec/format/format_common.h"
 #include "vec/exec/format/parquet/decoder.h"
+#include "vec/exec/format/parquet/parquet_common.h"
+
+namespace doris {
+namespace vectorized {
+template <typename T>
+class ColumnDecimal;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/format/parquet/decoder.cpp b/be/src/vec/exec/format/parquet/decoder.cpp
index e625994a21..a1fc3706fd 100644
--- a/be/src/vec/exec/format/parquet/decoder.cpp
+++ b/be/src/vec/exec/format/parquet/decoder.cpp
@@ -17,8 +17,9 @@
 
 #include "vec/exec/format/parquet/decoder.h"
 
-#include "vec/data_types/data_type_decimal.h"
-#include "vec/data_types/data_type_nullable.h"
+#include <cctz/time_zone.h>
+#include <gen_cpp/parquet_types.h>
+
 #include "vec/exec/format/parquet/bool_plain_decoder.h"
 #include "vec/exec/format/parquet/bool_rle_decoder.h"
 #include "vec/exec/format/parquet/byte_array_dict_decoder.h"
@@ -26,6 +27,7 @@
 #include "vec/exec/format/parquet/delta_bit_pack_decoder.h"
 #include "vec/exec/format/parquet/fix_length_dict_decoder.hpp"
 #include "vec/exec/format/parquet/fix_length_plain_decoder.h"
+#include "vec/exec/format/parquet/schema_desc.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/format/parquet/decoder.h b/be/src/vec/exec/format/parquet/decoder.h
index eb0d88bd63..f5126cd098 100644
--- a/be/src/vec/exec/format/parquet/decoder.h
+++ b/be/src/vec/exec/format/parquet/decoder.h
@@ -17,17 +17,41 @@
 
 #pragma once
 
+#include <gen_cpp/parquet_types.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
 #include <cstdint>
+#include <memory>
+#include <ostream>
+#include <vector>
 
 #include "common/status.h"
-#include "gen_cpp/parquet_types.h"
 #include "schema_desc.h"
 #include "util/rle_encoding.h"
+#include "util/slice.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_dictionary.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_decimal.h" // IWYU pragma: keep
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/exec/format/format_common.h"
 #include "vec/exec/format/parquet/parquet_common.h"
 
+namespace cctz {
+class time_zone;
+} // namespace cctz
+namespace doris {
+namespace vectorized {
+class ColumnString;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 #define FOR_LOGICAL_NUMERIC_TYPES(M)        \
diff --git a/be/src/vec/exec/format/parquet/delta_bit_pack_decoder.cpp b/be/src/vec/exec/format/parquet/delta_bit_pack_decoder.cpp
index c550e7262d..598c308761 100644
--- a/be/src/vec/exec/format/parquet/delta_bit_pack_decoder.cpp
+++ b/be/src/vec/exec/format/parquet/delta_bit_pack_decoder.cpp
@@ -17,6 +17,17 @@
 
 #include "delta_bit_pack_decoder.h"
 
+#include <string.h>
+
+#include <algorithm>
+#include <string_view>
+
+#include "vec/columns/column.h"
+#include "vec/common/arithmetic_overflow.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type_nullable.h"
+
 namespace doris::vectorized {
 
 Status DeltaDecoder::decode_byte_array(const std::vector<Slice>& decoded_vals,
diff --git a/be/src/vec/exec/format/parquet/delta_bit_pack_decoder.h b/be/src/vec/exec/format/parquet/delta_bit_pack_decoder.h
index f51941bb3a..a1d379d607 100644
--- a/be/src/vec/exec/format/parquet/delta_bit_pack_decoder.h
+++ b/be/src/vec/exec/format/parquet/delta_bit_pack_decoder.h
@@ -17,9 +17,27 @@
 
 #pragma once
 
+#include <gen_cpp/parquet_types.h>
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <memory>
+#include <ostream>
+#include <string>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
+#include "gutil/port.h"
 #include "util/bit_stream_utils.h"
-#include "vec/exec/format/parquet/byte_array_plain_decoder.h"
+#include "util/bit_stream_utils.inline.h"
+#include "util/slice.h"
+#include "vec/data_types/data_type.h"
+#include "vec/exec/format/parquet/decoder.h"
 #include "vec/exec/format/parquet/fix_length_plain_decoder.h"
+#include "vec/exec/format/parquet/parquet_common.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/format/parquet/fix_length_plain_decoder.cpp b/be/src/vec/exec/format/parquet/fix_length_plain_decoder.cpp
index 01bfe67981..f809264c46 100644
--- a/be/src/vec/exec/format/parquet/fix_length_plain_decoder.cpp
+++ b/be/src/vec/exec/format/parquet/fix_length_plain_decoder.cpp
@@ -17,9 +17,33 @@
 
 #include "vec/exec/format/parquet/fix_length_plain_decoder.h"
 
+#include <gen_cpp/parquet_types.h>
+#include <stdint.h>
+#include <string.h>
+
+#include <memory>
+#include <vector>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "gutil/endian.h"
-#include "vec/columns/column_nullable.h"
+#include "util/slice.h"
+#include "vec/columns/column.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_nullable.h"
+#include "vec/exec/format/format_common.h"
+#include "vec/exec/format/parquet/parquet_common.h"
+#include "vec/runtime/vdatetime_value.h"
+
+namespace doris {
+namespace vectorized {
+template <typename T>
+class ColumnDecimal;
+template <typename T>
+class ColumnVector;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/format/parquet/fix_length_plain_decoder.h b/be/src/vec/exec/format/parquet/fix_length_plain_decoder.h
index c35a97fc3e..f9d4b9290e 100644
--- a/be/src/vec/exec/format/parquet/fix_length_plain_decoder.h
+++ b/be/src/vec/exec/format/parquet/fix_length_plain_decoder.h
@@ -17,13 +17,18 @@
 
 #pragma once
 
+#include <gen_cpp/parquet_types.h>
+#include <stddef.h>
+
 #include "common/status.h"
-#include "gen_cpp/parquet_types.h"
-#include "util/slice.h"
-#include "vec/columns/column.h"
 #include "vec/data_types/data_type.h"
 #include "vec/exec/format/parquet/decoder.h"
-#include "vec/exec/format/parquet/parquet_common.h"
+
+namespace doris {
+namespace vectorized {
+class ColumnSelectVector;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/format/parquet/level_decoder.cpp b/be/src/vec/exec/format/parquet/level_decoder.cpp
index 387d032c05..1e7c03f87b 100644
--- a/be/src/vec/exec/format/parquet/level_decoder.cpp
+++ b/be/src/vec/exec/format/parquet/level_decoder.cpp
@@ -17,8 +17,14 @@
 
 #include "level_decoder.h"
 
+#include <gen_cpp/parquet_types.h>
+
+#include <algorithm>
+
+#include "util/bit_stream_utils.inline.h"
 #include "util/bit_util.h"
 #include "util/coding.h"
+#include "vec/exec/format/parquet/parquet_common.h"
 
 static constexpr size_t V1_LEVEL_SIZE = 4;
 
diff --git a/be/src/vec/exec/format/parquet/level_decoder.h b/be/src/vec/exec/format/parquet/level_decoder.h
index 324f914993..4f76ac0683 100644
--- a/be/src/vec/exec/format/parquet/level_decoder.h
+++ b/be/src/vec/exec/format/parquet/level_decoder.h
@@ -17,13 +17,16 @@
 
 #pragma once
 
+#include <gen_cpp/parquet_types.h>
+#include <stddef.h>
+
 #include <cstdint>
 
 #include "common/status.h"
-#include "gen_cpp/parquet_types.h"
 #include "parquet_common.h"
 #include "util/bit_stream_utils.h"
 #include "util/rle_encoding.h"
+#include "util/slice.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/format/parquet/parquet_common.cpp b/be/src/vec/exec/format/parquet/parquet_common.cpp
index 48a0a349eb..d58dbb712f 100644
--- a/be/src/vec/exec/format/parquet/parquet_common.cpp
+++ b/be/src/vec/exec/format/parquet/parquet_common.cpp
@@ -17,8 +17,10 @@
 
 #include "parquet_common.h"
 
-#include "util/coding.h"
-#include "vec/data_types/data_type_nullable.h"
+#include <glog/logging.h>
+
+#include "util/simd/bits.h"
+#include "vec/core/types.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/format/parquet/parquet_common.h b/be/src/vec/exec/format/parquet/parquet_common.h
index eb2e117bd6..eb0c128e26 100644
--- a/be/src/vec/exec/format/parquet/parquet_common.h
+++ b/be/src/vec/exec/format/parquet/parquet_common.h
@@ -17,25 +17,14 @@
 
 #pragma once
 
-#include <cstdint>
+#include <stddef.h>
+
+#include <cstdint>
+#include <ostream>
+#include <string>
+#include <vector>
 
-#include "common/status.h"
-#include "gen_cpp/parquet_types.h"
-#include "gutil/endian.h"
-#include "schema_desc.h"
-#include "util/bit_stream_utils.inline.h"
-#include "util/coding.h"
-#include "util/rle_encoding.h"
-#include "util/simd/bits.h"
-#include "vec/columns/column_array.h"
-#include "vec/columns/column_dictionary.h"
 #include "vec/columns/column_nullable.h"
-#include "vec/columns/column_string.h"
-#include "vec/common/int_exp.h"
-#include "vec/data_types/data_type.h"
-#include "vec/data_types/data_type_decimal.h"
-#include "vec/data_types/data_type_nullable.h"
-#include "vec/exec/format/format_common.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/format/parquet/parquet_pred_cmp.h b/be/src/vec/exec/format/parquet/parquet_pred_cmp.h
index 900465292b..60001f649e 100644
--- a/be/src/vec/exec/format/parquet/parquet_pred_cmp.h
+++ b/be/src/vec/exec/format/parquet/parquet_pred_cmp.h
@@ -25,6 +25,7 @@
 #include "parquet_common.h"
 #include "vec/data_types/data_type_decimal.h"
 #include "vec/exec/format/format_common.h"
+#include "vec/exec/format/parquet/schema_desc.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/format/parquet/parquet_thrift_util.h b/be/src/vec/exec/format/parquet/parquet_thrift_util.h
index 98b02609db..7dca35c850 100644
--- a/be/src/vec/exec/format/parquet/parquet_thrift_util.h
+++ b/be/src/vec/exec/format/parquet/parquet_thrift_util.h
@@ -18,11 +18,11 @@
 #pragma once
 
 #include <common/status.h>
+#include <gen_cpp/parquet_types.h>
 
 #include <cstdint>
 
 #include "common/logging.h"
-#include "gen_cpp/parquet_types.h"
 #include "io/fs/file_reader.h"
 #include "io/io_common.h"
 #include "olap/iterators.h"
diff --git a/be/src/vec/exec/format/parquet/schema_desc.cpp b/be/src/vec/exec/format/parquet/schema_desc.cpp
index ebd9a25352..0cf8bef696 100644
--- a/be/src/vec/exec/format/parquet/schema_desc.cpp
+++ b/be/src/vec/exec/format/parquet/schema_desc.cpp
@@ -17,7 +17,14 @@
 
 #include "schema_desc.h"
 
+#include <ctype.h>
+
+#include <algorithm>
+#include <ostream>
+#include <utility>
+
 #include "common/logging.h"
+#include "runtime/define_primitive_type.h"
 #include "util/slice.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/exec/format/parquet/schema_desc.h b/be/src/vec/exec/format/parquet/schema_desc.h
index 7d32749c68..ad6b8fb293 100644
--- a/be/src/vec/exec/format/parquet/schema_desc.h
+++ b/be/src/vec/exec/format/parquet/schema_desc.h
@@ -18,8 +18,12 @@
 #pragma once
 
 #include <gen_cpp/parquet_types.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <string>
 #include <unordered_map>
+#include <unordered_set>
 #include <vector>
 
 #include "common/status.h"
diff --git a/be/src/vec/exec/format/parquet/vparquet_column_chunk_reader.cpp b/be/src/vec/exec/format/parquet/vparquet_column_chunk_reader.cpp
index d0e5f25d71..225f910487 100644
--- a/be/src/vec/exec/format/parquet/vparquet_column_chunk_reader.cpp
+++ b/be/src/vec/exec/format/parquet/vparquet_column_chunk_reader.cpp
@@ -17,6 +17,33 @@
 
 #include "vparquet_column_chunk_reader.h"
 
+#include <gen_cpp/parquet_types.h>
+#include <glog/logging.h>
+#include <string.h>
+
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "util/bit_util.h"
+#include "util/block_compression.h"
+#include "util/runtime_profile.h"
+#include "vec/columns/column.h"
+#include "vec/exec/format/parquet/decoder.h"
+#include "vec/exec/format/parquet/level_decoder.h"
+#include "vec/exec/format/parquet/schema_desc.h"
+#include "vec/exec/format/parquet/vparquet_page_reader.h"
+
+namespace cctz {
+class time_zone;
+} // namespace cctz
+namespace doris {
+namespace io {
+class BufferedStreamReader;
+class IOContext;
+} // namespace io
+} // namespace doris
+
 namespace doris::vectorized {
 
 ColumnChunkReader::ColumnChunkReader(io::BufferedStreamReader* reader,
diff --git a/be/src/vec/exec/format/parquet/vparquet_column_chunk_reader.h b/be/src/vec/exec/format/parquet/vparquet_column_chunk_reader.h
index ad37e13815..9f62fddce7 100644
--- a/be/src/vec/exec/format/parquet/vparquet_column_chunk_reader.h
+++ b/be/src/vec/exec/format/parquet/vparquet_column_chunk_reader.h
@@ -17,6 +17,9 @@
 
 #pragma once
 
+#include <gen_cpp/parquet_types.h>
+#include <stddef.h>
+
 #include <cstdint>
 #include <memory>
 #include <unordered_map>
@@ -24,13 +27,29 @@
 
 #include "common/status.h"
 #include "decoder.h"
-#include "gen_cpp/parquet_types.h"
-#include "io/fs/buffered_reader.h"
 #include "level_decoder.h"
-#include "schema_desc.h"
-#include "util/block_compression.h"
+#include "util/slice.h"
+#include "vec/columns/columns_number.h"
+#include "vec/data_types/data_type.h"
+#include "vec/exec/format/parquet/parquet_common.h"
 #include "vparquet_page_reader.h"
 
+namespace cctz {
+class time_zone;
+} // namespace cctz
+namespace doris {
+class BlockCompressionCodec;
+
+namespace io {
+class BufferedStreamReader;
+class IOContext;
+} // namespace io
+namespace vectorized {
+class ColumnString;
+struct FieldSchema;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 /**
diff --git a/be/src/vec/exec/format/parquet/vparquet_column_reader.cpp b/be/src/vec/exec/format/parquet/vparquet_column_reader.cpp
index 1657f5ab3a..b9b0a9b536 100644
--- a/be/src/vec/exec/format/parquet/vparquet_column_reader.cpp
+++ b/be/src/vec/exec/format/parquet/vparquet_column_reader.cpp
@@ -19,8 +19,17 @@
 
 #include <common/status.h>
 #include <gen_cpp/parquet_types.h>
+#include <limits.h>
+#include <sys/types.h>
 
+#include <algorithm>
+#include <utility>
+
+#include "runtime/define_primitive_type.h"
+#include "runtime/types.h"
 #include "schema_desc.h"
+#include "util/runtime_profile.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
 #include "vec/columns/column_map.h"
 #include "vec/columns/column_nullable.h"
@@ -29,8 +38,21 @@
 #include "vec/data_types/data_type_map.h"
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_struct.h"
+#include "vec/exec/format/parquet/level_decoder.h"
 #include "vparquet_column_chunk_reader.h"
 
+namespace cctz {
+class time_zone;
+} // namespace cctz
+namespace doris {
+namespace io {
+class IOContext;
+} // namespace io
+namespace vectorized {
+class ColumnString;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 static void fill_struct_null_map(FieldSchema* field, NullMap& null_map,
diff --git a/be/src/vec/exec/format/parquet/vparquet_column_reader.h b/be/src/vec/exec/format/parquet/vparquet_column_reader.h
index 26cdbd9dd3..c9867f4e45 100644
--- a/be/src/vec/exec/format/parquet/vparquet_column_reader.h
+++ b/be/src/vec/exec/format/parquet/vparquet_column_reader.h
@@ -18,10 +18,35 @@
 #pragma once
 #include <common/status.h>
 #include <gen_cpp/parquet_types.h>
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
 
-#include "schema_desc.h"
+#include <list>
+#include <memory>
+#include <ostream>
+#include <vector>
+
+#include "io/fs/buffered_reader.h"
+#include "io/fs/file_reader_writer_fwd.h"
+#include "vec/columns/columns_number.h"
+#include "vec/data_types/data_type.h"
+#include "vec/exec/format/parquet/parquet_common.h"
 #include "vparquet_column_chunk_reader.h"
 
+namespace cctz {
+class time_zone;
+} // namespace cctz
+namespace doris {
+namespace io {
+class IOContext;
+} // namespace io
+namespace vectorized {
+class ColumnString;
+struct FieldSchema;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 class ParquetColumnReader {
diff --git a/be/src/vec/exec/format/parquet/vparquet_file_metadata.cpp b/be/src/vec/exec/format/parquet/vparquet_file_metadata.cpp
index c583e50022..c199e72da7 100644
--- a/be/src/vec/exec/format/parquet/vparquet_file_metadata.cpp
+++ b/be/src/vec/exec/format/parquet/vparquet_file_metadata.cpp
@@ -17,7 +17,10 @@
 
 #include "vparquet_file_metadata.h"
 
+#include <gen_cpp/parquet_types.h>
+
 #include <sstream>
+#include <vector>
 
 #include "schema_desc.h"
 
diff --git a/be/src/vec/exec/format/parquet/vparquet_file_metadata.h b/be/src/vec/exec/format/parquet/vparquet_file_metadata.h
index 19c6ce98f9..6f52ef5b4a 100644
--- a/be/src/vec/exec/format/parquet/vparquet_file_metadata.h
+++ b/be/src/vec/exec/format/parquet/vparquet_file_metadata.h
@@ -16,8 +16,11 @@
 // under the License.
 
 #pragma once
+#include <gen_cpp/parquet_types.h>
+
+#include <string>
+
 #include "common/status.h"
-#include "gen_cpp/parquet_types.h"
 #include "schema_desc.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/exec/format/parquet/vparquet_group_reader.cpp b/be/src/vec/exec/format/parquet/vparquet_group_reader.cpp
index 29c2f488a5..865e6152c2 100644
--- a/be/src/vec/exec/format/parquet/vparquet_group_reader.cpp
+++ b/be/src/vec/exec/format/parquet/vparquet_group_reader.cpp
@@ -17,16 +17,62 @@
 
 #include "vparquet_group_reader.h"
 
+#include <gen_cpp/Exprs_types.h>
+#include <gen_cpp/Opcodes_types.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/parquet_types.h>
+#include <string.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <ostream>
+
+#include "common/config.h"
+#include "common/logging.h"
+#include "common/object_pool.h"
+#include "common/status.h"
 #include "exprs/create_predicate_function.h"
+#include "exprs/hybrid_set.h"
+#include "gutil/stringprintf.h"
+#include "runtime/define_primitive_type.h"
+#include "runtime/descriptors.h"
+#include "runtime/thread_context.h"
+#include "runtime/types.h"
 #include "schema_desc.h"
 #include "util/simd/bits.h"
 #include "vec/columns/column_const.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/pod_array.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/data_types/data_type_number.h"
+#include "vec/data_types/data_type_string.h"
 #include "vec/exprs/vdirect_in_predicate.h"
 #include "vec/exprs/vectorized_fn_call.h"
+#include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
 #include "vec/exprs/vliteral.h"
 #include "vec/exprs/vslot_ref.h"
 #include "vparquet_column_reader.h"
 
+namespace cctz {
+class time_zone;
+} // namespace cctz
+namespace doris {
+class RuntimeState;
+
+namespace io {
+class IOContext;
+} // namespace io
+} // namespace doris
+
 namespace doris::vectorized {
 
 const std::vector<int64_t> RowGroupReader::NO_DELETE = {};
diff --git a/be/src/vec/exec/format/parquet/vparquet_group_reader.h b/be/src/vec/exec/format/parquet/vparquet_group_reader.h
index 33f8371699..910ab5ce1c 100644
--- a/be/src/vec/exec/format/parquet/vparquet_group_reader.h
+++ b/be/src/vec/exec/format/parquet/vparquet_group_reader.h
@@ -16,13 +16,49 @@
 // under the License.
 #pragma once
 #include <common/status.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <limits>
+#include <memory>
+#include <string>
+#include <tuple>
+#include <unordered_map>
+#include <utility>
+#include <vector>
 
 #include "exec/text_converter.h"
-#include "io/fs/file_reader.h"
-#include "vec/core/block.h"
-#include "vec/exprs/vexpr_context.h"
+#include "io/fs/file_reader_writer_fwd.h"
+#include "vec/columns/column.h"
+#include "vec/common/allocator.h"
+#include "vec/exec/format/parquet/parquet_common.h"
 #include "vparquet_column_reader.h"
 
+namespace cctz {
+class time_zone;
+} // namespace cctz
+namespace doris {
+class ObjectPool;
+class RowDescriptor;
+class RuntimeState;
+class SlotDescriptor;
+class TupleDescriptor;
+
+namespace io {
+class IOContext;
+} // namespace io
+namespace vectorized {
+class Block;
+class FieldDescriptor;
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+namespace tparquet {
+class ColumnMetaData;
+class OffsetIndex;
+class RowGroup;
+} // namespace tparquet
+
 namespace doris::vectorized {
 
 // TODO: we need to determine it by test.
@@ -142,8 +178,8 @@ private:
             const std::unordered_map<std::string, VExprContext*>& missing_columns);
     Status _build_pos_delete_filter(size_t read_rows);
     Status _filter_block(Block* block, int column_to_keep,
-                         const vector<uint32_t>& columns_to_filter);
-    Status _filter_block_internal(Block* block, const vector<uint32_t>& columns_to_filter,
+                         const std::vector<uint32_t>& columns_to_filter);
+    Status _filter_block_internal(Block* block, const std::vector<uint32_t>& columns_to_filter,
                                   const IColumn::Filter& filter);
 
     bool _can_filter_by_dict(int slot_id, const tparquet::ColumnMetaData& column_metadata);
diff --git a/be/src/vec/exec/format/parquet/vparquet_page_index.cpp b/be/src/vec/exec/format/parquet/vparquet_page_index.cpp
index ad70d8f8bc..bad05728eb 100644
--- a/be/src/vec/exec/format/parquet/vparquet_page_index.cpp
+++ b/be/src/vec/exec/format/parquet/vparquet_page_index.cpp
@@ -17,8 +17,27 @@
 
 #include "vparquet_page_index.h"
 
+#include <gen_cpp/parquet_types.h>
+
+#include <algorithm>
+#include <limits>
+#include <ostream>
+#include <string>
+
+#include "common/logging.h"
+#include "common/status.h"
 #include "parquet_pred_cmp.h"
 #include "util/thrift_util.h"
+#include "vec/exec/format/parquet/parquet_common.h"
+
+namespace cctz {
+class time_zone;
+} // namespace cctz
+namespace doris {
+namespace vectorized {
+struct FieldSchema;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/format/parquet/vparquet_page_index.h b/be/src/vec/exec/format/parquet/vparquet_page_index.h
index 708f79d3fb..1099870699 100644
--- a/be/src/vec/exec/format/parquet/vparquet_page_index.h
+++ b/be/src/vec/exec/format/parquet/vparquet_page_index.h
@@ -17,11 +17,26 @@
 
 #pragma once
 #include <common/status.h>
-#include <gen_cpp/parquet_types.h>
+#include <stdint.h>
+
+#include <vector>
 
 #include "exec/olap_common.h"
-#include "parquet_common.h"
-#include "schema_desc.h"
+
+namespace cctz {
+class time_zone;
+} // namespace cctz
+namespace doris {
+namespace vectorized {
+struct FieldSchema;
+struct RowRange;
+} // namespace vectorized
+} // namespace doris
+namespace tparquet {
+class ColumnChunk;
+class ColumnIndex;
+class OffsetIndex;
+} // namespace tparquet
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/format/parquet/vparquet_page_reader.cpp b/be/src/vec/exec/format/parquet/vparquet_page_reader.cpp
index 12ef8d0427..e71e3311fd 100644
--- a/be/src/vec/exec/format/parquet/vparquet_page_reader.cpp
+++ b/be/src/vec/exec/format/parquet/vparquet_page_reader.cpp
@@ -17,11 +17,26 @@
 
 #include "vparquet_page_reader.h"
 
+#include <gen_cpp/parquet_types.h>
+#include <stddef.h>
 #include <stdint.h>
 
+#include <algorithm>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/config.h"
+#include "io/fs/buffered_reader.h"
+#include "util/runtime_profile.h"
+#include "util/slice.h"
 #include "util/thrift_util.h"
 
+namespace doris {
+namespace io {
+class IOContext;
+} // namespace io
+} // namespace doris
+
 namespace doris::vectorized {
 
 static constexpr size_t INIT_PAGE_HEADER_SIZE = 128;
diff --git a/be/src/vec/exec/format/parquet/vparquet_page_reader.h b/be/src/vec/exec/format/parquet/vparquet_page_reader.h
index 846ab96f36..089409786f 100644
--- a/be/src/vec/exec/format/parquet/vparquet_page_reader.h
+++ b/be/src/vec/exec/format/parquet/vparquet_page_reader.h
@@ -17,9 +17,18 @@
 
 #pragma once
 
+#include <gen_cpp/parquet_types.h>
+#include <stdint.h>
+
 #include "common/status.h"
-#include "gen_cpp/parquet_types.h"
-#include "io/fs/buffered_reader.h"
+
+namespace doris {
+namespace io {
+class BufferedStreamReader;
+class IOContext;
+} // namespace io
+struct Slice;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/format/parquet/vparquet_reader.cpp b/be/src/vec/exec/format/parquet/vparquet_reader.cpp
index 6a7fb58362..8490d1985e 100644
--- a/be/src/vec/exec/format/parquet/vparquet_reader.cpp
+++ b/be/src/vec/exec/format/parquet/vparquet_reader.cpp
@@ -17,19 +17,54 @@
 
 #include "vparquet_reader.h"
 
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/parquet_types.h>
+#include <glog/logging.h>
+
 #include <algorithm>
+#include <functional>
+#include <ostream>
+#include <utility>
 
 #include "common/status.h"
 #include "io/file_factory.h"
-#include "olap/iterators.h"
 #include "parquet_pred_cmp.h"
 #include "parquet_thrift_util.h"
-#include "rapidjson/document.h"
+#include "runtime/define_primitive_type.h"
+#include "runtime/types.h"
+#include "util/slice.h"
+#include "vec/common/typeid_cast.h"
+#include "vec/exec/format/format_common.h"
+#include "vec/exec/format/parquet/schema_desc.h"
+#include "vec/exec/format/parquet/vparquet_file_metadata.h"
+#include "vec/exec/format/parquet/vparquet_group_reader.h"
+#include "vec/exec/format/parquet/vparquet_page_index.h"
 #include "vec/exprs/vbloom_predicate.h"
+#include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
 #include "vec/exprs/vin_predicate.h"
 #include "vec/exprs/vruntimefilter_wrapper.h"
 #include "vec/exprs/vslot_ref.h"
 
+namespace cctz {
+class time_zone;
+} // namespace cctz
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+class SlotDescriptor;
+class TupleDescriptor;
+namespace io {
+class IOContext;
+enum class FileCachePolicy : uint8_t;
+} // namespace io
+namespace vectorized {
+class Block;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 ParquetReader::ParquetReader(RuntimeProfile* profile, const TFileScanRangeParams& params,
diff --git a/be/src/vec/exec/format/parquet/vparquet_reader.h b/be/src/vec/exec/format/parquet/vparquet_reader.h
index 85f9b5c138..dab7ba4e76 100644
--- a/be/src/vec/exec/format/parquet/vparquet_reader.h
+++ b/be/src/vec/exec/format/parquet/vparquet_reader.h
@@ -17,25 +17,54 @@
 
 #pragma once
 
+#include <gen_cpp/parquet_types.h>
+#include <stddef.h>
 #include <stdint.h>
 
-#include <queue>
+#include <list>
+#include <map>
+#include <memory>
 #include <string>
+#include <tuple>
+#include <unordered_map>
+#include <unordered_set>
 #include <vector>
 
 #include "common/status.h"
 #include "exec/olap_common.h"
-#include "gen_cpp/parquet_types.h"
 #include "io/file_factory.h"
 #include "io/fs/file_reader.h"
-#include "io/fs/file_system.h"
-#include "vec/core/block.h"
+#include "io/fs/file_reader_writer_fwd.h"
+#include "util/runtime_profile.h"
 #include "vec/exec/format/generic_reader.h"
-#include "vec/exprs/vexpr_context.h"
+#include "vec/exec/format/parquet/parquet_common.h"
 #include "vparquet_column_reader.h"
-#include "vparquet_file_metadata.h"
 #include "vparquet_group_reader.h"
-#include "vparquet_page_index.h"
+
+namespace cctz {
+class time_zone;
+} // namespace cctz
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+class SlotDescriptor;
+class TFileRangeDesc;
+class TFileScanRangeParams;
+class TupleDescriptor;
+
+namespace io {
+class FileSystem;
+class IOContext;
+} // namespace io
+namespace vectorized {
+class Block;
+class FileMetaData;
+class PageIndex;
+class ShardedKVCache;
+class VExprContext;
+} // namespace vectorized
+struct TypeDescriptor;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/format/table/iceberg_reader.cpp b/be/src/vec/exec/format/table/iceberg_reader.cpp
index c21630b6c9..010f74fee5 100644
--- a/be/src/vec/exec/format/table/iceberg_reader.cpp
+++ b/be/src/vec/exec/format/table/iceberg_reader.cpp
@@ -17,12 +17,62 @@
 
 #include "iceberg_reader.h"
 
+#include <ctype.h>
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/parquet_types.h>
+#include <glog/logging.h>
+#include <parallel_hashmap/phmap.h>
+#include <rapidjson/allocators.h>
+#include <rapidjson/document.h>
+#include <string.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <functional>
+#include <memory>
+#include <mutex>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/status.h"
-#include "olap/iterators.h"
+#include "olap/olap_common.h"
+#include "runtime/define_primitive_type.h"
+#include "runtime/primitive_type.h"
+#include "runtime/runtime_state.h"
+#include "runtime/types.h"
+#include "util/string_util.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
 #include "vec/core/column_with_type_and_name.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_factory.hpp"
+#include "vec/exec/format/format_common.h"
+#include "vec/exec/format/generic_reader.h"
+#include "vec/exec/format/parquet/parquet_common.h"
 #include "vec/exec/format/parquet/vparquet_reader.h"
+#include "vec/exec/format/table/table_format_reader.h"
+
+namespace cctz {
+class time_zone;
+} // namespace cctz
+namespace doris {
+class RowDescriptor;
+class SlotDescriptor;
+class TupleDescriptor;
+
+namespace io {
+class IOContext;
+} // namespace io
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/format/table/iceberg_reader.h b/be/src/vec/exec/format/table/iceberg_reader.h
index 3477ee0f83..c4df688d28 100644
--- a/be/src/vec/exec/format/table/iceberg_reader.h
+++ b/be/src/vec/exec/format/table/iceberg_reader.h
@@ -17,19 +17,46 @@
 
 #pragma once
 
-#include <queue>
+#include <stddef.h>
+#include <stdint.h>
 
-#include "io/io_common.h"
+#include <string>
+#include <tuple>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
+#include "exec/olap_common.h"
 #include "table_format_reader.h"
+#include "util/runtime_profile.h"
 #include "vec/columns/column_dictionary.h"
-#include "vec/exec/format/format_common.h"
-#include "vec/exec/format/generic_reader.h"
-#include "vec/exec/format/parquet/parquet_common.h"
-#include "vec/exprs/vexpr.h"
+
+namespace tparquet {
+class KeyValue;
+} // namespace tparquet
 
 namespace doris {
+class RowDescriptor;
+class RuntimeState;
+class SlotDescriptor;
+class TFileRangeDesc;
+class TFileScanRangeParams;
+class TIcebergDeleteFileDesc;
+class TupleDescriptor;
+
+namespace io {
+class IOContext;
+} // namespace io
+struct TypeDescriptor;
 
 namespace vectorized {
+class Block;
+class ColumnString;
+class GenericReader;
+class ShardedKVCache;
+class VExprContext;
 
 class IcebergTableReader : public TableFormatReader {
 public:
diff --git a/be/src/vec/exec/format/table/table_format_reader.h b/be/src/vec/exec/format/table/table_format_reader.h
index 1c2ebec7d3..54e4a524e2 100644
--- a/be/src/vec/exec/format/table/table_format_reader.h
+++ b/be/src/vec/exec/format/table/table_format_reader.h
@@ -17,11 +17,24 @@
 
 #pragma once
 
-#include <string>
+#include <stddef.h>
 
-#include "runtime/runtime_state.h"
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+
+#include "common/status.h"
 #include "vec/exec/format/generic_reader.h"
-#include "vec/exec/format/parquet/parquet_common.h"
+
+namespace doris {
+class TFileRangeDesc;
+
+namespace vectorized {
+class Block;
+} // namespace vectorized
+struct TypeDescriptor;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/jni_connector.cpp b/be/src/vec/exec/jni_connector.cpp
index 80c4f4eb8a..ad984042c4 100644
--- a/be/src/vec/exec/jni_connector.cpp
+++ b/be/src/vec/exec/jni_connector.cpp
@@ -17,7 +17,27 @@
 
 #include "jni_connector.h"
 
+#include <glog/logging.h>
+#include <stdint.h>
+
 #include <sstream>
+#include <variant>
+
+#include "jni.h"
+#include "runtime/decimalv2_value.h"
+#include "runtime/runtime_state.h"
+#include "util/jni-util.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type_nullable.h"
+
+namespace doris {
+class RuntimeProfile;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/jni_connector.h b/be/src/vec/exec/jni_connector.h
index 7e74ecd6e3..14915ac533 100644
--- a/be/src/vec/exec/jni_connector.h
+++ b/be/src/vec/exec/jni_connector.h
@@ -18,9 +18,39 @@
 #pragma once
 
 #include <jni.h>
+#include <string.h>
 
-#include "runtime/runtime_state.h"
-#include "util/jni-util.h"
+#include <map>
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
+#include "exec/olap_common.h"
+#include "exec/olap_utils.h"
+#include "runtime/define_primitive_type.h"
+#include "runtime/primitive_type.h"
+#include "runtime/types.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/common/string_ref.h"
+#include "vec/data_types/data_type.h"
+
+namespace doris {
+class RuntimeProfile;
+class RuntimeState;
+
+namespace vectorized {
+class Block;
+template <typename T>
+class ColumnDecimal;
+template <typename T>
+class ColumnVector;
+template <typename T>
+struct Decimal;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/join/cross_join_impl.cpp b/be/src/vec/exec/join/cross_join_impl.cpp
index ce79087a7a..74fbba2027 100644
--- a/be/src/vec/exec/join/cross_join_impl.cpp
+++ b/be/src/vec/exec/join/cross_join_impl.cpp
@@ -15,7 +15,8 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include "gen_cpp/PlanNodes_types.h"
+#include <gen_cpp/PlanNodes_types.h>
+
 #include "process_hash_table_probe_impl.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/exec/join/full_outer_join_impl.cpp b/be/src/vec/exec/join/full_outer_join_impl.cpp
index 11486c3d0a..2786839097 100644
--- a/be/src/vec/exec/join/full_outer_join_impl.cpp
+++ b/be/src/vec/exec/join/full_outer_join_impl.cpp
@@ -15,7 +15,8 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include "gen_cpp/PlanNodes_types.h"
+#include <gen_cpp/PlanNodes_types.h>
+
 #include "process_hash_table_probe_impl.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/exec/join/inner_join_impl.cpp b/be/src/vec/exec/join/inner_join_impl.cpp
index cde5f141a6..6c1e0d399a 100644
--- a/be/src/vec/exec/join/inner_join_impl.cpp
+++ b/be/src/vec/exec/join/inner_join_impl.cpp
@@ -15,7 +15,8 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include "gen_cpp/PlanNodes_types.h"
+#include <gen_cpp/PlanNodes_types.h>
+
 #include "process_hash_table_probe_impl.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/exec/join/left_anti_join_impl.cpp b/be/src/vec/exec/join/left_anti_join_impl.cpp
index 762117e7c3..e1e1037eea 100644
--- a/be/src/vec/exec/join/left_anti_join_impl.cpp
+++ b/be/src/vec/exec/join/left_anti_join_impl.cpp
@@ -15,7 +15,8 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include "gen_cpp/PlanNodes_types.h"
+#include <gen_cpp/PlanNodes_types.h>
+
 #include "process_hash_table_probe_impl.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/exec/join/left_outer_join_impl.cpp b/be/src/vec/exec/join/left_outer_join_impl.cpp
index c663f64a94..681bd2b816 100644
--- a/be/src/vec/exec/join/left_outer_join_impl.cpp
+++ b/be/src/vec/exec/join/left_outer_join_impl.cpp
@@ -15,7 +15,8 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include "gen_cpp/PlanNodes_types.h"
+#include <gen_cpp/PlanNodes_types.h>
+
 #include "process_hash_table_probe_impl.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/exec/join/left_semi_join_impl.cpp b/be/src/vec/exec/join/left_semi_join_impl.cpp
index f62e6d5cd0..d06fe1783f 100644
--- a/be/src/vec/exec/join/left_semi_join_impl.cpp
+++ b/be/src/vec/exec/join/left_semi_join_impl.cpp
@@ -15,7 +15,8 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include "gen_cpp/PlanNodes_types.h"
+#include <gen_cpp/PlanNodes_types.h>
+
 #include "process_hash_table_probe_impl.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/exec/join/null_aware_left_anti_join_impl.cpp b/be/src/vec/exec/join/null_aware_left_anti_join_impl.cpp
index 7432e9ef1f..8b541d3749 100644
--- a/be/src/vec/exec/join/null_aware_left_anti_join_impl.cpp
+++ b/be/src/vec/exec/join/null_aware_left_anti_join_impl.cpp
@@ -15,7 +15,8 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include "gen_cpp/PlanNodes_types.h"
+#include <gen_cpp/PlanNodes_types.h>
+
 #include "process_hash_table_probe_impl.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/exec/join/process_hash_table_probe_impl.h b/be/src/vec/exec/join/process_hash_table_probe_impl.h
index c5fbc34b01..e1ce4c1c91 100644
--- a/be/src/vec/exec/join/process_hash_table_probe_impl.h
+++ b/be/src/vec/exec/join/process_hash_table_probe_impl.h
@@ -19,7 +19,9 @@
 
 #include "common/status.h"
 #include "process_hash_table_probe.h"
+#include "runtime/thread_context.h" // IWYU pragma: keep
 #include "util/simd/bits.h"
+#include "vec/exprs/vexpr_context.h"
 #include "vhash_join_node.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/exec/join/right_anti_join_impl.cpp b/be/src/vec/exec/join/right_anti_join_impl.cpp
index 79d7ab7a49..1c4eec5cf7 100644
--- a/be/src/vec/exec/join/right_anti_join_impl.cpp
+++ b/be/src/vec/exec/join/right_anti_join_impl.cpp
@@ -15,7 +15,8 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include "gen_cpp/PlanNodes_types.h"
+#include <gen_cpp/PlanNodes_types.h>
+
 #include "process_hash_table_probe_impl.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/exec/join/right_outer_join_impl.cpp b/be/src/vec/exec/join/right_outer_join_impl.cpp
index eea4bda8a1..e732639610 100644
--- a/be/src/vec/exec/join/right_outer_join_impl.cpp
+++ b/be/src/vec/exec/join/right_outer_join_impl.cpp
@@ -15,7 +15,8 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include "gen_cpp/PlanNodes_types.h"
+#include <gen_cpp/PlanNodes_types.h>
+
 #include "process_hash_table_probe_impl.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/exec/join/right_semi_join_impl.cpp b/be/src/vec/exec/join/right_semi_join_impl.cpp
index 354b3018b9..0ee9373443 100644
--- a/be/src/vec/exec/join/right_semi_join_impl.cpp
+++ b/be/src/vec/exec/join/right_semi_join_impl.cpp
@@ -15,7 +15,8 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include "gen_cpp/PlanNodes_types.h"
+#include <gen_cpp/PlanNodes_types.h>
+
 #include "process_hash_table_probe_impl.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/exec/join/vhash_join_node.cpp b/be/src/vec/exec/join/vhash_join_node.cpp
index 2dc9e09073..5c3eb3edd9 100644
--- a/be/src/vec/exec/join/vhash_join_node.cpp
+++ b/be/src/vec/exec/join/vhash_join_node.cpp
@@ -17,12 +17,52 @@
 
 #include "vec/exec/join/vhash_join_node.h"
 
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/Opcodes_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <glog/logging.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+
+#include <algorithm>
+#include <array>
+#include <boost/iterator/iterator_facade.hpp>
+#include <functional>
+#include <map>
+#include <new>
+#include <ostream>
+#include <type_traits>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/config.h"
+#include "common/object_pool.h"
+#include "exec/exec_node.h"
 #include "exprs/bloom_filter_func.h"
+#include "exprs/runtime_filter.h"
 #include "exprs/runtime_filter_slots.h"
-#include "gen_cpp/PlanNodes_types.h"
 #include "gutil/strings/substitute.h"
+#include "runtime/define_primitive_type.h"
+#include "runtime/descriptors.h"
+#include "runtime/query_fragments_ctx.h"
+#include "runtime/runtime_filter_mgr.h"
+#include "runtime/runtime_state.h"
+#include "runtime/thread_context.h"
 #include "util/defer_op.h"
+#include "util/telemetry/telemetry.h"
+#include "util/uid_util.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/hash_table/hash_map.h"
+#include "vec/common/uint128.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
+#include "vec/exec/join/join_op.h"
+#include "vec/exec/join/process_hash_table_probe.h"
+#include "vec/exec/join/vjoin_node_base.h"
 #include "vec/exprs/vexpr.h"
 #include "vec/exprs/vexpr_context.h"
 #include "vec/runtime/shared_hash_table_controller.h"
diff --git a/be/src/vec/exec/join/vhash_join_node.h b/be/src/vec/exec/join/vhash_join_node.h
index 1fc592daf2..960988e718 100644
--- a/be/src/vec/exec/join/vhash_join_node.h
+++ b/be/src/vec/exec/join/vhash_join_node.h
@@ -17,26 +17,51 @@
 
 #pragma once
 
-#include <future>
-#include <variant>
+#include <gen_cpp/PlanNodes_types.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <iosfwd>
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <variant>
+#include <vector>
+
+#include "common/global_types.h"
+#include "common/status.h"
 #include "exprs/runtime_filter_slots.h"
-#include "join_op.h"
-#include "process_hash_table_probe.h"
+#include "util/runtime_profile.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/aggregation_common.h"
+#include "vec/common/arena.h"
 #include "vec/common/columns_hashing.h"
-#include "vec/common/hash_table/hash_map.h"
 #include "vec/common/hash_table/partitioned_hash_map.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/types.h"
+#include "vec/exec/join/join_op.h" // IWYU pragma: keep
 #include "vec/runtime/shared_hash_table_controller.h"
 #include "vjoin_node_base.h"
 
+template <typename T>
+struct HashCRC32;
+
 namespace doris {
 
 class ObjectPool;
 class IRuntimeFilter;
+class DescriptorTbl;
+class RuntimeState;
 
 namespace vectorized {
 
-class SharedHashTableController;
+struct UInt128;
+struct UInt256;
+template <int JoinOpType>
+struct ProcessHashTableProbe;
 
 template <typename RowRefListType>
 struct SerializedHashTableContext {
@@ -173,7 +198,6 @@ using HashTableVariants = std::variant<
         I256FixedKeyHashTableContext<false, RowRefListWithFlags>>;
 
 class VExprContext;
-class HashJoinNode;
 
 using HashTableCtxVariants =
         std::variant<std::monostate, ProcessHashTableProbe<TJoinOp::INNER_JOIN>,
diff --git a/be/src/vec/exec/join/vjoin_node_base.cpp b/be/src/vec/exec/join/vjoin_node_base.cpp
index 7daf1fe237..e940495cab 100644
--- a/be/src/vec/exec/join/vjoin_node_base.cpp
+++ b/be/src/vec/exec/join/vjoin_node_base.cpp
@@ -17,12 +17,36 @@
 
 #include "vec/exec/join/vjoin_node_base.h"
 
-#include <sstream>
+#include <gen_cpp/Exprs_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <glog/logging.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+#include <opentelemetry/trace/span.h>
+#include <opentelemetry/trace/tracer.h>
+#include <stddef.h>
 
-#include "gen_cpp/PlanNodes_types.h"
+#include <sstream>
+#include <system_error>
+
+#include "runtime/exec_env.h"
 #include "runtime/runtime_state.h"
+#include "runtime/thread_context.h"
+#include "util/telemetry/telemetry.h"
+#include "util/threadpool.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/data_types/data_type.h"
+#include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
 #include "vec/utils/util.hpp"
 
+namespace doris {
+class ObjectPool;
+} // namespace doris
+
 namespace doris::vectorized {
 
 VJoinNodeBase::VJoinNodeBase(ObjectPool* pool, const TPlanNode& tnode, const DescriptorTbl& descs)
diff --git a/be/src/vec/exec/join/vjoin_node_base.h b/be/src/vec/exec/join/vjoin_node_base.h
index e7403e6522..9a33f41579 100644
--- a/be/src/vec/exec/join/vjoin_node_base.h
+++ b/be/src/vec/exec/join/vjoin_node_base.h
@@ -17,13 +17,29 @@
 
 #pragma once
 
-#include <future>
-#include <string>
-#include <variant>
+#include <gen_cpp/PlanNodes_types.h>
 
+#include <future>
+#include <memory>
+#include <type_traits>
+#include <variant>
+#include <vector>
+
+#include "common/status.h"
 #include "exec/exec_node.h"
-#include "gen_cpp/PlanNodes_types.h"
 #include "runtime/descriptors.h"
+#include "util/runtime_profile.h"
+#include "vec/core/block.h"
+#include "vec/data_types/data_type.h"
+
+namespace doris {
+class ObjectPool;
+class RuntimeState;
+
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/join/vnested_loop_join_node.cpp b/be/src/vec/exec/join/vnested_loop_join_node.cpp
index ee7d0cdc51..d0b308264e 100644
--- a/be/src/vec/exec/join/vnested_loop_join_node.cpp
+++ b/be/src/vec/exec/join/vnested_loop_join_node.cpp
@@ -17,21 +17,50 @@
 
 #include "vec/exec/join/vnested_loop_join_node.h"
 
+#include <gen_cpp/Exprs_types.h>
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/PlanNodes_types.h>
 #include <glog/logging.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+#include <string.h>
 
+#include <boost/iterator/iterator_facade.hpp>
+#include <functional>
 #include <sstream>
+#include <string>
+#include <type_traits>
+#include <utility>
+#include <variant>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/status.h"
+#include "exec/exec_node.h"
+#include "exprs/runtime_filter.h"
 #include "exprs/runtime_filter_slots_cross.h"
-#include "gen_cpp/PlanNodes_types.h"
+#include "gutil/integral_types.h"
+#include "runtime/descriptors.h"
+#include "runtime/runtime_filter_mgr.h"
 #include "runtime/runtime_state.h"
 #include "util/runtime_profile.h"
 #include "util/simd/bits.h"
+#include "util/telemetry/telemetry.h"
 #include "vec/columns/column_const.h"
-#include "vec/common/typeid_cast.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_number.h"
+#include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
 #include "vec/utils/template_helpers.hpp"
-#include "vec/utils/util.hpp"
+
+namespace doris {
+class ObjectPool;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/join/vnested_loop_join_node.h b/be/src/vec/exec/join/vnested_loop_join_node.h
index 537ed2d7f0..4249af9266 100644
--- a/be/src/vec/exec/join/vnested_loop_join_node.h
+++ b/be/src/vec/exec/join/vnested_loop_join_node.h
@@ -17,17 +17,34 @@
 
 #pragma once
 
-#include <boost/thread.hpp>
-#include <future>
-#include <stack>
-#include <string>
+#include <gen_cpp/PlanNodes_types.h>
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
 
-#include "exprs/runtime_filter.h"
-#include "gen_cpp/PlanNodes_types.h"
-#include "runtime/descriptors.h"
+#include <algorithm>
+#include <iosfwd>
+#include <memory>
+#include <stack>
+#include <vector>
+
+#include "common/status.h"
+#include "runtime/runtime_state.h"
+#include "runtime/thread_context.h"
+#include "vec/columns/column.h"
 #include "vec/core/block.h"
 #include "vec/exec/join/vjoin_node_base.h"
 
+namespace doris {
+class DescriptorTbl;
+class ObjectPool;
+class RowDescriptor;
+
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 // Node for nested loop joins.
diff --git a/be/src/vec/exec/scan/jni_reader.cpp b/be/src/vec/exec/scan/jni_reader.cpp
index 57818f5503..bbf9a5f8d0 100644
--- a/be/src/vec/exec/scan/jni_reader.cpp
+++ b/be/src/vec/exec/scan/jni_reader.cpp
@@ -17,6 +17,22 @@
 
 #include "jni_reader.h"
 
+#include <map>
+#include <ostream>
+
+#include "runtime/descriptors.h"
+#include "runtime/types.h"
+#include "vec/core/types.h"
+
+namespace doris {
+class RuntimeProfile;
+class RuntimeState;
+
+namespace vectorized {
+class Block;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 MockJniReader::MockJniReader(const std::vector<SlotDescriptor*>& file_slot_descs,
diff --git a/be/src/vec/exec/scan/jni_reader.h b/be/src/vec/exec/scan/jni_reader.h
index 1505aa7fc0..500301bd4b 100644
--- a/be/src/vec/exec/scan/jni_reader.h
+++ b/be/src/vec/exec/scan/jni_reader.h
@@ -17,9 +17,29 @@
 
 #pragma once
 
+#include <stddef.h>
+
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+#include "common/status.h"
+#include "exec/olap_common.h"
 #include "vec/exec/format/generic_reader.h"
 #include "vec/exec/jni_connector.h"
 
+namespace doris {
+class RuntimeProfile;
+class RuntimeState;
+class SlotDescriptor;
+namespace vectorized {
+class Block;
+} // namespace vectorized
+struct TypeDescriptor;
+} // namespace doris
+
 namespace doris::vectorized {
 
 /**
diff --git a/be/src/vec/exec/scan/new_es_scan_node.cpp b/be/src/vec/exec/scan/new_es_scan_node.cpp
index d513552857..efea42cb70 100644
--- a/be/src/vec/exec/scan/new_es_scan_node.cpp
+++ b/be/src/vec/exec/scan/new_es_scan_node.cpp
@@ -17,9 +17,26 @@
 
 #include "vec/exec/scan/new_es_scan_node.h"
 
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/Types_types.h>
+
+#include <ostream>
+
+#include "common/logging.h"
+#include "common/object_pool.h"
+#include "exec/es/es_scan_reader.h"
 #include "exec/es/es_scroll_query.h"
+#include "runtime/descriptors.h"
+#include "runtime/runtime_state.h"
+#include "service/backend_options.h"
 #include "vec/exec/scan/new_es_scanner.h"
-#include "vec/utils/util.hpp"
+
+namespace doris {
+namespace vectorized {
+class VScanner;
+} // namespace vectorized
+} // namespace doris
 
 static const std::string NEW_SCAN_NODE_TYPE = "NewEsScanNode";
 
diff --git a/be/src/vec/exec/scan/new_es_scan_node.h b/be/src/vec/exec/scan/new_es_scan_node.h
index 857f1d0067..a8e48b006c 100644
--- a/be/src/vec/exec/scan/new_es_scan_node.h
+++ b/be/src/vec/exec/scan/new_es_scan_node.h
@@ -17,9 +17,31 @@
 
 #pragma once
 
-#include "vec/exec/scan/new_es_scanner.h"
+#include <gen_cpp/PlanNodes_types.h>
+
+#include <list>
+#include <map>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "common/global_types.h"
+#include "common/status.h"
+#include "util/runtime_profile.h"
 #include "vec/exec/scan/vscan_node.h"
 
+namespace doris {
+class DescriptorTbl;
+class ObjectPool;
+class RuntimeState;
+class TScanRangeParams;
+class TupleDescriptor;
+
+namespace vectorized {
+class VScanner;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 class NewEsScanNode : public VScanNode {
diff --git a/be/src/vec/exec/scan/new_es_scanner.cpp b/be/src/vec/exec/scan/new_es_scanner.cpp
index 18233dde6c..a33d4a59d3 100644
--- a/be/src/vec/exec/scan/new_es_scanner.cpp
+++ b/be/src/vec/exec/scan/new_es_scanner.cpp
@@ -17,8 +17,26 @@
 
 #include "vec/exec/scan/new_es_scanner.h"
 
+#include <algorithm>
+#include <ostream>
+#include <utility>
+
+#include "common/logging.h"
+#include "runtime/descriptors.h"
+#include "runtime/runtime_state.h"
+#include "util/runtime_profile.h"
+#include "vec/columns/column.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/exec/scan/new_es_scan_node.h"
 
+namespace doris {
+namespace vectorized {
+class VExprContext;
+class VScanNode;
+} // namespace vectorized
+} // namespace doris
+
 static const std::string NEW_SCANNER_TYPE = "NewEsScanner";
 
 namespace doris::vectorized {
diff --git a/be/src/vec/exec/scan/new_es_scanner.h b/be/src/vec/exec/scan/new_es_scanner.h
index 77276d9ae1..db1c479712 100644
--- a/be/src/vec/exec/scan/new_es_scanner.h
+++ b/be/src/vec/exec/scan/new_es_scanner.h
@@ -17,11 +17,31 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include <map>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "common/global_types.h"
+#include "common/status.h"
 #include "exec/es/es_scan_reader.h"
 #include "exec/es/es_scroll_parser.h"
-#include "runtime/runtime_state.h"
+#include "vec/data_types/data_type.h"
 #include "vec/exec/scan/vscanner.h"
 
+namespace doris {
+class RuntimeProfile;
+class RuntimeState;
+class TupleDescriptor;
+
+namespace vectorized {
+class Block;
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 class NewEsScanNode;
diff --git a/be/src/vec/exec/scan/new_file_scan_node.cpp b/be/src/vec/exec/scan/new_file_scan_node.cpp
index 4a762ac1a4..aea5571822 100644
--- a/be/src/vec/exec/scan/new_file_scan_node.cpp
+++ b/be/src/vec/exec/scan/new_file_scan_node.cpp
@@ -17,7 +17,22 @@
 
 #include "vec/exec/scan/new_file_scan_node.h"
 
+#include <gen_cpp/PlanNodes_types.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <ostream>
+
+#include "common/config.h"
+#include "common/object_pool.h"
 #include "vec/exec/scan/vfile_scanner.h"
+#include "vec/exec/scan/vscanner.h"
+
+namespace doris {
+class DescriptorTbl;
+class RuntimeState;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/scan/new_file_scan_node.h b/be/src/vec/exec/scan/new_file_scan_node.h
index 2f67ca9713..5cd7458cae 100644
--- a/be/src/vec/exec/scan/new_file_scan_node.h
+++ b/be/src/vec/exec/scan/new_file_scan_node.h
@@ -17,9 +17,26 @@
 
 #pragma once
 
+#include <gen_cpp/PaloInternalService_types.h>
+
+#include <list>
+#include <memory>
+#include <vector>
+
+#include "common/status.h"
 #include "vec/exec/format/format_common.h"
 #include "vec/exec/scan/vscan_node.h"
 
+namespace doris {
+class DescriptorTbl;
+class ObjectPool;
+class RuntimeState;
+class TPlanNode;
+namespace vectorized {
+class VScanner;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 class NewFileScanNode : public VScanNode {
diff --git a/be/src/vec/exec/scan/new_jdbc_scan_node.cpp b/be/src/vec/exec/scan/new_jdbc_scan_node.cpp
index 19fd93b9d3..a4dff27f34 100644
--- a/be/src/vec/exec/scan/new_jdbc_scan_node.cpp
+++ b/be/src/vec/exec/scan/new_jdbc_scan_node.cpp
@@ -17,8 +17,24 @@
 
 #include "vec/exec/scan/new_jdbc_scan_node.h"
 
+#include <fmt/format.h>
+#include <gen_cpp/PlanNodes_types.h>
+
+#include <memory>
+#include <ostream>
+
+#include "common/logging.h"
+#include "common/object_pool.h"
+#include "runtime/runtime_state.h"
 #include "vec/exec/scan/new_jdbc_scanner.h"
-#include "vec/exec/scan/vscanner.h"
+
+namespace doris {
+class DescriptorTbl;
+namespace vectorized {
+class VScanner;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 NewJdbcScanNode::NewJdbcScanNode(ObjectPool* pool, const TPlanNode& tnode,
                                  const DescriptorTbl& descs)
diff --git a/be/src/vec/exec/scan/new_jdbc_scan_node.h b/be/src/vec/exec/scan/new_jdbc_scan_node.h
index 1cc34e83f0..7e109d1dd8 100644
--- a/be/src/vec/exec/scan/new_jdbc_scan_node.h
+++ b/be/src/vec/exec/scan/new_jdbc_scan_node.h
@@ -17,11 +17,24 @@
 
 #pragma once
 
-#include "runtime/runtime_state.h"
+#include <gen_cpp/Types_types.h>
+
+#include <list>
+#include <string>
+
+#include "common/global_types.h"
+#include "common/status.h"
 #include "vec/exec/scan/vscan_node.h"
 
 namespace doris {
+class DescriptorTbl;
+class ObjectPool;
+class RuntimeState;
+class TPlanNode;
+
 namespace vectorized {
+class VScanner;
+
 class NewJdbcScanNode : public VScanNode {
 public:
     NewJdbcScanNode(ObjectPool* pool, const TPlanNode& tnode, const DescriptorTbl& descs);
diff --git a/be/src/vec/exec/scan/new_jdbc_scanner.cpp b/be/src/vec/exec/scan/new_jdbc_scanner.cpp
index b74e279d97..1de4a10cf9 100644
--- a/be/src/vec/exec/scan/new_jdbc_scanner.cpp
+++ b/be/src/vec/exec/scan/new_jdbc_scanner.cpp
@@ -17,8 +17,23 @@
 
 #include "new_jdbc_scanner.h"
 
+#include <new>
+#include <ostream>
+#include <utility>
+#include <vector>
+
+#include "common/logging.h"
+#include "runtime/descriptors.h"
+#include "runtime/runtime_state.h"
 #include "util/runtime_profile.h"
+#include "vec/columns/column.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/data_types/data_type.h"
+#include "vec/exec/scan/new_jdbc_scan_node.h"
+#include "vec/exec/scan/vscan_node.h"
 #include "vec/exec/vjdbc_connector.h"
+#include "vec/exprs/vexpr_context.h"
 
 namespace doris::vectorized {
 NewJdbcScanner::NewJdbcScanner(RuntimeState* state, NewJdbcScanNode* parent, int64_t limit,
diff --git a/be/src/vec/exec/scan/new_jdbc_scanner.h b/be/src/vec/exec/scan/new_jdbc_scanner.h
index da5a3e7faf..1cf06387ff 100644
--- a/be/src/vec/exec/scan/new_jdbc_scanner.h
+++ b/be/src/vec/exec/scan/new_jdbc_scanner.h
@@ -17,13 +17,27 @@
 
 #pragma once
 
-#include "runtime/runtime_state.h"
+#include <gen_cpp/Types_types.h>
+#include <stdint.h>
+
+#include <memory>
+#include <string>
+
+#include "common/global_types.h"
+#include "common/status.h"
 #include "util/runtime_profile.h"
-#include "vec/exec/scan/new_jdbc_scan_node.h"
 #include "vec/exec/scan/vscanner.h"
 #include "vec/exec/vjdbc_connector.h"
+
 namespace doris {
+class RuntimeState;
+class TupleDescriptor;
+
 namespace vectorized {
+class Block;
+class NewJdbcScanNode;
+class VExprContext;
+
 class NewJdbcScanner : public VScanner {
 public:
     friend class JdbcConnector;
diff --git a/be/src/vec/exec/scan/new_odbc_scan_node.cpp b/be/src/vec/exec/scan/new_odbc_scan_node.cpp
index dbbf57a120..170122751e 100644
--- a/be/src/vec/exec/scan/new_odbc_scan_node.cpp
+++ b/be/src/vec/exec/scan/new_odbc_scan_node.cpp
@@ -17,8 +17,23 @@
 
 #include "vec/exec/scan/new_odbc_scan_node.h"
 
+#include <fmt/format.h>
+
+#include <memory>
+#include <ostream>
+
+#include "common/logging.h"
+#include "common/object_pool.h"
+#include "runtime/runtime_state.h"
 #include "vec/exec/scan/new_odbc_scanner.h"
 
+namespace doris {
+class DescriptorTbl;
+namespace vectorized {
+class VScanner;
+} // namespace vectorized
+} // namespace doris
+
 static const std::string NEW_SCAN_NODE_TYPE = "NewOdbcScanNode";
 
 namespace doris::vectorized {
diff --git a/be/src/vec/exec/scan/new_odbc_scan_node.h b/be/src/vec/exec/scan/new_odbc_scan_node.h
index 03d18ce58f..a6107ac402 100644
--- a/be/src/vec/exec/scan/new_odbc_scan_node.h
+++ b/be/src/vec/exec/scan/new_odbc_scan_node.h
@@ -17,7 +17,25 @@
 
 #pragma once
 
+#include <gen_cpp/PlanNodes_types.h>
+
+#include <list>
+#include <string>
+#include <vector>
+
+#include "common/status.h"
 #include "vec/exec/scan/vscan_node.h"
+
+namespace doris {
+class DescriptorTbl;
+class ObjectPool;
+class RuntimeState;
+class TScanRangeParams;
+namespace vectorized {
+class VScanner;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 class NewOdbcScanNode : public VScanNode {
 public:
diff --git a/be/src/vec/exec/scan/new_odbc_scanner.cpp b/be/src/vec/exec/scan/new_odbc_scanner.cpp
index f19b44c294..5c912909b4 100644
--- a/be/src/vec/exec/scan/new_odbc_scanner.cpp
+++ b/be/src/vec/exec/scan/new_odbc_scanner.cpp
@@ -17,10 +17,35 @@
 
 #include "vec/exec/scan/new_odbc_scanner.h"
 
+#include <gen_cpp/PlanNodes_types.h>
+#include <sql.h>
+
+#include <algorithm>
+#include <new>
+#include <ostream>
+#include <utility>
+#include <vector>
+
+#include "common/logging.h"
 #include "common/status.h"
 #include "exec/text_converter.hpp"
+#include "runtime/descriptors.h"
+#include "runtime/runtime_state.h"
+#include "runtime/types.h"
+#include "vec/columns/column.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/data_types/data_type.h"
 #include "vec/exec/scan/new_odbc_scan_node.h"
 #include "vec/exec/scan/vscanner.h"
+#include "vec/exprs/vexpr_context.h"
+
+namespace doris {
+class RuntimeProfile;
+namespace vectorized {
+class VScanNode;
+} // namespace vectorized
+} // namespace doris
 
 static const std::string NEW_SCANNER_TYPE = "NewOdbcScanner";
 
diff --git a/be/src/vec/exec/scan/new_odbc_scanner.h b/be/src/vec/exec/scan/new_odbc_scanner.h
index e238a544fe..7097215b0b 100644
--- a/be/src/vec/exec/scan/new_odbc_scanner.h
+++ b/be/src/vec/exec/scan/new_odbc_scanner.h
@@ -17,11 +17,30 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include <memory>
+#include <string>
+
+#include "common/global_types.h"
+#include "common/status.h"
 #include "exec/odbc_connector.h"
 #include "exec/text_converter.h"
-#include "vec/exec/scan/new_odbc_scan_node.h"
 #include "vec/exec/scan/vscanner.h"
 
+namespace doris {
+class RuntimeProfile;
+class RuntimeState;
+class TOdbcScanNode;
+class TupleDescriptor;
+
+namespace vectorized {
+class Block;
+class NewOdbcScanNode;
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 class NewOdbcScanner : public VScanner {
 public:
diff --git a/be/src/vec/exec/scan/new_olap_scan_node.cpp b/be/src/vec/exec/scan/new_olap_scan_node.cpp
index 3021ae3437..f9965260b3 100644
--- a/be/src/vec/exec/scan/new_olap_scan_node.cpp
+++ b/be/src/vec/exec/scan/new_olap_scan_node.cpp
@@ -17,16 +17,54 @@
 
 #include "vec/exec/scan/new_olap_scan_node.h"
 
-#include <charconv>
+#include <fmt/format.h>
+#include <gen_cpp/Exprs_types.h>
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/olap_file.pb.h>
+#include <opentelemetry/trace/tracer.h>
+#include <stdio.h>
 
+#include <algorithm>
+#include <charconv>
+#include <ostream>
+#include <shared_mutex>
+#include <unordered_map>
+#include <utility>
+#include <variant>
+
+#include "common/config.h"
+#include "common/logging.h"
+#include "common/object_pool.h"
 #include "common/status.h"
+#include "exec/exec_node.h"
+#include "olap/rowset/rowset.h"
+#include "olap/rowset/rowset_reader.h"
 #include "olap/storage_engine.h"
 #include "olap/tablet.h"
-#include "pipeline/pipeline.h"
-#include "pipeline/pipeline_fragment_context.h"
+#include "olap/tablet_manager.h"
+#include "runtime/decimalv2_value.h"
+#include "runtime/query_statistics.h"
+#include "runtime/runtime_state.h"
+#include "runtime/types.h"
+#include "service/backend_options.h"
+#include "util/time.h"
 #include "util/to_string.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
+#include "vec/common/string_ref.h"
 #include "vec/exec/scan/new_olap_scanner.h"
+#include "vec/exprs/vectorized_fn_call.h"
+#include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
+
+namespace doris {
+class DescriptorTbl;
+class FunctionContext;
+namespace vectorized {
+class VScanner;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/scan/new_olap_scan_node.h b/be/src/vec/exec/scan/new_olap_scan_node.h
index 48fde8436b..441fd4a3ae 100644
--- a/be/src/vec/exec/scan/new_olap_scan_node.h
+++ b/be/src/vec/exec/scan/new_olap_scan_node.h
@@ -17,15 +17,44 @@
 
 #pragma once
 
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <stdint.h>
+
+#include <list>
+#include <memory>
+#include <set>
+#include <string>
+#include <vector>
+
+#include "common/status.h"
+#include "exec/olap_common.h"
+#include "exec/olap_utils.h"
+#include "olap/olap_common.h"
+#include "util/runtime_profile.h"
 #include "vec/exec/scan/vscan_node.h"
 
+namespace doris {
+class DescriptorTbl;
+class FunctionContext;
+class ObjectPool;
+class QueryStatistics;
+class RuntimeState;
+
+namespace vectorized {
+class VExprContext;
+class VScanner;
+class VectorizedFnCall;
+} // namespace vectorized
+struct StringRef;
+} // namespace doris
+
 namespace doris::pipeline {
 class OlapScanOperator;
 }
 
 namespace doris::vectorized {
 
-class NewOlapScanner;
 class NewOlapScanNode : public VScanNode {
 public:
     NewOlapScanNode(ObjectPool* pool, const TPlanNode& tnode, const DescriptorTbl& descs);
diff --git a/be/src/vec/exec/scan/new_olap_scanner.cpp b/be/src/vec/exec/scan/new_olap_scanner.cpp
index df42e69636..9310db989c 100644
--- a/be/src/vec/exec/scan/new_olap_scanner.cpp
+++ b/be/src/vec/exec/scan/new_olap_scanner.cpp
@@ -17,9 +17,40 @@
 
 #include "vec/exec/scan/new_olap_scanner.h"
 
+#include <gen_cpp/Descriptors_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <stdlib.h>
+
+#include <algorithm>
+#include <iterator>
+#include <ostream>
+#include <set>
+#include <shared_mutex>
+
+#include "common/config.h"
+#include "common/consts.h"
+#include "exec/olap_utils.h"
+#include "exprs/function_filter.h"
 #include "io/cache/block/block_file_cache_profile.h"
+#include "io/io_common.h"
+#include "olap/olap_common.h"
+#include "olap/olap_tuple.h"
+#include "olap/rowset/rowset.h"
+#include "olap/rowset/rowset_meta.h"
 #include "olap/storage_engine.h"
+#include "olap/tablet_manager.h"
+#include "olap/tablet_meta.h"
+#include "runtime/descriptors.h"
+#include "runtime/runtime_state.h"
+#include "service/backend_options.h"
+#include "util/doris_metrics.h"
+#include "util/runtime_profile.h"
+#include "vec/core/block.h"
 #include "vec/exec/scan/new_olap_scan_node.h"
+#include "vec/exec/scan/vscan_node.h"
+#include "vec/exprs/vexpr_context.h"
 #include "vec/olap/block_reader.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/exec/scan/new_olap_scanner.h b/be/src/vec/exec/scan/new_olap_scanner.h
index 4b05880bcd..5eebf5d27c 100644
--- a/be/src/vec/exec/scan/new_olap_scanner.h
+++ b/be/src/vec/exec/scan/new_olap_scanner.h
@@ -17,20 +17,36 @@
 
 #pragma once
 
-#include "exec/olap_utils.h"
-#include "exprs/function_filter.h"
+#include <gen_cpp/PaloInternalService_types.h>
+#include <stdint.h>
+
+#include <memory>
+#include <string>
+#include <unordered_set>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
+#include "olap/data_dir.h"
 #include "olap/reader.h"
-#include "util/runtime_profile.h"
+#include "olap/rowset/rowset_reader.h"
+#include "olap/tablet.h"
+#include "olap/tablet_schema.h"
 #include "vec/exec/scan/vscanner.h"
 
 namespace doris {
 
 struct OlapScanRange;
+class FunctionFilter;
+class RuntimeProfile;
+class RuntimeState;
+class TPaloScanRange;
 
 namespace vectorized {
 
 class NewOlapScanNode;
 struct FilterPredicates;
+class Block;
 
 class NewOlapScanner : public VScanner {
 public:
diff --git a/be/src/vec/exec/scan/scanner_context.cpp b/be/src/vec/exec/scan/scanner_context.cpp
index 8f5d7aac91..ae017776f1 100644
--- a/be/src/vec/exec/scan/scanner_context.cpp
+++ b/be/src/vec/exec/scan/scanner_context.cpp
@@ -17,12 +17,25 @@
 
 #include "scanner_context.h"
 
+#include <bthread/bthread.h>
+#include <fmt/format.h>
+#include <gen_cpp/Metrics_types.h>
+#include <glog/logging.h>
+
+#include <algorithm>
 #include <mutex>
+#include <ostream>
+#include <utility>
 
 #include "common/config.h"
+#include "runtime/descriptors.h"
+#include "runtime/exec_env.h"
+#include "runtime/query_fragments_ctx.h"
 #include "runtime/runtime_state.h"
-#include "util/threadpool.h"
+#include "util/pretty_printer.h"
+#include "util/uid_util.h"
 #include "vec/core/block.h"
+#include "vec/exec/scan/scanner_scheduler.h"
 #include "vec/exec/scan/vscan_node.h"
 #include "vec/exec/scan/vscanner.h"
 
diff --git a/be/src/vec/exec/scan/scanner_context.h b/be/src/vec/exec/scan/scanner_context.h
index 08cb353776..05942ad73b 100644
--- a/be/src/vec/exec/scan/scanner_context.h
+++ b/be/src/vec/exec/scan/scanner_context.h
@@ -17,22 +17,26 @@
 
 #pragma once
 
+#include <bthread/types.h>
+#include <stdint.h>
+
 #include <atomic>
-#include <condition_variable>
+#include <list>
+#include <memory>
 #include <mutex>
+#include <string>
+#include <vector>
 
 #include "common/status.h"
-#include "runtime/descriptors.h"
 #include "util/lock.h"
 #include "util/runtime_profile.h"
-#include "util/uid_util.h"
 #include "vec/core/block.h"
 
 namespace doris {
 
-class PriorityThreadPool;
-class ThreadPool;
 class ThreadPoolToken;
+class RuntimeState;
+class TupleDescriptor;
 
 namespace vectorized {
 
diff --git a/be/src/vec/exec/scan/scanner_scheduler.cpp b/be/src/vec/exec/scan/scanner_scheduler.cpp
index a6a76e57d7..63d5c9282c 100644
--- a/be/src/vec/exec/scan/scanner_scheduler.cpp
+++ b/be/src/vec/exec/scan/scanner_scheduler.cpp
@@ -17,17 +17,34 @@
 
 #include "scanner_scheduler.h"
 
+#include <stdint.h>
+
+#include <algorithm>
+#include <functional>
+#include <list>
+#include <ostream>
+#include <string>
+#include <typeinfo>
+#include <utility>
+#include <vector>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/config.h"
-#include "util/async_io.h"
+#include "common/logging.h"
+#include "olap/options.h"
+#include "olap/tablet.h"
+#include "runtime/exec_env.h"
+#include "runtime/runtime_state.h"
+#include "runtime/thread_context.h"
+#include "util/blocking_queue.hpp"
 #include "util/priority_thread_pool.hpp"
 #include "util/priority_work_stealing_thread_pool.hpp"
-#include "util/telemetry/telemetry.h"
 #include "util/thread.h"
 #include "util/threadpool.h"
 #include "vec/core/block.h"
-#include "vec/exec/scan/new_olap_scanner.h"
+#include "vec/exec/scan/scanner_context.h"
 #include "vec/exec/scan/vscanner.h"
-#include "vec/exprs/vexpr.h"
 #include "vfile_scanner.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/exec/scan/scanner_scheduler.h b/be/src/vec/exec/scan/scanner_scheduler.h
index 1b9049f213..6f03ec5ea1 100644
--- a/be/src/vec/exec/scan/scanner_scheduler.h
+++ b/be/src/vec/exec/scan/scanner_scheduler.h
@@ -17,11 +17,22 @@
 
 #pragma once
 
+#include <atomic>
+#include <memory>
+
 #include "common/status.h"
-#include "runtime/exec_env.h"
-#include "util/blocking_queue.hpp"
 #include "util/threadpool.h"
-#include "vec/exec/scan/scanner_context.h"
+
+namespace doris {
+class ExecEnv;
+class PriorityThreadPool;
+
+namespace vectorized {
+class VScanner;
+} // namespace vectorized
+template <typename T>
+class BlockingQueue;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/scan/vfile_scanner.cpp b/be/src/vec/exec/scan/vfile_scanner.cpp
index 44c567ec3d..41d53cd4c3 100644
--- a/be/src/vec/exec/scan/vfile_scanner.cpp
+++ b/be/src/vec/exec/scan/vfile_scanner.cpp
@@ -18,26 +18,65 @@
 #include "vec/exec/scan/vfile_scanner.h"
 
 #include <fmt/format.h>
-#include <thrift/protocol/TDebugProtocol.h>
+#include <gen_cpp/Exprs_types.h>
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/PlanNodes_types.h>
 
-#include <vec/data_types/data_type_factory.hpp>
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <iterator>
+#include <map>
+#include <ostream>
+#include <tuple>
+#include <utility>
 
+#include "vec/data_types/data_type_factory.hpp"
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/config.h"
 #include "common/logging.h"
-#include "common/utils.h"
-#include "exec/text_converter.hpp"
+#include "common/object_pool.h"
 #include "io/cache/block/block_file_cache_profile.h"
-#include "olap/iterators.h"
 #include "runtime/descriptors.h"
 #include "runtime/runtime_state.h"
+#include "runtime/types.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/field.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/data_types/data_type_number.h"
+#include "vec/data_types/data_type_string.h"
 #include "vec/exec/format/csv/csv_reader.h"
 #include "vec/exec/format/json/new_json_reader.h"
 #include "vec/exec/format/orc/vorc_reader.h"
 #include "vec/exec/format/parquet/vparquet_reader.h"
 #include "vec/exec/format/table/iceberg_reader.h"
 #include "vec/exec/scan/new_file_scan_node.h"
+#include "vec/exec/scan/vscan_node.h"
+#include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
 #include "vec/exprs/vslot_ref.h"
+#include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace cctz {
+class time_zone;
+} // namespace cctz
+namespace doris {
+namespace vectorized {
+class ShardedKVCache;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 using namespace ErrorCode;
 
diff --git a/be/src/vec/exec/scan/vfile_scanner.h b/be/src/vec/exec/scan/vfile_scanner.h
index cd25367848..ca42e9c1a6 100644
--- a/be/src/vec/exec/scan/vfile_scanner.h
+++ b/be/src/vec/exec/scan/vfile_scanner.h
@@ -17,15 +17,41 @@
 
 #pragma once
 
+#include <stddef.h>
+#include <stdint.h>
+
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+#include "common/global_types.h"
+#include "common/status.h"
 #include "exec/olap_common.h"
 #include "exec/text_converter.h"
-#include "exprs/function_filter.h"
-#include "io/file_factory.h"
+#include "io/io_common.h"
+#include "runtime/descriptors.h"
+#include "util/runtime_profile.h"
 #include "vec/common/schema_util.h"
-#include "vec/exec/format/format_common.h"
+#include "vec/core/block.h"
 #include "vec/exec/format/generic_reader.h"
 #include "vec/exec/scan/vscanner.h"
 
+namespace doris {
+class RuntimeState;
+class TFileRangeDesc;
+class TFileScanRange;
+class TFileScanRangeParams;
+
+namespace vectorized {
+class ShardedKVCache;
+class VExpr;
+class VExprContext;
+} // namespace vectorized
+struct TypeDescriptor;
+} // namespace doris
+
 namespace doris::vectorized {
 
 class NewFileScanNode;
diff --git a/be/src/vec/exec/scan/vmeta_scan_node.cpp b/be/src/vec/exec/scan/vmeta_scan_node.cpp
index 6a9e86636c..b3744311ac 100644
--- a/be/src/vec/exec/scan/vmeta_scan_node.cpp
+++ b/be/src/vec/exec/scan/vmeta_scan_node.cpp
@@ -17,8 +17,19 @@
 
 #include "vmeta_scan_node.h"
 
+#include <memory>
+
+#include "common/object_pool.h"
 #include "vmeta_scanner.h"
 
+namespace doris {
+class DescriptorTbl;
+class RuntimeState;
+namespace vectorized {
+class VScanner;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 VMetaScanNode::VMetaScanNode(ObjectPool* pool, const TPlanNode& tnode, const DescriptorTbl& descs)
diff --git a/be/src/vec/exec/scan/vmeta_scan_node.h b/be/src/vec/exec/scan/vmeta_scan_node.h
index 745882a2d3..8f8554883b 100644
--- a/be/src/vec/exec/scan/vmeta_scan_node.h
+++ b/be/src/vec/exec/scan/vmeta_scan_node.h
@@ -17,9 +17,26 @@
 
 #pragma once
 
-#include "runtime/runtime_state.h"
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+
+#include <list>
+#include <vector>
+
+#include "common/global_types.h"
+#include "common/status.h"
 #include "vec/exec/scan/vscan_node.h"
 
+namespace doris {
+class DescriptorTbl;
+class ObjectPool;
+class RuntimeState;
+
+namespace vectorized {
+class VScanner;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 class VMetaScanNode : public VScanNode {
diff --git a/be/src/vec/exec/scan/vmeta_scanner.cpp b/be/src/vec/exec/scan/vmeta_scanner.cpp
index 0ac07f9dd0..647c4c7de6 100644
--- a/be/src/vec/exec/scan/vmeta_scanner.cpp
+++ b/be/src/vec/exec/scan/vmeta_scanner.cpp
@@ -17,14 +17,42 @@
 
 #include "vmeta_scanner.h"
 
+#include <fmt/format.h>
+#include <gen_cpp/FrontendService.h>
 #include <gen_cpp/FrontendService_types.h>
 #include <gen_cpp/HeartbeatService_types.h>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Types_types.h>
 
-#include "gen_cpp/FrontendService.h"
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/logging.h"
 #include "runtime/client_cache.h"
 #include "runtime/define_primitive_type.h"
+#include "runtime/descriptors.h"
+#include "runtime/exec_env.h"
+#include "runtime/runtime_state.h"
+#include "runtime/types.h"
 #include "util/thrift_rpc_helper.h"
-#include "vec/runtime/vdatetime_value.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/exec/scan/vmeta_scan_node.h"
+
+namespace doris {
+class RuntimeProfile;
+namespace vectorized {
+class VExprContext;
+class VScanNode;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/scan/vmeta_scanner.h b/be/src/vec/exec/scan/vmeta_scanner.h
index 7b547e5ad1..c875d90b3c 100644
--- a/be/src/vec/exec/scan/vmeta_scanner.h
+++ b/be/src/vec/exec/scan/vmeta_scanner.h
@@ -17,9 +17,31 @@
 
 #pragma once
 
-#include "runtime/runtime_state.h"
+#include <gen_cpp/Data_types.h>
+#include <stdint.h>
+
+#include <vector>
+
+#include "common/global_types.h"
+#include "common/status.h"
+#include "vec/data_types/data_type.h"
 #include "vec/exec/scan/vscanner.h"
-#include "vmeta_scan_node.h"
+
+namespace doris {
+class RuntimeProfile;
+class RuntimeState;
+class TFetchSchemaTableDataRequest;
+class TMetaScanRange;
+class TScanRange;
+class TScanRangeParams;
+class TupleDescriptor;
+
+namespace vectorized {
+class Block;
+class VExprContext;
+class VMetaScanNode;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/scan/vscan_node.cpp b/be/src/vec/exec/scan/vscan_node.cpp
index a1d07b1f15..4789f46b86 100644
--- a/be/src/vec/exec/scan/vscan_node.cpp
+++ b/be/src/vec/exec/scan/vscan_node.cpp
@@ -17,21 +17,51 @@
 
 #include "vec/exec/scan/vscan_node.h"
 
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/Opcodes_types.h>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/Types_types.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+#include <string.h>
+
+#include <algorithm>
+#include <mutex>
+#include <ostream>
+#include <variant>
+
+#include "common/config.h"
 #include "common/consts.h"
+#include "common/logging.h"
 #include "common/status.h"
+#include "exec/olap_utils.h"
 #include "exprs/bloom_filter_func.h"
 #include "exprs/hybrid_set.h"
+#include "exprs/runtime_filter.h"
+#include "runtime/descriptors.h"
+#include "runtime/exec_env.h"
+#include "runtime/primitive_type.h"
 #include "runtime/runtime_filter_mgr.h"
+#include "runtime/types.h"
+#include "udf/udf.h"
 #include "util/defer_op.h"
 #include "util/runtime_profile.h"
+#include "util/telemetry/telemetry.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
+#include "vec/columns/column_vector.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/types.h"
 #include "vec/exec/scan/pip_scanner_context.h"
 #include "vec/exec/scan/scanner_scheduler.h"
-#include "vec/exec/scan/vscanner.h"
 #include "vec/exprs/vcompound_pred.h"
-#include "vec/exprs/vdirect_in_predicate.h"
+#include "vec/exprs/vectorized_fn_call.h"
+#include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
+#include "vec/exprs/vin_predicate.h"
 #include "vec/exprs/vslot_ref.h"
 #include "vec/functions/in.h"
+#include "vec/runtime/vdatetime_value.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/scan/vscan_node.h b/be/src/vec/exec/scan/vscan_node.h
index 996a8ea6d7..294ccf6437 100644
--- a/be/src/vec/exec/scan/vscan_node.h
+++ b/be/src/vec/exec/scan/vscan_node.h
@@ -17,15 +17,55 @@
 
 #pragma once
 
+#include <fmt/format.h>
+#include <gen_cpp/Exprs_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <parallel_hashmap/phmap.h>
+#include <stdint.h>
+
+#include <functional>
+#include <list>
+#include <map>
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include "common/global_types.h"
+#include "common/object_pool.h"
+#include "common/status.h"
 #include "exec/exec_node.h"
 #include "exec/olap_common.h"
 #include "exprs/function_filter.h"
-#include "exprs/runtime_filter.h"
+#include "runtime/define_primitive_type.h"
+#include "runtime/query_fragments_ctx.h"
+#include "runtime/runtime_state.h"
 #include "util/lock.h"
+#include "util/runtime_profile.h"
 #include "vec/exec/scan/scanner_context.h"
-#include "vec/exprs/vectorized_fn_call.h"
-#include "vec/exprs/vexpr.h"
-#include "vec/exprs/vin_predicate.h"
+#include "vec/runtime/shared_scanner_controller.h"
+
+namespace doris {
+class BitmapFilterFuncBase;
+class BloomFilterFuncBase;
+class DescriptorTbl;
+class FunctionContext;
+class HybridSetBase;
+class IRuntimeFilter;
+class SlotDescriptor;
+class TScanRangeParams;
+class TupleDescriptor;
+
+namespace vectorized {
+class Block;
+class VExpr;
+class VExprContext;
+class VInPredicate;
+class VectorizedFnCall;
+} // namespace vectorized
+struct StringRef;
+} // namespace doris
 
 namespace doris::pipeline {
 class ScanOperator;
diff --git a/be/src/vec/exec/scan/vscanner.cpp b/be/src/vec/exec/scan/vscanner.cpp
index 180b573614..e2c9135397 100644
--- a/be/src/vec/exec/scan/vscanner.cpp
+++ b/be/src/vec/exec/scan/vscanner.cpp
@@ -17,7 +17,14 @@
 
 #include "vec/exec/scan/vscanner.h"
 
+#include <glog/logging.h>
+
+#include "common/config.h"
+#include "runtime/descriptors.h"
+#include "util/runtime_profile.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/exec/scan/vscan_node.h"
+#include "vec/exprs/vexpr_context.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/scan/vscanner.h b/be/src/vec/exec/scan/vscanner.h
index 26248d9b54..1733ee3acf 100644
--- a/be/src/vec/exec/scan/vscanner.h
+++ b/be/src/vec/exec/scan/vscanner.h
@@ -17,14 +17,29 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include <algorithm>
+#include <vector>
+
 #include "common/status.h"
 #include "olap/tablet.h"
+#include "runtime/exec_env.h"
 #include "runtime/runtime_state.h"
-#include "vec/exprs/vexpr_context.h"
+#include "util/stopwatch.hpp"
+#include "vec/core/block.h"
+
+namespace doris {
+class RuntimeProfile;
+class TupleDescriptor;
+
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
-class Block;
 class VScanNode;
 
 // Counter for load
diff --git a/be/src/vec/exec/vaggregation_node.cpp b/be/src/vec/exec/vaggregation_node.cpp
index c19a4b7730..76bad7dc43 100644
--- a/be/src/vec/exec/vaggregation_node.cpp
+++ b/be/src/vec/exec/vaggregation_node.cpp
@@ -17,16 +17,39 @@
 
 #include "vec/exec/vaggregation_node.h"
 
+#include <gen_cpp/Exprs_types.h>
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+
+#include <array>
+#include <atomic>
 #include <memory>
 
 #include "exec/exec_node.h"
+#include "runtime/define_primitive_type.h"
+#include "runtime/descriptors.h"
+#include "runtime/memory/mem_tracker.h"
+#include "runtime/runtime_state.h"
+#include "runtime/thread_context.h"
+#include "util/telemetry/telemetry.h"
+#include "vec/common/hash_table/hash_table_key_holder.h"
+#include "vec/common/hash_table/hash_table_utils.h"
+#include "vec/common/hash_table/string_hash_table.h"
+#include "vec/common/string_buffer.hpp"
 #include "vec/core/block.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_string.h"
 #include "vec/exprs/vexpr.h"
 #include "vec/exprs/vexpr_context.h"
 #include "vec/utils/util.hpp"
 
+namespace doris {
+class ObjectPool;
+} // namespace doris
+
 namespace doris::vectorized {
 
 // Here is an empirical value.
diff --git a/be/src/vec/exec/vaggregation_node.h b/be/src/vec/exec/vaggregation_node.h
index 34bcb7efed..aa9b149ef2 100644
--- a/be/src/vec/exec/vaggregation_node.h
+++ b/be/src/vec/exec/vaggregation_node.h
@@ -17,22 +17,61 @@
 
 #pragma once
 
-#include <functional>
-#include <variant>
+#include <assert.h>
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
 
-#include "common/object_pool.h"
+#include <algorithm>
+#include <functional>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <type_traits>
+#include <utility>
+#include <variant>
+#include <vector>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/global_types.h"
+#include "common/status.h"
 #include "exec/exec_node.h"
+#include "util/runtime_profile.h"
 #include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector_helper.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/aggregation_common.h"
+#include "vec/common/allocator.h"
+#include "vec/common/arena.h"
+#include "vec/common/assert_cast.h"
 #include "vec/common/columns_hashing.h"
 #include "vec/common/hash_table/fixed_hash_map.h"
+#include "vec/common/hash_table/fixed_hash_table.h"
+#include "vec/common/hash_table/hash.h"
 #include "vec/common/hash_table/partitioned_hash_map.h"
+#include "vec/common/hash_table/ph_hash_map.h"
 #include "vec/common/hash_table/string_hash_map.h"
+#include "vec/common/pod_array.h"
+#include "vec/common/string_ref.h"
+#include "vec/common/uint128.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
 #include "vec/exprs/vectorized_agg_fn.h"
+#include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
 #include "vec/exprs/vslot_ref.h"
 
 namespace doris {
 class TPlanNode;
 class DescriptorTbl;
+class ObjectPool;
+class RuntimeState;
+class TupleDescriptor;
 
 namespace pipeline {
 class AggSinkOperator;
@@ -42,7 +81,6 @@ class StreamingAggSourceOperator;
 } // namespace pipeline
 
 namespace vectorized {
-class VExprContext;
 
 /** Aggregates by concatenating serialized key values.
   * The serialized value differs in that it uniquely allows to deserialize it, having only the position with which it starts.
diff --git a/be/src/vec/exec/vanalytic_eval_node.cpp b/be/src/vec/exec/vanalytic_eval_node.cpp
index 932baa370d..dba8f6e126 100644
--- a/be/src/vec/exec/vanalytic_eval_node.cpp
+++ b/be/src/vec/exec/vanalytic_eval_node.cpp
@@ -17,9 +17,33 @@
 
 #include "vec/exec/vanalytic_eval_node.h"
 
+#include <gen_cpp/Exprs_types.h>
+#include <gen_cpp/Metrics_types.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+#include <thrift/protocol/TDebugProtocol.h>
+
+#include <algorithm>
+#include <ostream>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/logging.h"
 #include "runtime/descriptors.h"
+#include "runtime/memory/mem_tracker.h"
 #include "runtime/runtime_state.h"
+#include "util/telemetry/telemetry.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/exprs/vectorized_agg_fn.h"
 #include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
+
+namespace doris {
+class ObjectPool;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/vanalytic_eval_node.h b/be/src/vec/exec/vanalytic_eval_node.h
index ca7d723974..642ceb858c 100644
--- a/be/src/vec/exec/vanalytic_eval_node.h
+++ b/be/src/vec/exec/vanalytic_eval_node.h
@@ -17,16 +17,38 @@
 
 #pragma once
 
-#include <thrift/protocol/TDebugProtocol.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+#include <stdint.h>
 
 #include <atomic>
+#include <functional>
+#include <memory>
 #include <string>
+#include <vector>
 
+#include "common/global_types.h"
+#include "common/status.h"
 #include "exec/exec_node.h"
+#include "util/runtime_profile.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/common/arena.h"
 #include "vec/core/block.h"
-#include "vec/exprs/vectorized_agg_fn.h"
-#include "vec/exprs/vexpr_context.h"
+#include "vec/data_types/data_type.h"
+
+namespace doris {
+class DescriptorTbl;
+class ObjectPool;
+class RuntimeState;
+class TupleDescriptor;
+
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 struct BlockRowPos {
@@ -46,6 +68,7 @@ struct BlockRowPos {
 };
 
 class AggFnEvaluator;
+
 class VAnalyticEvalNode : public ExecNode {
 public:
     ~VAnalyticEvalNode() override = default;
diff --git a/be/src/vec/exec/varrow_scanner.cpp b/be/src/vec/exec/varrow_scanner.cpp
index eae0bb0deb..c705f1fd69 100644
--- a/be/src/vec/exec/varrow_scanner.cpp
+++ b/be/src/vec/exec/varrow_scanner.cpp
@@ -17,15 +17,48 @@
 
 #include "vec/exec/varrow_scanner.h"
 
-#include "exec/arrow/parquet_reader.h"
+#include <arrow/array/array_base.h>
+#include <arrow/record_batch.h>
+#include <arrow/type.h>
+#include <fmt/format.h>
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+
+#include <algorithm>
+#include <map>
+#include <sstream>
+#include <string>
+#include <utility>
+
+#include "exec/arrow/arrow_reader.h"
 #include "io/file_factory.h"
-#include "olap/iterators.h"
+#include "io/fs/file_reader.h"
 #include "runtime/descriptors.h"
-#include "runtime/exec_env.h"
+#include "runtime/runtime_state.h"
+#include "runtime/thread_context.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/field.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_factory.hpp"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/data_types/data_type_string.h"
+#include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 #include "vec/utils/arrow_column_to_doris_column.h"
 
+namespace doris {
+class TExpr;
+namespace io {
+enum class FileCachePolicy : uint8_t;
+} // namespace io
+} // namespace doris
+
 namespace doris::vectorized {
 using namespace ErrorCode;
 
diff --git a/be/src/vec/exec/varrow_scanner.h b/be/src/vec/exec/varrow_scanner.h
index fc02c5c2c0..e8d55542d6 100644
--- a/be/src/vec/exec/varrow_scanner.h
+++ b/be/src/vec/exec/varrow_scanner.h
@@ -17,23 +17,39 @@
 
 #pragma once
 
-#include <arrow/array.h>
-#include <exec/arrow/arrow_reader.h>
+#include <stddef.h>
+#include <stdint.h>
 
-#include <map>
 #include <memory>
-#include <sstream>
-#include <string>
-#include <unordered_map>
 #include <vector>
 
 #include "common/status.h"
-#include "exec/arrow/parquet_reader.h"
 #include "exec/base_scanner.h"
-#include "gen_cpp/Types_types.h"
 #include "io/file_factory.h"
+#include "io/fs/file_reader_writer_fwd.h"
 #include "util/runtime_profile.h"
 
+namespace arrow {
+class RecordBatch;
+} // namespace arrow
+namespace doris {
+class ArrowReaderWrap;
+class RuntimeState;
+class SlotDescriptor;
+class TBrokerRangeDesc;
+class TBrokerScanRangeParams;
+class TExpr;
+class TNetworkAddress;
+
+namespace io {
+class FileSystem;
+} // namespace io
+namespace vectorized {
+class Block;
+} // namespace vectorized
+struct Statistics;
+} // namespace doris
+
 namespace doris::vectorized {
 
 // VArrow scanner convert the data read from orc|parquet to doris's columns.
diff --git a/be/src/vec/exec/vassert_num_rows_node.cpp b/be/src/vec/exec/vassert_num_rows_node.cpp
index 9ffe69c22e..4b57ed0a05 100644
--- a/be/src/vec/exec/vassert_num_rows_node.cpp
+++ b/be/src/vec/exec/vassert_num_rows_node.cpp
@@ -17,11 +17,27 @@
 
 #include "vec/exec/vassert_num_rows_node.h"
 
-#include "gen_cpp/PlanNodes_types.h"
+#include <gen_cpp/PlanNodes_types.h>
+#include <glog/logging.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+
+#include <functional>
+#include <map>
+#include <memory>
+#include <ostream>
+#include <utility>
+#include <vector>
+
 #include "runtime/runtime_state.h"
 #include "util/runtime_profile.h"
+#include "util/telemetry/telemetry.h"
 #include "vec/core/block.h"
 
+namespace doris {
+class DescriptorTbl;
+class ObjectPool;
+} // namespace doris
+
 namespace doris::vectorized {
 
 VAssertNumRowsNode::VAssertNumRowsNode(ObjectPool* pool, const TPlanNode& tnode,
diff --git a/be/src/vec/exec/vassert_num_rows_node.h b/be/src/vec/exec/vassert_num_rows_node.h
index 6b0432b6db..90b74b0acb 100644
--- a/be/src/vec/exec/vassert_num_rows_node.h
+++ b/be/src/vec/exec/vassert_num_rows_node.h
@@ -15,9 +15,19 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <gen_cpp/PlanNodes_types.h>
+#include <stdint.h>
+
+#include <string>
+
 #include "common/status.h"
 #include "exec/exec_node.h"
-#include "gen_cpp/PlanNodes_types.h"
+
+namespace doris {
+class DescriptorTbl;
+class ObjectPool;
+class RuntimeState;
+} // namespace doris
 
 namespace doris::vectorized {
 class Block;
diff --git a/be/src/vec/exec/vdata_gen_scan_node.cpp b/be/src/vec/exec/vdata_gen_scan_node.cpp
index 74c2dd814c..9fccc16aaf 100644
--- a/be/src/vec/exec/vdata_gen_scan_node.cpp
+++ b/be/src/vec/exec/vdata_gen_scan_node.cpp
@@ -17,13 +17,25 @@
 
 #include "vec/exec/vdata_gen_scan_node.h"
 
+#include <gen_cpp/PlanNodes_types.h>
+
 #include <sstream>
 
+#include "common/logging.h"
 #include "common/status.h"
-#include "gen_cpp/PlanNodes_types.h"
+#include "exec/exec_node.h"
+#include "runtime/descriptors.h"
 #include "runtime/runtime_state.h"
 #include "util/runtime_profile.h"
+#include "vec/core/block.h"
+#include "vec/exec/data_gen_functions/vdata_gen_function_inf.h"
 #include "vec/exec/data_gen_functions/vnumbers_tvf.h"
+#include "vec/exprs/vexpr_context.h"
+
+namespace doris {
+class ObjectPool;
+class TScanRangeParams;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/vdata_gen_scan_node.h b/be/src/vec/exec/vdata_gen_scan_node.h
index 4c699cc293..27ecb7d519 100644
--- a/be/src/vec/exec/vdata_gen_scan_node.h
+++ b/be/src/vec/exec/vdata_gen_scan_node.h
@@ -18,19 +18,24 @@
 #pragma once
 
 #include <memory>
+#include <vector>
 
+#include "common/global_types.h"
 #include "exec/scan_node.h"
-#include "runtime/descriptors.h"
-#include "vec/exec/data_gen_functions/vdata_gen_function_inf.h"
 
 namespace doris {
 
-class TextConverter;
 class TupleDescriptor;
 class RuntimeState;
 class Status;
+class DescriptorTbl;
+class ObjectPool;
+class TPlanNode;
+class TScanRangeParams;
 
 namespace vectorized {
+class Block;
+class VDataGenFunctionInf;
 
 class VDataGenFunctionScanNode : public ScanNode {
 public:
diff --git a/be/src/vec/exec/vempty_set_node.cpp b/be/src/vec/exec/vempty_set_node.cpp
index fb4970f5bb..1bee0aa5a5 100644
--- a/be/src/vec/exec/vempty_set_node.cpp
+++ b/be/src/vec/exec/vempty_set_node.cpp
@@ -17,8 +17,19 @@
 
 #include "vec/exec/vempty_set_node.h"
 
+#include <opentelemetry/nostd/shared_ptr.h>
+
+#include "runtime/runtime_state.h"
+#include "util/telemetry/telemetry.h"
+
 namespace doris {
+class DescriptorTbl;
+class ObjectPool;
+class TPlanNode;
+
 namespace vectorized {
+class Block;
+
 VEmptySetNode::VEmptySetNode(ObjectPool* pool, const TPlanNode& tnode, const DescriptorTbl& descs)
         : ExecNode(pool, tnode, descs) {}
 
diff --git a/be/src/vec/exec/vempty_set_node.h b/be/src/vec/exec/vempty_set_node.h
index 80b1d2775f..eec107018b 100644
--- a/be/src/vec/exec/vempty_set_node.h
+++ b/be/src/vec/exec/vempty_set_node.h
@@ -17,10 +17,18 @@
 
 #pragma once
 
+#include "common/status.h"
 #include "exec/exec_node.h"
 
 namespace doris {
+class DescriptorTbl;
+class ObjectPool;
+class RuntimeState;
+class TPlanNode;
+
 namespace vectorized {
+class Block;
+
 /// Node that returns an empty result set, i.e., just sets eos_ in GetNext().
 /// Corresponds to EmptySetNode.java in the FE.
 class VEmptySetNode : public ExecNode {
diff --git a/be/src/vec/exec/vexchange_node.cpp b/be/src/vec/exec/vexchange_node.cpp
index 0f4bb58bfa..a00f9cd839 100644
--- a/be/src/vec/exec/vexchange_node.cpp
+++ b/be/src/vec/exec/vexchange_node.cpp
@@ -17,15 +17,20 @@
 
 #include "vec/exec/vexchange_node.h"
 
-#include "common/consts.h"
+#include <gen_cpp/PlanNodes_types.h>
+#include <glog/logging.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+
+#include "common/object_pool.h"
 #include "exec/rowid_fetcher.h"
-#include "pipeline/exec/exchange_source_operator.h"
-#include "pipeline/pipeline.h"
-#include "pipeline/pipeline_fragment_context.h"
+#include "exec/tablet_info.h"
 #include "runtime/exec_env.h"
+#include "runtime/query_statistics.h"
 #include "runtime/runtime_state.h"
-#include "runtime/thread_context.h"
-#include "util/defer_op.h"
+#include "util/runtime_profile.h"
+#include "util/telemetry/telemetry.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/runtime/vdata_stream_mgr.h"
 #include "vec/runtime/vdata_stream_recvr.h"
 
diff --git a/be/src/vec/exec/vexchange_node.h b/be/src/vec/exec/vexchange_node.h
index 1d2738fda0..46f5aa619c 100644
--- a/be/src/vec/exec/vexchange_node.h
+++ b/be/src/vec/exec/vexchange_node.h
@@ -17,19 +17,31 @@
 
 #pragma once
 
-#include <memory>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <memory>
+#include <vector>
+
+#include "common/status.h"
 #include "exec/exec_node.h"
-#include "exec/tablet_info.h" // DorisNodesInfo
 #include "runtime/descriptors.h"
 #include "vec/common/sort/vsort_exec_exprs.h"
 
 namespace doris {
+class DorisNodesInfo;
+class ObjectPool;
+class QueryStatistics;
+class QueryStatisticsRecvr;
+class RuntimeState;
+class TPlanNode;
+
 namespace pipeline {
 class ExchangeSourceOperator;
 }
 namespace vectorized {
 class VDataStreamRecvr;
+class Block;
 
 class VExchangeNode : public ExecNode {
 public:
diff --git a/be/src/vec/exec/vjdbc_connector.cpp b/be/src/vec/exec/vjdbc_connector.cpp
index 413a5a9679..d544227f02 100644
--- a/be/src/vec/exec/vjdbc_connector.cpp
+++ b/be/src/vec/exec/vjdbc_connector.cpp
@@ -17,24 +17,46 @@
 
 #include "vec/exec/vjdbc_connector.h"
 
-#include <cstring>
+#include <gen_cpp/Types_types.h>
 
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+// IWYU pragma: no_include <bits/std_abs.h>
+#include <cmath> // IWYU pragma: keep
+#include <cstring>
+#include <memory>
+#include <ostream>
+#include <utility>
+
+#include "common/logging.h"
 #include "common/status.h"
 #include "exec/table_connector.h"
-#include "gen_cpp/Types_types.h"
 #include "gutil/strings/substitute.h"
 #include "jni.h"
+#include "jni_md.h"
 #include "runtime/define_primitive_type.h"
+#include "runtime/descriptors.h"
+#include "runtime/runtime_state.h"
+#include "runtime/types.h"
 #include "runtime/user_function_cache.h"
 #include "util/jni-util.h"
 #include "util/runtime_profile.h"
-#include "vec/columns/column_array.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_nullable.h"
 #include "vec/columns/column_string.h"
-#include "vec/data_types/data_type_factory.hpp"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/data_types/data_type_number.h"
 #include "vec/data_types/data_type_string.h"
 #include "vec/exec/jni_connector.h"
-#include "vec/exec/scan/new_jdbc_scanner.h"
+#include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
+#include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 
 namespace doris {
diff --git a/be/src/vec/exec/vjdbc_connector.h b/be/src/vec/exec/vjdbc_connector.h
index 3af73bf109..d113cceb6e 100644
--- a/be/src/vec/exec/vjdbc_connector.h
+++ b/be/src/vec/exec/vjdbc_connector.h
@@ -17,19 +17,30 @@
 
 #pragma once
 
+#include <fmt/format.h>
+#include <gen_cpp/Types_types.h>
 #include <jni.h>
+#include <stdint.h>
 
-#include <string_view>
+#include <map>
+#include <string>
+#include <vector>
 
 #include "common/status.h"
 #include "exec/table_connector.h"
-#include "runtime/define_primitive_type.h"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/data_types/data_type.h"
 
 namespace doris {
+class RuntimeState;
+class SlotDescriptor;
+class TupleDescriptor;
+
 namespace vectorized {
 
-class NewJdbcScanner;
+class Block;
+class IColumn;
+class VExprContext;
 
 struct JdbcConnectorParam {
     std::string driver_path;
diff --git a/be/src/vec/exec/vmysql_scan_node.cpp b/be/src/vec/exec/vmysql_scan_node.cpp
index c0428b5e51..36251e24d7 100644
--- a/be/src/vec/exec/vmysql_scan_node.cpp
+++ b/be/src/vec/exec/vmysql_scan_node.cpp
@@ -17,9 +17,10 @@
 
 #include "vec/exec/vmysql_scan_node.h"
 
+#include <gen_cpp/PlanNodes_types.h>
+
 #include "exec/text_converter.h"
 #include "exec/text_converter.hpp"
-#include "gen_cpp/PlanNodes_types.h"
 #include "runtime/runtime_state.h"
 #include "util/runtime_profile.h"
 #include "util/types.h"
diff --git a/be/src/vec/exec/vparquet_scanner.cpp b/be/src/vec/exec/vparquet_scanner.cpp
index 85886e28c1..70d8bd1408 100644
--- a/be/src/vec/exec/vparquet_scanner.cpp
+++ b/be/src/vec/exec/vparquet_scanner.cpp
@@ -18,6 +18,19 @@
 #include "vec/exec/vparquet_scanner.h"
 
 #include "exec/arrow/parquet_reader.h"
+#include "vec/exec/varrow_scanner.h"
+
+namespace doris {
+class ArrowReaderWrap;
+class RuntimeProfile;
+class RuntimeState;
+class SlotDescriptor;
+class TBrokerRangeDesc;
+class TBrokerScanRangeParams;
+class TExpr;
+class TNetworkAddress;
+struct ScannerCounter;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/vparquet_scanner.h b/be/src/vec/exec/vparquet_scanner.h
index 3ba440c9ef..dc71f43b9e 100644
--- a/be/src/vec/exec/vparquet_scanner.h
+++ b/be/src/vec/exec/vparquet_scanner.h
@@ -17,20 +17,24 @@
 
 #pragma once
 
-#include <arrow/array.h>
-#include <vec/exec/varrow_scanner.h>
+#include <stdint.h>
 
-#include <map>
-#include <memory>
-#include <sstream>
-#include <string>
-#include <unordered_map>
 #include <vector>
 
-#include "common/status.h"
-#include "gen_cpp/PlanNodes_types.h"
-#include "gen_cpp/Types_types.h"
-#include "util/runtime_profile.h"
+#include "io/fs/file_reader_writer_fwd.h"
+#include "vec/exec/varrow_scanner.h"
+
+namespace doris {
+class ArrowReaderWrap;
+class RuntimeProfile;
+class RuntimeState;
+class SlotDescriptor;
+class TBrokerRangeDesc;
+class TBrokerScanRangeParams;
+class TExpr;
+class TNetworkAddress;
+struct ScannerCounter;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/vrepeat_node.cpp b/be/src/vec/exec/vrepeat_node.cpp
index c3fe0e8d0f..5bfa907ccc 100644
--- a/be/src/vec/exec/vrepeat_node.cpp
+++ b/be/src/vec/exec/vrepeat_node.cpp
@@ -17,10 +17,32 @@
 
 #include "vec/exec/vrepeat_node.h"
 
+#include <gen_cpp/PlanNodes_types.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+#include <string.h>
+
+#include <functional>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/logging.h"
+#include "common/status.h"
 #include "gutil/strings/join.h"
+#include "runtime/descriptors.h"
 #include "runtime/runtime_state.h"
+#include "runtime/types.h"
 #include "util/runtime_profile.h"
+#include "util/telemetry/telemetry.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
 
 namespace doris::vectorized {
 VRepeatNode::VRepeatNode(ObjectPool* pool, const TPlanNode& tnode, const DescriptorTbl& descs)
diff --git a/be/src/vec/exec/vrepeat_node.h b/be/src/vec/exec/vrepeat_node.h
index 5f038c606e..4a2c242e6c 100644
--- a/be/src/vec/exec/vrepeat_node.h
+++ b/be/src/vec/exec/vrepeat_node.h
@@ -17,7 +17,16 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include <iosfwd>
+#include <memory>
+#include <set>
+#include <vector>
+
+#include "common/global_types.h"
 #include "exec/exec_node.h"
+#include "vec/core/block.h"
 
 namespace doris {
 
@@ -26,6 +35,8 @@ class TPlanNode;
 class DescriptorTbl;
 class RuntimeState;
 class Status;
+class SlotDescriptor;
+class TupleDescriptor;
 
 namespace vectorized {
 class VExprContext;
diff --git a/be/src/vec/exec/vschema_scan_node.cpp b/be/src/vec/exec/vschema_scan_node.cpp
index 65dcd8f17e..f7d98dc6ce 100644
--- a/be/src/vec/exec/vschema_scan_node.cpp
+++ b/be/src/vec/exec/vschema_scan_node.cpp
@@ -17,21 +17,37 @@
 
 #include "vec/exec/vschema_scan_node.h"
 
-#include <arrow/type.h>
-#include <arrow/type_fwd.h>
+#include <gen_cpp/FrontendService_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Types_types.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+#include <opentelemetry/trace/span.h>
+#include <opentelemetry/trace/span_metadata.h>
 
+#include <boost/algorithm/string/predicate.hpp>
+#include <ostream>
+#include <utility>
+
+#include "common/logging.h"
+#include "common/object_pool.h"
 #include "common/status.h"
-#include "exec/text_converter.h"
-#include "exec/text_converter.hpp"
-#include "gen_cpp/PlanNodes_types.h"
+#include "exec/exec_node.h"
 #include "runtime/descriptors.h"
 #include "runtime/runtime_state.h"
+#include "runtime/types.h"
 #include "util/runtime_profile.h"
-#include "util/types.h"
+#include "util/telemetry/telemetry.h"
 #include "vec/columns/column.h"
-#include "vec/common/string_ref.h"
-#include "vec/core/types.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_factory.hpp"
+#include "vec/exprs/vexpr_context.h"
+
+namespace doris {
+class TScanRangeParams;
+} // namespace doris
+
 namespace doris::vectorized {
 
 VSchemaScanNode::VSchemaScanNode(ObjectPool* pool, const TPlanNode& tnode,
diff --git a/be/src/vec/exec/vschema_scan_node.h b/be/src/vec/exec/vschema_scan_node.h
index 10db793d8a..5e0784dd2b 100644
--- a/be/src/vec/exec/vschema_scan_node.h
+++ b/be/src/vec/exec/vschema_scan_node.h
@@ -17,20 +17,27 @@
 
 #pragma once
 
+#include <iosfwd>
 #include <memory>
+#include <string>
+#include <vector>
 
+#include "common/global_types.h"
 #include "exec/scan_node.h"
 #include "exec/schema_scanner.h"
-#include "gen_cpp/Descriptors_types.h"
-#include "runtime/descriptors.h"
+
 namespace doris {
 
-class TextConverter;
 class TupleDescriptor;
 class RuntimeState;
 class Status;
+class DescriptorTbl;
+class ObjectPool;
+class TPlanNode;
+class TScanRangeParams;
 
 namespace vectorized {
+class Block;
 
 class VSchemaScanNode : public ScanNode {
 public:
diff --git a/be/src/vec/exec/vselect_node.cpp b/be/src/vec/exec/vselect_node.cpp
index cd8924abb4..accc99e94c 100644
--- a/be/src/vec/exec/vselect_node.cpp
+++ b/be/src/vec/exec/vselect_node.cpp
@@ -17,7 +17,23 @@
 
 #include "vec/exec/vselect_node.h"
 
+#include <opentelemetry/nostd/shared_ptr.h>
+
+#include <functional>
+#include <memory>
+#include <vector>
+
+#include "runtime/runtime_state.h"
+#include "util/runtime_profile.h"
+#include "util/telemetry/telemetry.h"
+#include "vec/core/block.h"
+#include "vec/exprs/vexpr_context.h"
+
 namespace doris {
+class DescriptorTbl;
+class ObjectPool;
+class TPlanNode;
+
 namespace vectorized {
 
 VSelectNode::VSelectNode(ObjectPool* pool, const TPlanNode& tnode, const DescriptorTbl& descs)
diff --git a/be/src/vec/exec/vselect_node.h b/be/src/vec/exec/vselect_node.h
index 984e4578fb..d6783d7237 100644
--- a/be/src/vec/exec/vselect_node.h
+++ b/be/src/vec/exec/vselect_node.h
@@ -16,10 +16,17 @@
 // under the License.
 
 #pragma once
+#include "common/status.h"
 #include "exec/exec_node.h"
 
 namespace doris {
+class DescriptorTbl;
+class ObjectPool;
+class RuntimeState;
+class TPlanNode;
+
 namespace vectorized {
+class Block;
 
 class VSelectNode final : public ExecNode {
 public:
diff --git a/be/src/vec/exec/vset_operation_node.cpp b/be/src/vec/exec/vset_operation_node.cpp
index 5665fcc5f4..c60581276a 100644
--- a/be/src/vec/exec/vset_operation_node.cpp
+++ b/be/src/vec/exec/vset_operation_node.cpp
@@ -17,9 +17,40 @@
 
 #include "vec/exec/vset_operation_node.h"
 
+#include <fmt/format.h>
+#include <gen_cpp/Exprs_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <glog/logging.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+
+#include <array>
+#include <atomic>
+#include <ostream>
+#include <string>
+#include <type_traits>
+#include <utility>
+
+#include "runtime/define_primitive_type.h"
+#include "runtime/runtime_state.h"
 #include "util/defer_op.h"
+#include "util/telemetry/telemetry.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/common/aggregation_common.h"
+#include "vec/common/columns_hashing.h"
+#include "vec/common/uint128.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/materialize_block.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/exec/join/join_op.h"
 #include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
+
 namespace doris {
+class DescriptorTbl;
+class ObjectPool;
+
 namespace vectorized {
 
 //build hash table for operation node, intersect/except node
diff --git a/be/src/vec/exec/vset_operation_node.h b/be/src/vec/exec/vset_operation_node.h
index ad9cbab7e7..b833658c7d 100644
--- a/be/src/vec/exec/vset_operation_node.h
+++ b/be/src/vec/exec/vset_operation_node.h
@@ -17,18 +17,37 @@
 
 #pragma once
 
+#include <stddef.h>
+#include <stdint.h>
+
+#include <functional>
+#include <iosfwd>
+#include <memory>
+#include <unordered_map>
+#include <variant>
+#include <vector>
+
+#include "common/status.h"
 #include "exec/exec_node.h"
-#include "runtime/runtime_state.h"
-#include "vec/core/materialize_block.h"
-#include "vec/exec/join/join_op.h"
-#include "vec/exec/join/vacquire_list.hpp"
+#include "util/runtime_profile.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/common/arena.h"
+#include "vec/common/hash_table/hash_map.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/exec/join/process_hash_table_probe.h"
 #include "vec/exec/join/vhash_join_node.h"
-#include "vec/functions/function.h"
-#include "vec/utils/util.hpp"
 
 namespace doris {
+class DescriptorTbl;
+class ObjectPool;
+class RuntimeState;
+class TPlanNode;
 
 namespace vectorized {
+class VExprContext;
+struct RowRefListWithFlags;
 
 template <bool is_intersect>
 class VSetOperationNode final : public ExecNode {
diff --git a/be/src/vec/exec/vsort_node.cpp b/be/src/vec/exec/vsort_node.cpp
index f5111ad0af..5901db2ce6 100644
--- a/be/src/vec/exec/vsort_node.cpp
+++ b/be/src/vec/exec/vsort_node.cpp
@@ -17,13 +17,33 @@
 
 #include "vec/exec/vsort_node.h"
 
-#include "common/config.h"
-#include "pipeline/pipeline.h"
+#include <gen_cpp/Exprs_types.h>
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+
+#include <atomic>
+#include <functional>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "runtime/descriptors.h"
+#include "runtime/memory/mem_tracker.h"
+#include "runtime/query_fragments_ctx.h"
+#include "runtime/runtime_predicate.h"
 #include "runtime/runtime_state.h"
-#include "util/debug_util.h"
+#include "runtime/types.h"
+#include "util/telemetry/telemetry.h"
 #include "vec/common/sort/heap_sorter.h"
 #include "vec/common/sort/topn_sorter.h"
-#include "vec/core/sort_block.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/sort_description.h"
+
+namespace doris {
+class ObjectPool;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/vsort_node.h b/be/src/vec/exec/vsort_node.h
index 1ab3b4e5cc..2196e5e318 100644
--- a/be/src/vec/exec/vsort_node.h
+++ b/be/src/vec/exec/vsort_node.h
@@ -17,13 +17,30 @@
 
 #pragma once
 
-#include <queue>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <iosfwd>
+#include <memory>
+#include <vector>
+
+#include "common/status.h"
 #include "exec/exec_node.h"
+#include "util/runtime_profile.h"
 #include "vec/common/sort/sorter.h"
 #include "vec/common/sort/vsort_exec_exprs.h"
-#include "vec/core/block.h"
-#include "vec/core/sort_cursor.h"
+#include "vec/core/field.h"
+
+namespace doris {
+class DescriptorTbl;
+class ObjectPool;
+class RuntimeState;
+class TPlanNode;
+
+namespace vectorized {
+class Block;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 // Node that implements a full sort of its input with a fixed memory budget
diff --git a/be/src/vec/exec/vtable_function_node.cpp b/be/src/vec/exec/vtable_function_node.cpp
index 39ffb4fad7..9f3fbccb28 100644
--- a/be/src/vec/exec/vtable_function_node.cpp
+++ b/be/src/vec/exec/vtable_function_node.cpp
@@ -17,9 +17,26 @@
 
 #include "vec/exec/vtable_function_node.h"
 
+#include <gen_cpp/Exprs_types.h>
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <functional>
+#include <memory>
+#include <string>
+#include <utility>
+
 #include "vec/exprs/table_function/table_function.h"
 #include "vec/exprs/table_function/table_function_factory.h"
 #include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
+
+namespace doris {
+class ObjectPool;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exec/vtable_function_node.h b/be/src/vec/exec/vtable_function_node.h
index 2aad138f4a..acfd3e4e72 100644
--- a/be/src/vec/exec/vtable_function_node.h
+++ b/be/src/vec/exec/vtable_function_node.h
@@ -17,10 +17,34 @@
 
 #pragma once
 
+#include <opentelemetry/nostd/shared_ptr.h>
+#include <stdint.h>
+
+#include <vector>
+
+#include "common/global_types.h"
+#include "common/status.h"
 #include "exec/exec_node.h"
+#include "runtime/descriptors.h"
+#include "runtime/runtime_state.h"
+#include "util/runtime_profile.h"
+#include "util/telemetry/telemetry.h"
+#include "vec/columns/column.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/data_types/data_type.h"
 #include "vec/exprs/table_function/table_function.h"
 #include "vec/exprs/vexpr.h"
 
+namespace doris {
+class ObjectPool;
+class TPlanNode;
+
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 class VTableFunctionNode final : public ExecNode {
diff --git a/be/src/vec/exec/vunion_node.cpp b/be/src/vec/exec/vunion_node.cpp
index 67866b09c5..94bb800d9f 100644
--- a/be/src/vec/exec/vunion_node.cpp
+++ b/be/src/vec/exec/vunion_node.cpp
@@ -17,19 +17,34 @@
 
 #include "vec/exec/vunion_node.h"
 
-#include <gen_cpp/AgentService_types.h>
+#include <gen_cpp/Exprs_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <functional>
+#include <memory>
+#include <ostream>
+#include <string>
 
 #include "common/status.h"
-#include "gen_cpp/PlanNodes_types.h"
 #include "runtime/runtime_state.h"
 #include "util/runtime_profile.h"
+#include "util/telemetry/telemetry.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
 #include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/exprs/vexpr.h"
 #include "vec/exprs/vexpr_context.h"
 #include "vec/utils/util.hpp"
 
 namespace doris {
+class DescriptorTbl;
+class ObjectPool;
 
 namespace vectorized {
 
diff --git a/be/src/vec/exec/vunion_node.h b/be/src/vec/exec/vunion_node.h
index fa517df2ff..7532c00875 100644
--- a/be/src/vec/exec/vunion_node.h
+++ b/be/src/vec/exec/vunion_node.h
@@ -17,10 +17,25 @@
 
 #pragma once
 
-#include "vec/exec/vset_operation_node.h"
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <iosfwd>
+#include <vector>
+
+#include "common/status.h"
+#include "exec/exec_node.h"
+#include "runtime/runtime_state.h"
+#include "util/runtime_profile.h"
+#include "vec/core/block.h"
 
 namespace doris {
+class DescriptorTbl;
+class ObjectPool;
+class TPlanNode;
+
 namespace vectorized {
+class VExprContext;
 
 class VUnionNode final : public ExecNode {
 public:
diff --git a/be/src/vec/exprs/lambda_function/varray_filter_function.cpp b/be/src/vec/exprs/lambda_function/varray_filter_function.cpp
index 8d4acdf8f8..0e76522e41 100644
--- a/be/src/vec/exprs/lambda_function/varray_filter_function.cpp
+++ b/be/src/vec/exprs/lambda_function/varray_filter_function.cpp
@@ -15,20 +15,36 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <fmt/core.h>
+#include <glog/logging.h>
+
+#include <memory>
+#include <string>
+#include <utility>
+#include <vector>
 
 #include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
 #include "vec/core/block.h"
-#include "vec/data_types/data_type_array.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/data_types/data_type.h"
 #include "vec/exprs/lambda_function/lambda_function.h"
 #include "vec/exprs/lambda_function/lambda_function_factory.h"
 #include "vec/exprs/vexpr.h"
-#include "vec/exprs/vexpr_context.h"
 #include "vec/utils/util.hpp"
 
+namespace doris {
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 class ArrayFilterFunction : public LambdaFunction {
diff --git a/be/src/vec/exprs/lambda_function/varray_map_function.cpp b/be/src/vec/exprs/lambda_function/varray_map_function.cpp
index 7b14be2260..2c4fc977ae 100644
--- a/be/src/vec/exprs/lambda_function/varray_map_function.cpp
+++ b/be/src/vec/exprs/lambda_function/varray_map_function.cpp
@@ -15,18 +15,36 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <fmt/core.h>
+#include <memory>
+#include <string>
+#include <utility>
+#include <vector>
 
 #include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
 #include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_array.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/exprs/lambda_function/lambda_function.h"
 #include "vec/exprs/lambda_function/lambda_function_factory.h"
 #include "vec/exprs/vexpr.h"
-#include "vec/exprs/vexpr_context.h"
 #include "vec/utils/util.hpp"
 
+namespace doris {
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 class ArrayMapFunction : public LambdaFunction {
diff --git a/be/src/vec/exprs/table_function/table_function_factory.cpp b/be/src/vec/exprs/table_function/table_function_factory.cpp
index c09a3b365a..8cdeafba22 100644
--- a/be/src/vec/exprs/table_function/table_function_factory.cpp
+++ b/be/src/vec/exprs/table_function/table_function_factory.cpp
@@ -17,7 +17,15 @@
 
 #include "vec/exprs/table_function/table_function_factory.h"
 
+#include <utility>
+
 #include "common/object_pool.h"
+#include "vec/exprs/table_function/table_function.h"
+#include "vec/exprs/table_function/vexplode.h"
+#include "vec/exprs/table_function/vexplode_bitmap.h"
+#include "vec/exprs/table_function/vexplode_json_array.h"
+#include "vec/exprs/table_function/vexplode_numbers.h"
+#include "vec/exprs/table_function/vexplode_split.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exprs/table_function/table_function_factory.h b/be/src/vec/exprs/table_function/table_function_factory.h
index d299af0a9c..4bbc0e7939 100644
--- a/be/src/vec/exprs/table_function/table_function_factory.h
+++ b/be/src/vec/exprs/table_function/table_function_factory.h
@@ -18,22 +18,17 @@
 #pragma once
 
 #include <functional>
+#include <string>
 #include <unordered_map>
 
-#include "common/object_pool.h"
 #include "common/status.h"
-#include "vec/exprs/table_function/table_function.h"
-#include "vec/exprs/table_function/vexplode.h"
-#include "vec/exprs/table_function/vexplode_bitmap.h"
-#include "vec/exprs/table_function/vexplode_json_array.h"
-#include "vec/exprs/table_function/vexplode_numbers.h"
-#include "vec/exprs/table_function/vexplode_split.h"
 
 namespace doris {
 class ObjectPool;
 
 namespace vectorized {
 class TableFunction;
+
 class TableFunctionFactory {
 public:
     TableFunctionFactory() = delete;
diff --git a/be/src/vec/exprs/table_function/vexplode.cpp b/be/src/vec/exprs/table_function/vexplode.cpp
index b2c93196a8..8c17ceeb25 100644
--- a/be/src/vec/exprs/table_function/vexplode.cpp
+++ b/be/src/vec/exprs/table_function/vexplode.cpp
@@ -17,8 +17,18 @@
 
 #include "vec/exprs/table_function/vexplode.h"
 
+#include <glog/logging.h>
+
+#include <ostream>
+#include <vector>
+
 #include "common/status.h"
+#include "vec/columns/column.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exprs/table_function/vexplode.h b/be/src/vec/exprs/table_function/vexplode.h
index 8911c8f225..9c540a186c 100644
--- a/be/src/vec/exprs/table_function/vexplode.h
+++ b/be/src/vec/exprs/table_function/vexplode.h
@@ -17,13 +17,19 @@
 
 #pragma once
 
-#include "vec/columns/column.h"
-#include "vec/columns/column_array.h"
-#include "vec/columns/column_nullable.h"
-#include "vec/common/string_ref.h"
+#include <stddef.h>
+
+#include "common/status.h"
+#include "vec/data_types/data_type.h"
 #include "vec/exprs/table_function/table_function.h"
 #include "vec/functions/array/function_array_utils.h"
 
+namespace doris {
+namespace vectorized {
+class Block;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 class VExplodeTableFunction : public TableFunction {
diff --git a/be/src/vec/exprs/table_function/vexplode_bitmap.cpp b/be/src/vec/exprs/table_function/vexplode_bitmap.cpp
index 8dabbac8c3..7ffbb346f8 100644
--- a/be/src/vec/exprs/table_function/vexplode_bitmap.cpp
+++ b/be/src/vec/exprs/table_function/vexplode_bitmap.cpp
@@ -17,11 +17,22 @@
 
 #include "vec/exprs/table_function/vexplode_bitmap.h"
 
+#include <glog/logging.h>
+
+#include <ostream>
+#include <vector>
+
 #include "common/status.h"
 #include "util/bitmap_value.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
 #include "vec/columns/columns_number.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/exprs/table_function/table_function.h"
 #include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exprs/table_function/vexplode_bitmap.h b/be/src/vec/exprs/table_function/vexplode_bitmap.h
index dd47261c38..1b459ad6bf 100644
--- a/be/src/vec/exprs/table_function/vexplode_bitmap.h
+++ b/be/src/vec/exprs/table_function/vexplode_bitmap.h
@@ -17,10 +17,21 @@
 
 #pragma once
 
+#include <stddef.h>
+
+#include <memory>
+
+#include "common/status.h"
 #include "util/bitmap_value.h"
-#include "vec/columns/column.h"
+#include "vec/data_types/data_type.h"
 #include "vec/exprs/table_function/table_function.h"
 
+namespace doris {
+namespace vectorized {
+class Block;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 class VExplodeBitmapTableFunction final : public TableFunction {
diff --git a/be/src/vec/exprs/table_function/vexplode_json_array.cpp b/be/src/vec/exprs/table_function/vexplode_json_array.cpp
index b67467ea98..96e6b10065 100644
--- a/be/src/vec/exprs/table_function/vexplode_json_array.cpp
+++ b/be/src/vec/exprs/table_function/vexplode_json_array.cpp
@@ -17,8 +17,18 @@
 
 #include "vec/exprs/table_function/vexplode_json_array.h"
 
+#include <inttypes.h>
+#include <rapidjson/rapidjson.h>
+#include <stdio.h>
+
+#include <algorithm>
+
 #include "common/status.h"
+#include "vec/columns/column.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exprs/table_function/vexplode_json_array.h b/be/src/vec/exprs/table_function/vexplode_json_array.h
index 9d7de170ff..c32aaed29e 100644
--- a/be/src/vec/exprs/table_function/vexplode_json_array.h
+++ b/be/src/vec/exprs/table_function/vexplode_json_array.h
@@ -17,13 +17,27 @@
 
 #pragma once
 
+#include <glog/logging.h>
 #include <rapidjson/document.h>
-#include <rapidjson/stringbuffer.h>
+#include <stddef.h>
+#include <stdint.h>
 
-#include "vec/columns/column.h"
+#include <ostream>
+#include <string>
+#include <vector>
+
+#include "common/status.h"
+#include "gutil/integral_types.h"
 #include "vec/common/string_ref.h"
+#include "vec/data_types/data_type.h"
 #include "vec/exprs/table_function/table_function.h"
 
+namespace doris {
+namespace vectorized {
+class Block;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 enum ExplodeJsonArrayType { INT = 0, DOUBLE, STRING };
diff --git a/be/src/vec/exprs/table_function/vexplode_numbers.cpp b/be/src/vec/exprs/table_function/vexplode_numbers.cpp
index d1a5eb07dc..50bc6cca7d 100644
--- a/be/src/vec/exprs/table_function/vexplode_numbers.cpp
+++ b/be/src/vec/exprs/table_function/vexplode_numbers.cpp
@@ -17,11 +17,22 @@
 
 #include "vec/exprs/table_function/vexplode_numbers.h"
 
+#include <glog/logging.h>
+
+#include <ostream>
+#include <vector>
+
 #include "common/status.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
 #include "vec/columns/column_nullable.h"
 #include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exprs/table_function/vexplode_numbers.h b/be/src/vec/exprs/table_function/vexplode_numbers.h
index 3e93471b47..b374e00750 100644
--- a/be/src/vec/exprs/table_function/vexplode_numbers.h
+++ b/be/src/vec/exprs/table_function/vexplode_numbers.h
@@ -17,9 +17,23 @@
 
 #pragma once
 
-#include "vec/columns/column.h"
+#include <stddef.h>
+
+#include <algorithm>
+
+#include "common/status.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/data_types/data_type.h"
 #include "vec/exprs/table_function/table_function.h"
 
+namespace doris {
+namespace vectorized {
+class Block;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 class VExplodeNumbersTableFunction : public TableFunction {
diff --git a/be/src/vec/exprs/table_function/vexplode_split.cpp b/be/src/vec/exprs/table_function/vexplode_split.cpp
index d2438d8c9a..8d21a76a05 100644
--- a/be/src/vec/exprs/table_function/vexplode_split.cpp
+++ b/be/src/vec/exprs/table_function/vexplode_split.cpp
@@ -17,10 +17,21 @@
 
 #include "vec/exprs/table_function/vexplode_split.h"
 
+#include <glog/logging.h>
+
+#include <algorithm>
+#include <iterator>
+#include <ostream>
+
 #include "common/status.h"
-#include "gutil/strings/split.h"
 #include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exprs/table_function/vexplode_split.h b/be/src/vec/exprs/table_function/vexplode_split.h
index 5881ef9b6e..901af08616 100644
--- a/be/src/vec/exprs/table_function/vexplode_split.h
+++ b/be/src/vec/exprs/table_function/vexplode_split.h
@@ -17,11 +17,24 @@
 
 #pragma once
 
-#include "gutil/strings/stringpiece.h"
-#include "vec/columns/column.h"
+#include <stddef.h>
+#include <stdint.h>
+
+#include <string_view>
+#include <vector>
+
+#include "common/status.h"
 #include "vec/common/string_ref.h"
+#include "vec/data_types/data_type.h"
 #include "vec/exprs/table_function/table_function.h"
 
+namespace doris {
+namespace vectorized {
+class Block;
+class ColumnString;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 class VExplodeSplitTableFunction final : public TableFunction {
diff --git a/be/src/vec/exprs/varray_literal.cpp b/be/src/vec/exprs/varray_literal.cpp
index e986e927a4..e3f883c1ef 100644
--- a/be/src/vec/exprs/varray_literal.cpp
+++ b/be/src/vec/exprs/varray_literal.cpp
@@ -17,6 +17,27 @@
 
 #include "vec/exprs/varray_literal.h"
 
+#include <gen_cpp/Exprs_types.h>
+#include <glog/logging.h>
+
+#include <memory>
+#include <ostream>
+#include <vector>
+
+#include "runtime/types.h"
+#include "vec/columns/column.h"
+#include "vec/core/field.h"
+#include "vec/data_types/data_type.h"
+#include "vec/exprs/vexpr.h"
+
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 Status VArrayLiteral::prepare(RuntimeState* state, const RowDescriptor& row_desc,
diff --git a/be/src/vec/exprs/varray_literal.h b/be/src/vec/exprs/varray_literal.h
index 56b5f8e331..ff3cedc852 100644
--- a/be/src/vec/exprs/varray_literal.h
+++ b/be/src/vec/exprs/varray_literal.h
@@ -17,11 +17,17 @@
 
 #pragma once
 
+#include "common/status.h"
 #include "vec/exprs/vliteral.h"
 
 namespace doris {
+class RowDescriptor;
+class RuntimeState;
+class TExprNode;
 
 namespace vectorized {
+class VExprContext;
+
 class VArrayLiteral : public VLiteral {
 public:
     VArrayLiteral(const TExprNode& node) : VLiteral(node, false) {}
diff --git a/be/src/vec/exprs/vbitmap_predicate.cpp b/be/src/vec/exprs/vbitmap_predicate.cpp
index 2599ee925c..81e08022d0 100644
--- a/be/src/vec/exprs/vbitmap_predicate.cpp
+++ b/be/src/vec/exprs/vbitmap_predicate.cpp
@@ -17,7 +17,34 @@
 
 #include "vec/exprs/vbitmap_predicate.h"
 
+#include <stddef.h>
+
+#include <algorithm>
+#include <utility>
+#include <vector>
+
+#include "exprs/bitmapfilter_predicate.h"
+#include "gutil/integral_types.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
 #include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+class TExprNode;
+
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exprs/vbitmap_predicate.h b/be/src/vec/exprs/vbitmap_predicate.h
index 629a240218..de1783871d 100644
--- a/be/src/vec/exprs/vbitmap_predicate.h
+++ b/be/src/vec/exprs/vbitmap_predicate.h
@@ -17,8 +17,27 @@
 
 #pragma once
 
+#include <fmt/format.h>
+
+#include <memory>
+#include <string>
+
+#include "common/object_pool.h"
+#include "common/status.h"
+#include "udf/udf.h"
 #include "vec/exprs/vexpr.h"
 
+namespace doris {
+class BitmapFilterFuncBase;
+class RowDescriptor;
+class RuntimeState;
+class TExprNode;
+namespace vectorized {
+class Block;
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 // used for bitmap runtime filter
diff --git a/be/src/vec/exprs/vbloom_predicate.cpp b/be/src/vec/exprs/vbloom_predicate.cpp
index 9ca49a9c7e..1722f96231 100644
--- a/be/src/vec/exprs/vbloom_predicate.cpp
+++ b/be/src/vec/exprs/vbloom_predicate.cpp
@@ -17,10 +17,35 @@
 
 #include "vec/exprs/vbloom_predicate.h"
 
+#include <stddef.h>
+
+#include <utility>
+#include <vector>
+
 #include "common/status.h"
 #include "exprs/bloom_filter_func.h"
+#include "gutil/integral_types.h"
+#include "runtime/runtime_state.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_nullable.h"
 
+namespace doris {
+class RowDescriptor;
+class TExprNode;
+
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 VBloomPredicate::VBloomPredicate(const TExprNode& node)
diff --git a/be/src/vec/exprs/vbloom_predicate.h b/be/src/vec/exprs/vbloom_predicate.h
index 51835a19c0..3226b4db9a 100644
--- a/be/src/vec/exprs/vbloom_predicate.h
+++ b/be/src/vec/exprs/vbloom_predicate.h
@@ -17,8 +17,25 @@
 
 #pragma once
 
+#include <memory>
+#include <string>
+
+#include "common/object_pool.h"
+#include "common/status.h"
+#include "udf/udf.h"
 #include "vec/exprs/vexpr.h"
 
+namespace doris {
+class BloomFilterFuncBase;
+class RowDescriptor;
+class RuntimeState;
+class TExprNode;
+namespace vectorized {
+class Block;
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 class VBloomPredicate final : public VExpr {
 public:
diff --git a/be/src/vec/exprs/vcase_expr.cpp b/be/src/vec/exprs/vcase_expr.cpp
index 955964d1c2..4a6b013185 100644
--- a/be/src/vec/exprs/vcase_expr.cpp
+++ b/be/src/vec/exprs/vcase_expr.cpp
@@ -17,8 +17,28 @@
 
 #include "vec/exprs/vcase_expr.h"
 
+#include <gen_cpp/Exprs_types.h>
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <memory>
+#include <ostream>
+#include <vector>
+
 #include "common/status.h"
-#include "vec/columns/column_nullable.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/exprs/vexpr_context.h"
+#include "vec/functions/simple_function_factory.h"
+
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exprs/vcase_expr.h b/be/src/vec/exprs/vcase_expr.h
index 6b056f9431..61199b0f47 100644
--- a/be/src/vec/exprs/vcase_expr.h
+++ b/be/src/vec/exprs/vcase_expr.h
@@ -17,8 +17,24 @@
 
 #pragma once
 
+#include <string>
+
+#include "common/object_pool.h"
+#include "common/status.h"
+#include "udf/udf.h"
 #include "vec/exprs/vexpr.h"
-#include "vec/functions/function_case.h"
+#include "vec/functions/function.h"
+
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+class TExprNode;
+
+namespace vectorized {
+class Block;
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exprs/vcast_expr.cpp b/be/src/vec/exprs/vcast_expr.cpp
index 8253216de4..65fa18e89b 100644
--- a/be/src/vec/exprs/vcast_expr.cpp
+++ b/be/src/vec/exprs/vcast_expr.cpp
@@ -17,14 +17,29 @@
 
 #include "vec/exprs/vcast_expr.h"
 
-#include <string_view>
+#include <fmt/format.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <memory>
+#include <ostream>
+#include <vector>
 
 #include "common/status.h"
-#include "vec/core/field.h"
-#include "vec/data_types/data_type_factory.hpp"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
 #include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+} // namespace doris
+
 namespace doris::vectorized {
 
 doris::Status VCastExpr::prepare(doris::RuntimeState* state, const doris::RowDescriptor& desc,
diff --git a/be/src/vec/exprs/vcast_expr.h b/be/src/vec/exprs/vcast_expr.h
index 9ac0dfe9db..33146eecf5 100644
--- a/be/src/vec/exprs/vcast_expr.h
+++ b/be/src/vec/exprs/vcast_expr.h
@@ -16,9 +16,26 @@
 // under the License.
 
 #pragma once
+#include <string>
+
+#include "common/object_pool.h"
+#include "common/status.h"
+#include "udf/udf.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/data_types/data_type.h"
 #include "vec/exprs/vexpr.h"
 #include "vec/functions/function.h"
 
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+class TExprNode;
+namespace vectorized {
+class Block;
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 class VCastExpr final : public VExpr {
 public:
diff --git a/be/src/vec/exprs/vcolumn_ref.h b/be/src/vec/exprs/vcolumn_ref.h
index b5985fe82c..b470807ab5 100644
--- a/be/src/vec/exprs/vcolumn_ref.h
+++ b/be/src/vec/exprs/vcolumn_ref.h
@@ -16,6 +16,7 @@
 // under the License.
 
 #pragma once
+#include "runtime/descriptors.h"
 #include "runtime/runtime_state.h"
 #include "vec/exprs/vexpr.h"
 #include "vec/functions/function.h"
diff --git a/be/src/vec/exprs/vectorized_agg_fn.cpp b/be/src/vec/exprs/vectorized_agg_fn.cpp
index 555ce77f0f..80a1785349 100644
--- a/be/src/vec/exprs/vectorized_agg_fn.cpp
+++ b/be/src/vec/exprs/vectorized_agg_fn.cpp
@@ -17,16 +17,37 @@
 
 #include "vec/exprs/vectorized_agg_fn.h"
 
-#include "fmt/format.h"
-#include "fmt/ranges.h"
-#include "runtime/descriptors.h"
+#include <fmt/format.h>
+#include <fmt/ranges.h> // IWYU pragma: keep
+#include <gen_cpp/Exprs_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <glog/logging.h>
+
+#include <memory>
+#include <ostream>
+#include <string_view>
+
+#include "common/config.h"
+#include "common/object_pool.h"
 #include "vec/aggregate_functions/aggregate_function_java_udaf.h"
 #include "vec/aggregate_functions/aggregate_function_rpc.h"
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
 #include "vec/aggregate_functions/aggregate_function_sort.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/core/materialize_block.h"
 #include "vec/data_types/data_type_factory.hpp"
 #include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
+
+namespace doris {
+class RowDescriptor;
+namespace vectorized {
+class Arena;
+class BufferWritable;
+class IColumn;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exprs/vectorized_agg_fn.h b/be/src/vec/exprs/vectorized_agg_fn.h
index 2a7f572af1..3ee25a878f 100644
--- a/be/src/vec/exprs/vectorized_agg_fn.h
+++ b/be/src/vec/exprs/vectorized_agg_fn.h
@@ -16,18 +16,35 @@
 // under the License.
 
 #pragma once
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+
+#include <string>
+#include <vector>
+
+#include "common/status.h"
 #include "runtime/types.h"
 #include "util/runtime_profile.h"
 #include "vec/aggregate_functions/aggregate_function.h"
-#include "vec/core/block.h"
 #include "vec/core/sort_description.h"
 #include "vec/data_types/data_type.h"
-#include "vec/exprs/vexpr_context.h"
 
 namespace doris {
 class RuntimeState;
 class SlotDescriptor;
+class ObjectPool;
+class RowDescriptor;
+class TExpr;
+class TExprNode;
+class TSortInfo;
+
 namespace vectorized {
+class Arena;
+class Block;
+class BufferWritable;
+class IColumn;
+class VExprContext;
+
 class AggFnEvaluator {
 public:
     static Status create(ObjectPool* pool, const TExpr& desc, const TSortInfo& sort_info,
diff --git a/be/src/vec/exprs/vectorized_fn_call.cpp b/be/src/vec/exprs/vectorized_fn_call.cpp
index af551b7dcc..bec7d0c1a3 100644
--- a/be/src/vec/exprs/vectorized_fn_call.cpp
+++ b/be/src/vec/exprs/vectorized_fn_call.cpp
@@ -17,19 +17,37 @@
 
 #include "vec/exprs/vectorized_fn_call.h"
 
-#include <string_view>
+#include <fmt/format.h>
+#include <fmt/ranges.h> // IWYU pragma: keep
+#include <gen_cpp/Types_types.h>
 
+#include <algorithm>
+#include <memory>
+#include <ostream>
+#include <string_view>
+#include <utility>
+
+#include "common/config.h"
 #include "common/consts.h"
 #include "common/status.h"
-#include "fmt/format.h"
-#include "fmt/ranges.h"
+#include "runtime/runtime_state.h"
 #include "udf/udf.h"
-#include "vec/data_types/data_type_nullable.h"
-#include "vec/data_types/data_type_number.h"
+#include "vec/columns/column.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/data_types/data_type.h"
+#include "vec/exprs/vexpr_context.h"
 #include "vec/functions/function_java_udf.h"
 #include "vec/functions/function_rpc.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+class TExprNode;
+} // namespace doris
+
 namespace doris::vectorized {
 
 VectorizedFnCall::VectorizedFnCall(const doris::TExprNode& node) : VExpr(node) {}
diff --git a/be/src/vec/exprs/vectorized_fn_call.h b/be/src/vec/exprs/vectorized_fn_call.h
index c2fba08ccc..d50ae3dd0b 100644
--- a/be/src/vec/exprs/vectorized_fn_call.h
+++ b/be/src/vec/exprs/vectorized_fn_call.h
@@ -16,10 +16,29 @@
 // under the License.
 
 #pragma once
-#include "runtime/runtime_state.h"
+#include <stddef.h>
+
+#include <string>
+#include <vector>
+
+#include "common/object_pool.h"
+#include "common/status.h"
+#include "udf/udf.h"
+#include "vec/core/column_numbers.h"
 #include "vec/exprs/vexpr.h"
 #include "vec/functions/function.h"
 
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+class TExprNode;
+
+namespace vectorized {
+class Block;
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 class VectorizedFnCall : public VExpr {
 public:
diff --git a/be/src/vec/exprs/vexpr.cpp b/be/src/vec/exprs/vexpr.cpp
index b1724c9af8..49995b89a8 100644
--- a/be/src/vec/exprs/vexpr.cpp
+++ b/be/src/vec/exprs/vexpr.cpp
@@ -17,31 +17,44 @@
 
 #include "vec/exprs/vexpr.h"
 
-#include <fmt/format.h>
+#include <gen_cpp/Exprs_types.h>
 #include <thrift/protocol/TDebugProtocol.h>
 
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
 #include <memory>
+#include <stack>
 
-#include "gen_cpp/Exprs_types.h"
+#include "common/config.h"
+#include "common/exception.h"
+#include "common/object_pool.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
 #include "vec/data_types/data_type_factory.hpp"
+#include "vec/data_types/data_type_number.h"
 #include "vec/exprs/varray_literal.h"
 #include "vec/exprs/vcase_expr.h"
 #include "vec/exprs/vcast_expr.h"
 #include "vec/exprs/vcolumn_ref.h"
 #include "vec/exprs/vcompound_pred.h"
 #include "vec/exprs/vectorized_fn_call.h"
+#include "vec/exprs/vexpr_context.h"
 #include "vec/exprs/vin_predicate.h"
 #include "vec/exprs/vinfo_func.h"
 #include "vec/exprs/vlambda_function_call_expr.h"
 #include "vec/exprs/vlambda_function_expr.h"
 #include "vec/exprs/vliteral.h"
 #include "vec/exprs/vmap_literal.h"
-#include "vec/exprs/vruntimefilter_wrapper.h"
 #include "vec/exprs/vschema_change_expr.h"
 #include "vec/exprs/vslot_ref.h"
 #include "vec/exprs/vstruct_literal.h"
 #include "vec/exprs/vtuple_is_null_predicate.h"
 
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+} // namespace doris
+
 namespace doris::vectorized {
 using doris::Status;
 using doris::RuntimeState;
diff --git a/be/src/vec/exprs/vexpr.h b/be/src/vec/exprs/vexpr.h
index b45bf7bbb4..168975d449 100644
--- a/be/src/vec/exprs/vexpr.h
+++ b/be/src/vec/exprs/vexpr.h
@@ -17,21 +17,39 @@
 
 #pragma once
 
+#include <gen_cpp/Exprs_types.h>
+#include <gen_cpp/Opcodes_types.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
 #include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
 #include <vector>
 
 #include "common/status.h"
-#include "exprs/bitmapfilter_predicate.h"
-#include "exprs/hybrid_set.h"
-#include "gen_cpp/Exprs_types.h"
+#include "runtime/define_primitive_type.h"
 #include "runtime/types.h"
 #include "udf/udf.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/data_types/data_type.h"
-#include "vec/exprs/vexpr_context.h"
 #include "vec/functions/function.h"
 
 namespace doris {
+class BitmapFilterFuncBase;
+class BloomFilterFuncBase;
+class HybridSetBase;
+class ObjectPool;
+class RowDescriptor;
+class RuntimeState;
+
 namespace vectorized {
+class VExprContext;
 
 #define RETURN_IF_ERROR_OR_PREPARED(stmt) \
     if (_prepared) {                      \
diff --git a/be/src/vec/exprs/vexpr_context.cpp b/be/src/vec/exprs/vexpr_context.cpp
index ab8ee1d696..73cd294e2d 100644
--- a/be/src/vec/exprs/vexpr_context.cpp
+++ b/be/src/vec/exprs/vexpr_context.cpp
@@ -17,10 +17,25 @@
 
 #include "vec/exprs/vexpr_context.h"
 
+#include <algorithm>
+#include <ostream>
+#include <string>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/object_pool.h"
+#include "runtime/runtime_state.h"
+#include "runtime/thread_context.h"
 #include "udf/udf.h"
 #include "util/stack_util.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
 #include "vec/exprs/vexpr.h"
 
+namespace doris {
+class RowDescriptor;
+} // namespace doris
+
 namespace doris::vectorized {
 VExprContext::VExprContext(VExpr* expr)
         : _root(expr),
diff --git a/be/src/vec/exprs/vexpr_context.h b/be/src/vec/exprs/vexpr_context.h
index e1b9bf69d7..d9b61b7732 100644
--- a/be/src/vec/exprs/vexpr_context.h
+++ b/be/src/vec/exprs/vexpr_context.h
@@ -17,10 +17,21 @@
 
 #pragma once
 
+#include <glog/logging.h>
+
+#include <memory>
+#include <vector>
+
 #include "common/status.h"
-#include "runtime/runtime_state.h"
+#include "runtime/types.h"
+#include "udf/udf.h"
 #include "vec/core/block.h"
 
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+} // namespace doris
+
 namespace doris::vectorized {
 class VExpr;
 
diff --git a/be/src/vec/exprs/vin_predicate.cpp b/be/src/vec/exprs/vin_predicate.cpp
index e9bb6355d5..3aecb5fbca 100644
--- a/be/src/vec/exprs/vin_predicate.cpp
+++ b/be/src/vec/exprs/vin_predicate.cpp
@@ -17,16 +17,29 @@
 
 #include "vec/exprs/vin_predicate.h"
 
-#include <string_view>
+#include <fmt/format.h>
+#include <gen_cpp/Exprs_types.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <memory>
+#include <ostream>
+#include <vector>
 
 #include "common/status.h"
-#include "exprs/create_predicate_function.h"
-#include "vec/columns/column_set.h"
-#include "vec/core/field.h"
-#include "vec/data_types/data_type_factory.hpp"
-#include "vec/data_types/data_type_nullable.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/exprs/vexpr_context.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+} // namespace doris
+
 namespace doris::vectorized {
 
 VInPredicate::VInPredicate(const TExprNode& node)
diff --git a/be/src/vec/exprs/vin_predicate.h b/be/src/vec/exprs/vin_predicate.h
index d8853ab488..2826308158 100644
--- a/be/src/vec/exprs/vin_predicate.h
+++ b/be/src/vec/exprs/vin_predicate.h
@@ -17,9 +17,24 @@
 
 #pragma once
 
+#include <string>
+
+#include "common/object_pool.h"
+#include "common/status.h"
+#include "udf/udf.h"
 #include "vec/exprs/vexpr.h"
 #include "vec/functions/function.h"
 
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+class TExprNode;
+namespace vectorized {
+class Block;
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 class VInPredicate final : public VExpr {
 public:
diff --git a/be/src/vec/exprs/vinfo_func.cpp b/be/src/vec/exprs/vinfo_func.cpp
index 0978385751..c262882b31 100644
--- a/be/src/vec/exprs/vinfo_func.cpp
+++ b/be/src/vec/exprs/vinfo_func.cpp
@@ -17,11 +17,26 @@
 
 #include "vec/exprs/vinfo_func.h"
 
-#include <fmt/format.h>
+#include <gen_cpp/Exprs_types.h>
+#include <glog/logging.h>
+#include <stddef.h>
 
-#include "util/string_parser.hpp"
+#include <algorithm>
+#include <iostream>
+#include <memory>
+
+#include "runtime/define_primitive_type.h"
+#include "runtime/types.h"
+#include "vec/core/block.h"
 #include "vec/core/field.h"
-#include "vec/data_types/data_type_nullable.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+
+namespace doris {
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exprs/vinfo_func.h b/be/src/vec/exprs/vinfo_func.h
index e2f1fb09af..3cbf13ea0f 100644
--- a/be/src/vec/exprs/vinfo_func.h
+++ b/be/src/vec/exprs/vinfo_func.h
@@ -17,17 +17,20 @@
 
 #pragma once
 
-#include <iostream>
 #include <string>
 
-#include "vec/columns/column.h"
-#include "vec/columns/column_const.h"
+#include "common/object_pool.h"
+#include "common/status.h"
+#include "vec/data_types/data_type.h"
 #include "vec/exprs/vexpr.h"
 
 namespace doris {
 class TExprNode;
 
 namespace vectorized {
+class Block;
+class VExprContext;
+
 class VInfoFunc : public VExpr {
 public:
     VInfoFunc(const TExprNode& node);
diff --git a/be/src/vec/exprs/vliteral.cpp b/be/src/vec/exprs/vliteral.cpp
index 69db0e2456..8025134c51 100644
--- a/be/src/vec/exprs/vliteral.cpp
+++ b/be/src/vec/exprs/vliteral.cpp
@@ -18,18 +18,39 @@
 #include "vec/exprs/vliteral.h"
 
 #include <fmt/format.h>
+#include <gen_cpp/Exprs_types.h>
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <math.h>
+#include <stdint.h>
+#include <sys/types.h>
 
+#include <algorithm>
+#include <ostream>
+#include <vector>
+
+#include "common/exception.h"
+#include "olap/olap_common.h"
+#include "runtime/decimalv2_value.h"
+#include "runtime/define_primitive_type.h"
 #include "runtime/jsonb_value.h"
 #include "runtime/large_int_value.h"
+#include "runtime/types.h"
 #include "util/string_parser.hpp"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/common/string_ref.h"
+#include "vec/common/typeid_cast.h"
+#include "vec/core/block.h"
 #include "vec/core/field.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_decimal.h"
-#include "vec/io/io_helper.h"
 #include "vec/runtime/vdatetime_value.h"
 
 namespace doris {
 
 namespace vectorized {
+class VExprContext;
 
 void VLiteral::init(const TExprNode& node) {
     Field field;
diff --git a/be/src/vec/exprs/vliteral.h b/be/src/vec/exprs/vliteral.h
index fcea4d9ffd..456375487a 100644
--- a/be/src/vec/exprs/vliteral.h
+++ b/be/src/vec/exprs/vliteral.h
@@ -17,13 +17,21 @@
 
 #pragma once
 
-#include "vec/columns/column.h"
+#include <memory>
+#include <string>
+
+#include "common/object_pool.h"
+#include "common/status.h"
+#include "vec/data_types/data_type.h"
 #include "vec/exprs/vexpr.h"
 
 namespace doris {
 class TExprNode;
 
 namespace vectorized {
+class Block;
+class VExprContext;
+
 class VLiteral : public VExpr {
 public:
     VLiteral(const TExprNode& node, bool should_init = true)
diff --git a/be/src/vec/exprs/vmap_literal.cpp b/be/src/vec/exprs/vmap_literal.cpp
index 1304569e49..9ae1193afa 100644
--- a/be/src/vec/exprs/vmap_literal.cpp
+++ b/be/src/vec/exprs/vmap_literal.cpp
@@ -17,6 +17,26 @@
 
 #include "vec/exprs/vmap_literal.h"
 
+#include <glog/logging.h>
+
+#include <memory>
+#include <ostream>
+#include <vector>
+
+#include "runtime/types.h"
+#include "vec/columns/column.h"
+#include "vec/core/field.h"
+#include "vec/data_types/data_type.h"
+#include "vec/exprs/vexpr.h"
+
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+
 //insert into table_map values ({'name':'zhangsan', 'gender':'male'}), ({'name':'lisi', 'gender':'female'});
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exprs/vmap_literal.h b/be/src/vec/exprs/vmap_literal.h
index 6206d4c58f..173ff0dcd3 100644
--- a/be/src/vec/exprs/vmap_literal.h
+++ b/be/src/vec/exprs/vmap_literal.h
@@ -16,11 +16,17 @@
 // under the License.
 #pragma once
 
+#include "common/status.h"
 #include "vec/exprs/vliteral.h"
 
 namespace doris {
+class RowDescriptor;
+class RuntimeState;
+class TExprNode;
 
 namespace vectorized {
+class VExprContext;
+
 class VMapLiteral : public VLiteral {
 public:
     VMapLiteral(const TExprNode& node) : VLiteral(node, false) {}
diff --git a/be/src/vec/exprs/vruntimefilter_wrapper.cpp b/be/src/vec/exprs/vruntimefilter_wrapper.cpp
index 456035c396..554a15231f 100644
--- a/be/src/vec/exprs/vruntimefilter_wrapper.cpp
+++ b/be/src/vec/exprs/vruntimefilter_wrapper.cpp
@@ -17,13 +17,30 @@
 
 #include "vec/exprs/vruntimefilter_wrapper.h"
 
-#include <string_view>
+#include <fmt/format.h>
+#include <stddef.h>
+
+#include <memory>
+#include <utility>
 
 #include "util/simd/bits.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_nullable.h"
-#include "vec/core/field.h"
-#include "vec/data_types/data_type_factory.hpp"
-#include "vec/functions/simple_function_factory.h"
+#include "vec/columns/column_vector.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+class TExprNode;
+
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exprs/vruntimefilter_wrapper.h b/be/src/vec/exprs/vruntimefilter_wrapper.h
index d78880a636..b2a03db8ba 100644
--- a/be/src/vec/exprs/vruntimefilter_wrapper.h
+++ b/be/src/vec/exprs/vruntimefilter_wrapper.h
@@ -17,8 +17,28 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include <atomic>
+#include <string>
+#include <vector>
+
+#include "common/config.h"
+#include "common/object_pool.h"
+#include "common/status.h"
+#include "udf/udf.h"
 #include "vec/exprs/vexpr.h"
 
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+class TExprNode;
+namespace vectorized {
+class Block;
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 class VRuntimeFilterWrapper final : public VExpr {
 public:
diff --git a/be/src/vec/exprs/vschema_change_expr.cpp b/be/src/vec/exprs/vschema_change_expr.cpp
index cfe5e1c654..0e0e71e167 100644
--- a/be/src/vec/exprs/vschema_change_expr.cpp
+++ b/be/src/vec/exprs/vschema_change_expr.cpp
@@ -17,9 +17,28 @@
 
 #include "vec/exprs/vschema_change_expr.h"
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+
+#include <memory>
+#include <vector>
+
+#include "runtime/descriptors.h"
+#include "runtime/runtime_state.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_object.h"
+#include "vec/common/assert_cast.h"
 #include "vec/common/schema_util.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/core/columns_with_type_and_name.h"
+#include "vec/json/path_in_data.h"
+
+namespace doris {
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exprs/vschema_change_expr.h b/be/src/vec/exprs/vschema_change_expr.h
index 52a64b8f50..afa35f6b4c 100644
--- a/be/src/vec/exprs/vschema_change_expr.h
+++ b/be/src/vec/exprs/vschema_change_expr.h
@@ -16,10 +16,25 @@
 // under the License.
 
 #pragma once
-#include "vec/common/schema_util.h"
-#include "vec/data_types/data_type_object.h"
+#include <gen_cpp/Exprs_types.h>
+#include <stdint.h>
+
+#include <string>
+
+#include "common/object_pool.h"
+#include "common/status.h"
+#include "udf/udf.h"
 #include "vec/exprs/vexpr.h"
-#include "vec/functions/function.h"
+
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+
+namespace vectorized {
+class Block;
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exprs/vslot_ref.cpp b/be/src/vec/exprs/vslot_ref.cpp
index a1ff659511..4b2e182563 100644
--- a/be/src/vec/exprs/vslot_ref.cpp
+++ b/be/src/vec/exprs/vslot_ref.cpp
@@ -17,8 +17,22 @@
 
 #include "vec/exprs/vslot_ref.h"
 
+#include <gen_cpp/Exprs_types.h>
+#include <glog/logging.h>
+
+#include <ostream>
+#include <vector>
+
 #include "common/status.h"
 #include "runtime/descriptors.h"
+#include "runtime/runtime_state.h"
+#include "vec/core/block.h"
+
+namespace doris {
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exprs/vslot_ref.h b/be/src/vec/exprs/vslot_ref.h
index 50d9d24ece..658e0f05ac 100644
--- a/be/src/vec/exprs/vslot_ref.h
+++ b/be/src/vec/exprs/vslot_ref.h
@@ -16,13 +16,22 @@
 // under the License.
 
 #pragma once
-#include "runtime/runtime_state.h"
+#include <string>
+
+#include "common/object_pool.h"
+#include "common/status.h"
 #include "vec/exprs/vexpr.h"
-#include "vec/functions/function.h"
 
 namespace doris {
 class SlotDescriptor;
+class RowDescriptor;
+class RuntimeState;
+class TExprNode;
+
 namespace vectorized {
+class Block;
+class VExprContext;
+
 class VSlotRef final : public VExpr {
 public:
     VSlotRef(const doris::TExprNode& node);
diff --git a/be/src/vec/exprs/vstruct_literal.cpp b/be/src/vec/exprs/vstruct_literal.cpp
index 3ae034f64d..0e8c197268 100644
--- a/be/src/vec/exprs/vstruct_literal.cpp
+++ b/be/src/vec/exprs/vstruct_literal.cpp
@@ -17,6 +17,22 @@
 
 #include "vec/exprs/vstruct_literal.h"
 
+#include <memory>
+#include <vector>
+
+#include "vec/columns/column.h"
+#include "vec/core/field.h"
+#include "vec/data_types/data_type.h"
+#include "vec/exprs/vexpr.h"
+
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 Status VStructLiteral::prepare(RuntimeState* state, const RowDescriptor& row_desc,
diff --git a/be/src/vec/exprs/vstruct_literal.h b/be/src/vec/exprs/vstruct_literal.h
index 4cc4fe78d6..7af79946b7 100644
--- a/be/src/vec/exprs/vstruct_literal.h
+++ b/be/src/vec/exprs/vstruct_literal.h
@@ -17,11 +17,17 @@
 
 #pragma once
 
+#include "common/status.h"
 #include "vec/exprs/vliteral.h"
 
 namespace doris {
+class RowDescriptor;
+class RuntimeState;
+class TExprNode;
 
 namespace vectorized {
+class VExprContext;
+
 class VStructLiteral : public VLiteral {
 public:
     VStructLiteral(const TExprNode& node) : VLiteral(node, false) {}
diff --git a/be/src/vec/exprs/vtuple_is_null_predicate.cpp b/be/src/vec/exprs/vtuple_is_null_predicate.cpp
index acb41f73b0..b17428bfc0 100644
--- a/be/src/vec/exprs/vtuple_is_null_predicate.cpp
+++ b/be/src/vec/exprs/vtuple_is_null_predicate.cpp
@@ -17,13 +17,21 @@
 
 #include "vec/exprs/vtuple_is_null_predicate.h"
 
-#include <string_view>
+#include <gen_cpp/Exprs_types.h>
+#include <glog/logging.h>
 
-#include "exprs/create_predicate_function.h"
-#include "vec/core/field.h"
-#include "vec/data_types/data_type_factory.hpp"
-#include "vec/data_types/data_type_nullable.h"
-#include "vec/functions/simple_function_factory.h"
+#include <ostream>
+#include <vector>
+
+#include "runtime/descriptors.h"
+
+namespace doris {
+class RuntimeState;
+namespace vectorized {
+class Block;
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/exprs/vtuple_is_null_predicate.h b/be/src/vec/exprs/vtuple_is_null_predicate.h
index 2e83fb88ad..91d217a756 100644
--- a/be/src/vec/exprs/vtuple_is_null_predicate.h
+++ b/be/src/vec/exprs/vtuple_is_null_predicate.h
@@ -17,8 +17,24 @@
 
 #pragma once
 
+#include <stdint.h>
+
+#include <string>
+
+#include "common/object_pool.h"
+#include "common/status.h"
 #include "vec/exprs/vexpr.h"
 
+namespace doris {
+class RowDescriptor;
+class RuntimeState;
+class TExprNode;
+namespace vectorized {
+class Block;
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 class VTupleIsNullPredicate final : public VExpr {
 public:
diff --git a/be/src/vec/functions/array/function_array_aggregation.cpp b/be/src/vec/functions/array/function_array_aggregation.cpp
index c0d45acea1..102acb9d39 100644
--- a/be/src/vec/functions/array/function_array_aggregation.cpp
+++ b/be/src/vec/functions/array/function_array_aggregation.cpp
@@ -18,18 +18,30 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Functions/array/arrayAggregation.cpp
 // and modified by Doris
 
-#include <type_traits>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <memory>
+#include <type_traits>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/aggregate_functions/aggregate_function_avg.h"
 #include "vec/aggregate_functions/aggregate_function_min_max.h"
-#include "vec/aggregate_functions/aggregate_function_null.h"
 #include "vec/aggregate_functions/aggregate_function_product.h"
 #include "vec/aggregate_functions/aggregate_function_sum.h"
 #include "vec/aggregate_functions/helpers.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_array.h"
+#include "vec/columns/column_decimal.h"
 #include "vec/columns/column_nullable.h"
 #include "vec/common/arena.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
 #include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_array.h"
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/functions/array/function_array_join.h"
 #include "vec/functions/array/function_array_mapped.h"
diff --git a/be/src/vec/functions/array/function_array_apply.cpp b/be/src/vec/functions/array/function_array_apply.cpp
index a47a28f281..a9de042e98 100644
--- a/be/src/vec/functions/array/function_array_apply.cpp
+++ b/be/src/vec/functions/array/function_array_apply.cpp
@@ -15,14 +15,39 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <vec/columns/column_array.h>
-#include <vec/columns/column_nullable.h>
-#include <vec/columns/columns_number.h>
-#include <vec/data_types/data_type_array.h>
-#include <vec/data_types/data_type_number.h>
-#include <vec/functions/function.h>
-#include <vec/functions/function_helpers.h>
-#include <vec/functions/simple_function_factory.h>
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
+#include "runtime/thread_context.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_array.h"
+#include "vec/columns/column_const.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_array.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/functions/function.h"
+#include "vec/functions/simple_function_factory.h"
+
+namespace doris {
+class FunctionContext;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/array/function_array_compact.h b/be/src/vec/functions/array/function_array_compact.h
index 55488d7631..57c07b38d8 100644
--- a/be/src/vec/functions/array/function_array_compact.h
+++ b/be/src/vec/functions/array/function_array_compact.h
@@ -17,11 +17,32 @@
 
 #pragma once
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_array.h"
-#include "vec/data_types/data_type_number.h"
 #include "vec/functions/function.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 class FunctionArrayCompact : public IFunction {
diff --git a/be/src/vec/functions/array/function_array_concat.cpp b/be/src/vec/functions/array/function_array_concat.cpp
index 02377272e9..fe7cfa8eca 100644
--- a/be/src/vec/functions/array/function_array_concat.cpp
+++ b/be/src/vec/functions/array/function_array_concat.cpp
@@ -15,13 +15,33 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <cstddef>
+#include <glog/logging.h>
 
+#include <cstddef>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_array.h"
 #include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 // array_concat([1, 2], [7, 8], [5, 6]) -> [1, 2, 7, 8, 5, 6]
diff --git a/be/src/vec/functions/array/function_array_constructor.cpp b/be/src/vec/functions/array/function_array_constructor.cpp
index 54bb79c2e6..b7133ec0db 100644
--- a/be/src/vec/functions/array/function_array_constructor.cpp
+++ b/be/src/vec/functions/array/function_array_constructor.cpp
@@ -14,18 +14,35 @@
 // KIND, either express or implied.  See the License for the
 // specific language governing permissions and limitations
 // under the License.
-#include <string_view>
+#include <glog/logging.h>
+#include <stddef.h>
 
+#include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
-#include "vec/columns/column_string.h"
-#include "vec/common/string_ref.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_array.h"
-#include "vec/data_types/data_type_number.h"
-#include "vec/data_types/get_least_supertype.h"
-#include "vec/functions/array/function_array_utils.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 // construct an array
diff --git a/be/src/vec/functions/array/function_array_difference.cpp b/be/src/vec/functions/array/function_array_difference.cpp
index 589d0c5b4e..5cb66ba36c 100644
--- a/be/src/vec/functions/array/function_array_difference.cpp
+++ b/be/src/vec/functions/array/function_array_difference.cpp
@@ -17,7 +17,6 @@
 
 #include "vec/functions/array/function_array_difference.h"
 
-#include "vec/functions/array/function_array_mapped.h"
 #include "vec/functions/simple_function_factory.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/functions/array/function_array_difference.h b/be/src/vec/functions/array/function_array_difference.h
index 7e6c957a96..242d62adbf 100644
--- a/be/src/vec/functions/array/function_array_difference.h
+++ b/be/src/vec/functions/array/function_array_difference.h
@@ -17,8 +17,29 @@
 
 #pragma once
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
+#include "vec/columns/column_decimal.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_array.h"
@@ -27,6 +48,10 @@
 #include "vec/functions/function.h"
 #include "vec/utils/util.hpp"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 class FunctionArrayDifference : public IFunction {
diff --git a/be/src/vec/functions/array/function_array_distinct.h b/be/src/vec/functions/array/function_array_distinct.h
index 7707438dce..ae8e3ac600 100644
--- a/be/src/vec/functions/array/function_array_distinct.h
+++ b/be/src/vec/functions/array/function_array_distinct.h
@@ -19,12 +19,42 @@
 // and modified by Doris
 #pragma once
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <string.h>
+
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/hash_table/hash.h"
 #include "vec/common/hash_table/hash_set.h"
-#include "vec/common/hash_table/hash_table.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_array.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/functions/function.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+template <typename, typename>
+struct DefaultHash;
+
 namespace doris::vectorized {
 
 class FunctionArrayDistinct : public IFunction {
diff --git a/be/src/vec/functions/array/function_array_element.h b/be/src/vec/functions/array/function_array_element.h
index 9b84a274a9..b6ef0d305f 100644
--- a/be/src/vec/functions/array/function_array_element.h
+++ b/be/src/vec/functions/array/function_array_element.h
@@ -19,18 +19,41 @@
 // and modified by Doris
 #pragma once
 
-#include <string_view>
+#include <glog/logging.h>
+#include <string.h>
 
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
 #include "vec/columns/column_map.h"
+#include "vec/columns/column_nullable.h"
 #include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_array.h"
 #include "vec/data_types/data_type_map.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/functions/function.h"
 #include "vec/functions/function_helpers.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 class FunctionArrayElement : public IFunction {
diff --git a/be/src/vec/functions/array/function_array_enumerate.cpp b/be/src/vec/functions/array/function_array_enumerate.cpp
index 8f63e46ab1..9594f87ba8 100644
--- a/be/src/vec/functions/array/function_array_enumerate.cpp
+++ b/be/src/vec/functions/array/function_array_enumerate.cpp
@@ -15,14 +15,39 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <vec/columns/column_array.h>
-#include <vec/columns/column_nullable.h>
-#include <vec/columns/columns_number.h>
-#include <vec/data_types/data_type_array.h>
-#include <vec/data_types/data_type_number.h>
-#include <vec/functions/function.h>
-#include <vec/functions/function_helpers.h>
-#include <vec/functions/simple_function_factory.h>
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_array.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_array.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/data_types/data_type_number.h"
+#include "vec/functions/function.h"
+#include "vec/functions/function_helpers.h"
+#include "vec/functions/simple_function_factory.h"
+
+namespace doris {
+class FunctionContext;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/array/function_array_enumerate_uniq.cpp b/be/src/vec/functions/array/function_array_enumerate_uniq.cpp
index d84fea03d6..3ca1084bf2 100644
--- a/be/src/vec/functions/array/function_array_enumerate_uniq.cpp
+++ b/be/src/vec/functions/array/function_array_enumerate_uniq.cpp
@@ -15,16 +15,52 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <vec/columns/column_array.h>
-#include <vec/columns/column_nullable.h>
-#include <vec/columns/columns_number.h>
-#include <vec/common/columns_hashing.h>
-#include <vec/common/hash_table/hash_map.h>
-#include <vec/data_types/data_type_array.h>
-#include <vec/data_types/data_type_number.h>
-#include <vec/functions/function.h>
-#include <vec/functions/function_helpers.h>
-#include <vec/functions/simple_function_factory.h>
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <new>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_array.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/arena.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/columns_hashing.h"
+#include "vec/common/hash_table/hash.h"
+#include "vec/common/hash_table/hash_map.h"
+#include "vec/common/hash_table/hash_table.h"
+#include "vec/common/hash_table/hash_table_allocator.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/common/string_ref.h"
+#include "vec/common/uint128.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_array.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/data_types/data_type_number.h"
+#include "vec/functions/function.h"
+#include "vec/functions/function_helpers.h"
+#include "vec/functions/simple_function_factory.h"
+
+namespace doris {
+class FunctionContext;
+} // namespace doris
+template <typename, typename>
+struct DefaultHash;
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/array/function_array_except.cpp b/be/src/vec/functions/array/function_array_except.cpp
index 93894b0e47..2e8d0532f4 100644
--- a/be/src/vec/functions/array/function_array_except.cpp
+++ b/be/src/vec/functions/array/function_array_except.cpp
@@ -15,6 +15,11 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include "vec/columns/column_const.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/string_ref.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/functions/array/function_array_binary.h"
 #include "vec/functions/array/function_array_set.h"
 #include "vec/functions/simple_function_factory.h"
diff --git a/be/src/vec/functions/array/function_array_exists.cpp b/be/src/vec/functions/array/function_array_exists.cpp
index 4f4fa593e5..2e0288f869 100644
--- a/be/src/vec/functions/array/function_array_exists.cpp
+++ b/be/src/vec/functions/array/function_array_exists.cpp
@@ -15,16 +15,33 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <vec/columns/column_array.h>
-#include <vec/columns/column_nullable.h>
-#include <vec/columns/columns_number.h>
-#include <vec/data_types/data_type_array.h>
-#include <vec/data_types/data_type_number.h>
-#include <vec/functions/function.h>
-#include <vec/functions/function_helpers.h>
-#include <vec/functions/simple_function_factory.h>
+#include <glog/logging.h>
+#include <stddef.h>
 
-#include <vec/utils/util.hpp>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_array.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/functions/function.h"
+#include "vec/functions/simple_function_factory.h"
+
+namespace doris {
+class FunctionContext;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/array/function_array_first_index.cpp b/be/src/vec/functions/array/function_array_first_index.cpp
index 3bf4379f6a..c17f3e0a4d 100644
--- a/be/src/vec/functions/array/function_array_first_index.cpp
+++ b/be/src/vec/functions/array/function_array_first_index.cpp
@@ -15,16 +15,32 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <vec/columns/column_array.h>
-#include <vec/columns/column_nullable.h>
-#include <vec/columns/columns_number.h>
-#include <vec/data_types/data_type_array.h>
-#include <vec/data_types/data_type_number.h>
-#include <vec/functions/function.h>
-#include <vec/functions/function_helpers.h>
-#include <vec/functions/simple_function_factory.h>
+#include <stddef.h>
 
-#include <vec/utils/util.hpp>
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_array.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type_number.h"
+#include "vec/functions/function.h"
+#include "vec/functions/simple_function_factory.h"
+
+namespace doris {
+class FunctionContext;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/array/function_array_index.h b/be/src/vec/functions/array/function_array_index.h
index 3472035747..8e4d73687a 100644
--- a/be/src/vec/functions/array/function_array_index.h
+++ b/be/src/vec/functions/array/function_array_index.h
@@ -19,14 +19,34 @@
 // and modified by Doris
 #pragma once
 
+#include <stddef.h>
+
+#include <memory>
+#include <utility>
+
+#include "common/status.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
+#include "vec/columns/column_nullable.h"
 #include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
 #include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_array.h"
-#include "vec/data_types/data_type_number.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/data_types/data_type_number.h" // IWYU pragma: keep
 #include "vec/functions/function.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 struct ArrayContainsAction {
diff --git a/be/src/vec/functions/array/function_array_intersect.cpp b/be/src/vec/functions/array/function_array_intersect.cpp
index cf726fce6b..ed14195bca 100644
--- a/be/src/vec/functions/array/function_array_intersect.cpp
+++ b/be/src/vec/functions/array/function_array_intersect.cpp
@@ -15,6 +15,11 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include "vec/columns/column_const.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/string_ref.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/functions/array/function_array_binary.h"
 #include "vec/functions/array/function_array_set.h"
 #include "vec/functions/simple_function_factory.h"
diff --git a/be/src/vec/functions/array/function_array_popback.cpp b/be/src/vec/functions/array/function_array_popback.cpp
index ae8c76df66..89be7b2263 100644
--- a/be/src/vec/functions/array/function_array_popback.cpp
+++ b/be/src/vec/functions/array/function_array_popback.cpp
@@ -14,17 +14,33 @@
 // KIND, either express or implied.  See the License for the
 // specific language governing permissions and limitations
 // under the License.
-#include <string_view>
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stddef.h>
 
-#include "vec/columns/column_array.h"
-#include "vec/columns/column_string.h"
-#include "vec/common/string_ref.h"
-#include "vec/data_types/data_type_array.h"
-#include "vec/data_types/data_type_number.h"
+#include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/functions/array/function_array_utils.h"
 #include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 class FunctionArrayPopback : public IFunction {
diff --git a/be/src/vec/functions/array/function_array_popfront.cpp b/be/src/vec/functions/array/function_array_popfront.cpp
index 8c4ed89db4..c156c8dca6 100644
--- a/be/src/vec/functions/array/function_array_popfront.cpp
+++ b/be/src/vec/functions/array/function_array_popfront.cpp
@@ -14,17 +14,33 @@
 // KIND, either express or implied.  See the License for the
 // specific language governing permissions and limitations
 // under the License.
-#include <string_view>
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stddef.h>
 
-#include "vec/columns/column_array.h"
-#include "vec/columns/column_string.h"
-#include "vec/common/string_ref.h"
-#include "vec/data_types/data_type_array.h"
-#include "vec/data_types/data_type_number.h"
+#include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/functions/array/function_array_utils.h"
 #include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 class FunctionArrayPopfront : public IFunction {
diff --git a/be/src/vec/functions/array/function_array_pushfront.cpp b/be/src/vec/functions/array/function_array_pushfront.cpp
index 5da81776a7..4e19ad8c33 100644
--- a/be/src/vec/functions/array/function_array_pushfront.cpp
+++ b/be/src/vec/functions/array/function_array_pushfront.cpp
@@ -18,15 +18,29 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Functions/array/arrayIndex.h
 // and modified by Doris
 
+#include <stddef.h>
+
+#include <memory>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
-#include "vec/columns/column_const.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_nullable.h"
-#include "vec/functions/array/function_array_utils.h"
 #include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 class FunctionArrayPushfront : public IFunction {
diff --git a/be/src/vec/functions/array/function_array_range.cpp b/be/src/vec/functions/array/function_array_range.cpp
index 15fb78e578..06cb8fda15 100644
--- a/be/src/vec/functions/array/function_array_range.cpp
+++ b/be/src/vec/functions/array/function_array_range.cpp
@@ -15,19 +15,39 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <utility>
+
 #include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
-#include "vec/columns/column_const.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_array.h"
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/functions/function.h"
-#include "vec/functions/function_helpers.h"
 #include "vec/functions/simple_function_factory.h"
 #include "vec/utils/util.hpp"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 template <typename Impl>
diff --git a/be/src/vec/functions/array/function_array_remove.h b/be/src/vec/functions/array/function_array_remove.h
index 58566f34cb..065f348fa8 100644
--- a/be/src/vec/functions/array/function_array_remove.h
+++ b/be/src/vec/functions/array/function_array_remove.h
@@ -17,15 +17,38 @@
 
 #pragma once
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <string.h>
+
+#include <memory>
+#include <ostream>
+#include <string>
+#include <string_view>
+#include <utility>
+
+#include "common/status.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
-#include "vec/columns/column_const.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_array.h"
-#include "vec/data_types/data_type_number.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/functions/function.h"
-#include "vec/functions/function_helpers.h"
+
+namespace doris {
+class FunctionContext;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/array/function_array_shuffle.cpp b/be/src/vec/functions/array/function_array_shuffle.cpp
index 64cf845517..6f080b0cb9 100644
--- a/be/src/vec/functions/array/function_array_shuffle.cpp
+++ b/be/src/vec/functions/array/function_array_shuffle.cpp
@@ -14,16 +14,35 @@
 // KIND, either express or implied.  See the License for the
 // specific language governing permissions and limitations
 // under the License.
-#include <unistd.h>
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stdint.h>
+#include <time.h>
 
+#include <algorithm>
+#include <memory>
+#include <ostream>
 #include <random>
+#include <string>
+#include <utility>
 
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
-#include "vec/data_types/data_type_array.h"
-#include "vec/data_types/data_type_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 class FunctionArrayShuffle : public IFunction {
diff --git a/be/src/vec/functions/array/function_array_size.h b/be/src/vec/functions/array/function_array_size.h
index 1e375814b7..deb46970b8 100644
--- a/be/src/vec/functions/array/function_array_size.h
+++ b/be/src/vec/functions/array/function_array_size.h
@@ -17,12 +17,31 @@
 
 #pragma once
 
+#include <glog/logging.h>
+#include <sys/types.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <ostream>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
-#include "vec/columns/column_const.h"
-#include "vec/data_types/data_type_array.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/functions/function.h"
-#include "vec/functions/function_helpers.h"
+
+namespace doris {
+class FunctionContext;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/array/function_array_slice.h b/be/src/vec/functions/array/function_array_slice.h
index c80a6834d4..43a84d4653 100644
--- a/be/src/vec/functions/array/function_array_slice.h
+++ b/be/src/vec/functions/array/function_array_slice.h
@@ -16,16 +16,29 @@
 // under the License.
 #pragma once
 
-#include <string_view>
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stddef.h>
 
-#include "vec/columns/column_array.h"
-#include "vec/columns/column_string.h"
-#include "vec/common/string_ref.h"
-#include "vec/data_types/data_type_array.h"
-#include "vec/data_types/data_type_number.h"
+#include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/columns/column.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/functions/array/function_array_utils.h"
 #include "vec/functions/function.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 class FunctionArraySlice : public IFunction {
diff --git a/be/src/vec/functions/array/function_array_sort.cpp b/be/src/vec/functions/array/function_array_sort.cpp
index e1688d7471..13f590d519 100644
--- a/be/src/vec/functions/array/function_array_sort.cpp
+++ b/be/src/vec/functions/array/function_array_sort.cpp
@@ -17,6 +17,8 @@
 
 #include "vec/functions/array/function_array_sort.h"
 
+#include <algorithm>
+
 #include "vec/functions/simple_function_factory.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/functions/array/function_array_sort.h b/be/src/vec/functions/array/function_array_sort.h
index 78b6943c70..5178780d82 100644
--- a/be/src/vec/functions/array/function_array_sort.h
+++ b/be/src/vec/functions/array/function_array_sort.h
@@ -19,11 +19,27 @@
 // and modified by Doris
 #pragma once
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <memory>
+#include <ostream>
+
+#include "common/status.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
-#include "vec/columns/columns_number.h"
-#include "vec/data_types/data_type_array.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/functions/function.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 template <typename Name, bool Positive>
diff --git a/be/src/vec/functions/array/function_array_sortby.cpp b/be/src/vec/functions/array/function_array_sortby.cpp
index da0e0e4d49..a493a4ed88 100644
--- a/be/src/vec/functions/array/function_array_sortby.cpp
+++ b/be/src/vec/functions/array/function_array_sortby.cpp
@@ -15,14 +15,36 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <string>
+#include <glog/logging.h>
+#include <stddef.h>
 
+#include <algorithm>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
-#include "vec/data_types/data_type_array.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 class FunctionArraySortBy : public IFunction {
diff --git a/be/src/vec/functions/array/function_array_union.cpp b/be/src/vec/functions/array/function_array_union.cpp
index 990f621435..641063103c 100644
--- a/be/src/vec/functions/array/function_array_union.cpp
+++ b/be/src/vec/functions/array/function_array_union.cpp
@@ -15,6 +15,11 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include "vec/columns/column_const.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/string_ref.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/functions/array/function_array_binary.h"
 #include "vec/functions/array/function_array_set.h"
 #include "vec/functions/simple_function_factory.h"
diff --git a/be/src/vec/functions/array/function_array_utils.cpp b/be/src/vec/functions/array/function_array_utils.cpp
index ae5baac811..ab999aa21c 100644
--- a/be/src/vec/functions/array/function_array_utils.cpp
+++ b/be/src/vec/functions/array/function_array_utils.cpp
@@ -17,6 +17,15 @@
 
 #include "vec/functions/array/function_array_utils.h"
 
+#include <stddef.h>
+
+#include <algorithm>
+#include <utility>
+
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+
 namespace doris::vectorized {
 
 bool extract_column_array_info(const IColumn& src, ColumnArrayExecutionData& data) {
diff --git a/be/src/vec/functions/array/function_array_utils.h b/be/src/vec/functions/array/function_array_utils.h
index f488e30786..a9b303934c 100644
--- a/be/src/vec/functions/array/function_array_utils.h
+++ b/be/src/vec/functions/array/function_array_utils.h
@@ -17,8 +17,15 @@
 #pragma once
 
 #include "vec/columns/column_array.h"
-#include "vec/columns/column_nullable.h"
-#include "vec/data_types/data_type_array.h"
+#include "vec/columns/columns_number.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+
+namespace doris {
+namespace vectorized {
+class IColumn;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/array/function_array_with_constant.cpp b/be/src/vec/functions/array/function_array_with_constant.cpp
index bb8e789525..e1a0c4f15e 100644
--- a/be/src/vec/functions/array/function_array_with_constant.cpp
+++ b/be/src/vec/functions/array/function_array_with_constant.cpp
@@ -15,15 +15,39 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <memory>
+#include <utility>
+#include <vector>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/status.h"
+#include "runtime/thread_context.h"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column_array.h"
-#include "vec/columns/column_const.h"
-#include "vec/data_types/data_type_array.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type_array.h" // IWYU pragma: keep
 #include "vec/data_types/data_type_nullable.h"
-#include "vec/data_types/data_type_number.h"
 #include "vec/functions/function.h"
-#include "vec/functions/function_helpers.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+
+namespace vectorized {
+class Block;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 /* array_with_constant(num, T) / array_repeat(T, num)  - return array of constants with length num.
diff --git a/be/src/vec/functions/array/function_array_zip.cpp b/be/src/vec/functions/array/function_array_zip.cpp
index f07698e73d..7f82492ce8 100644
--- a/be/src/vec/functions/array/function_array_zip.cpp
+++ b/be/src/vec/functions/array/function_array_zip.cpp
@@ -18,13 +18,40 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Functions/array/arrayZip.cpp
 // and modified by Doris
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_array.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_struct.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_array.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_struct.h"
-#include "vec/functions/array/function_array_utils.h"
 #include "vec/functions/function.h"
 #include "vec/functions/function_helpers.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 // Combines multiple arrays into a single array
diff --git a/be/src/vec/functions/array/function_arrays_overlap.h b/be/src/vec/functions/array/function_arrays_overlap.h
index 4256dd3229..969773481a 100644
--- a/be/src/vec/functions/array/function_arrays_overlap.h
+++ b/be/src/vec/functions/array/function_arrays_overlap.h
@@ -16,17 +16,48 @@
 // under the License.
 #pragma once
 
-#include <string_view>
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <sys/types.h>
 
-#include "vec/columns/column_array.h"
-#include "vec/columns/column_string.h"
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <new>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/hash_table/hash.h"
 #include "vec/common/hash_table/hash_set.h"
 #include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_array.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/functions/array/function_array_utils.h"
 #include "vec/functions/function.h"
 
+namespace doris {
+class FunctionContext;
+
+namespace vectorized {
+class ColumnString;
+} // namespace vectorized
+} // namespace doris
+template <typename, typename>
+struct DefaultHash;
+
 namespace doris::vectorized {
 
 template <typename T>
diff --git a/be/src/vec/functions/comparison_equal_for_null.cpp b/be/src/vec/functions/comparison_equal_for_null.cpp
index dff52a2309..ae08523a59 100644
--- a/be/src/vec/functions/comparison_equal_for_null.cpp
+++ b/be/src/vec/functions/comparison_equal_for_null.cpp
@@ -15,12 +15,34 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include "common/compiler_util.h"
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_const.h"
 #include "vec/columns/column_nullable.h"
-#include "vec/data_types/get_least_supertype.h"
-#include "vec/functions/function_string.h"
+#include "vec/columns/column_vector.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/data_types/data_type_number.h"
+#include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
-#include "vec/utils/util.hpp"
+
+namespace doris {
+class FunctionContext;
+} // namespace doris
 
 namespace doris::vectorized {
 // Operator <=>
diff --git a/be/src/vec/functions/comparison_equals.cpp b/be/src/vec/functions/comparison_equals.cpp
index 087e5dae89..87bd8f6e33 100644
--- a/be/src/vec/functions/comparison_equals.cpp
+++ b/be/src/vec/functions/comparison_equals.cpp
@@ -15,6 +15,9 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include "vec/columns/column.h"
+#include "vec/core/accurate_comparison.h"
+#include "vec/data_types/data_type.h"
 #include "vec/functions/functions_comparison.h"
 #include "vec/functions/simple_function_factory.h"
 
diff --git a/be/src/vec/functions/comparison_greater.cpp b/be/src/vec/functions/comparison_greater.cpp
index a2c8e6fe06..f17380ca4b 100644
--- a/be/src/vec/functions/comparison_greater.cpp
+++ b/be/src/vec/functions/comparison_greater.cpp
@@ -15,6 +15,9 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include "vec/columns/column.h"
+#include "vec/core/accurate_comparison.h"
+#include "vec/data_types/data_type.h"
 #include "vec/functions/functions_comparison.h"
 #include "vec/functions/simple_function_factory.h"
 
diff --git a/be/src/vec/functions/comparison_less.cpp b/be/src/vec/functions/comparison_less.cpp
index 05e4a8c86d..f2e9196405 100644
--- a/be/src/vec/functions/comparison_less.cpp
+++ b/be/src/vec/functions/comparison_less.cpp
@@ -15,6 +15,9 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include "vec/columns/column.h"
+#include "vec/core/accurate_comparison.h"
+#include "vec/data_types/data_type.h"
 #include "vec/functions/functions_comparison.h"
 #include "vec/functions/simple_function_factory.h"
 
diff --git a/be/src/vec/functions/divide.cpp b/be/src/vec/functions/divide.cpp
index a602cb8b33..e9de80a121 100644
--- a/be/src/vec/functions/divide.cpp
+++ b/be/src/vec/functions/divide.cpp
@@ -18,6 +18,14 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Functions/divide.cpp
 // and modified by Doris
 
+#include <string.h>
+
+#include <utility>
+
+#include "runtime/decimalv2_value.h"
+#include "vec/columns/column_vector.h"
+#include "vec/core/types.h"
+#include "vec/data_types/number_traits.h"
 #include "vec/functions/function_binary_arithmetic.h"
 #include "vec/functions/simple_function_factory.h"
 
diff --git a/be/src/vec/functions/function.cpp b/be/src/vec/functions/function.cpp
index 753d84cec5..101365c2a8 100644
--- a/be/src/vec/functions/function.cpp
+++ b/be/src/vec/functions/function.cpp
@@ -20,13 +20,19 @@
 
 #include "vec/functions/function.h"
 
+#include <algorithm>
 #include <memory>
-#include <optional>
+#include <numeric>
+#include <vector>
 
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
 #include "vec/common/assert_cast.h"
-#include "vec/common/typeid_cast.h"
+#include "vec/core/field.h"
 #include "vec/data_types/data_type_array.h"
 #include "vec/data_types/data_type_nothing.h"
 #include "vec/data_types/data_type_nullable.h"
diff --git a/be/src/vec/functions/function.h b/be/src/vec/functions/function.h
index 94decc1a8e..4d233f6915 100644
--- a/be/src/vec/functions/function.h
+++ b/be/src/vec/functions/function.h
@@ -20,12 +20,24 @@
 
 #pragma once
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
 #include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
 
 #include "common/status.h"
+#include "udf/udf.h"
 #include "vec/core/block.h"
 #include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/function_bit.cpp b/be/src/vec/functions/function_bit.cpp
index 002341bcb2..3a56d5c0cd 100644
--- a/be/src/vec/functions/function_bit.cpp
+++ b/be/src/vec/functions/function_bit.cpp
@@ -18,6 +18,15 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Functions/bitAnd.cpp
 // and modified by Doris
 
+#include <utility>
+
+#include "common/status.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type_number.h"
 #include "vec/data_types/number_traits.h"
 #include "vec/functions/function_binary_arithmetic.h"
 #include "vec/functions/function_totype.h"
diff --git a/be/src/vec/functions/function_bitmap.cpp b/be/src/vec/functions/function_bitmap.cpp
index 8f81e8c2a6..680aea73dd 100644
--- a/be/src/vec/functions/function_bitmap.cpp
+++ b/be/src/vec/functions/function_bitmap.cpp
@@ -18,21 +18,63 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Functions/FunctionBitmap.h
 // and modified by Doris
 
+#include <glog/logging.h>
+#include <stdint.h>
+#include <string.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <functional>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
+#include <vector>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/status.h"
 #include "gutil/strings/numbers.h"
 #include "gutil/strings/split.h"
+#include "util/bitmap_value.h"
+#include "util/hash_util.hpp"
+#include "util/murmur_hash3.h"
 #include "util/string_parser.hpp"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
+#include "vec/columns/column_complex.h"
+#include "vec/columns/column_const.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_array.h"
+#include "vec/data_types/data_type_bitmap.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/data_types/data_type_string.h"
+#include "vec/functions/function.h"
 #include "vec/functions/function_always_not_nullable.h"
 #include "vec/functions/function_bitmap_min_or_max.h"
 #include "vec/functions/function_const.h"
+#include "vec/functions/function_helpers.h"
 #include "vec/functions/function_string.h"
 #include "vec/functions/function_totype.h"
 #include "vec/functions/simple_function_factory.h"
+#include "vec/utils/util.hpp"
+
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 struct BitmapEmpty {
diff --git a/be/src/vec/functions/function_bitmap_variadic.cpp b/be/src/vec/functions/function_bitmap_variadic.cpp
index fbd5942422..44b18d108a 100644
--- a/be/src/vec/functions/function_bitmap_variadic.cpp
+++ b/be/src/vec/functions/function_bitmap_variadic.cpp
@@ -15,14 +15,39 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <stddef.h>
+
+#include <algorithm>
+#include <functional>
+#include <memory>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
+#include "util/bitmap_value.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_complex.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_bitmap.h"
-#include "vec/functions/function_const.h"
-#include "vec/functions/function_string.h"
-#include "vec/functions/function_totype.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/data_types/data_type_number.h"
+#include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 // currently only bitmap_or and bitmap_or_count will call this function,
diff --git a/be/src/vec/functions/function_case.cpp b/be/src/vec/functions/function_case.cpp
index 6ce94ab5aa..8f0278b317 100644
--- a/be/src/vec/functions/function_case.cpp
+++ b/be/src/vec/functions/function_case.cpp
@@ -17,6 +17,8 @@
 
 #include "vec/functions/function_case.h"
 
+#include "vec/functions/simple_function_factory.h"
+
 namespace doris::vectorized {
 
 void register_function_case(SimpleFunctionFactory& factory) {
diff --git a/be/src/vec/functions/function_case.h b/be/src/vec/functions/function_case.h
index 0f59075e05..bd7e8f68a5 100644
--- a/be/src/vec/functions/function_case.h
+++ b/be/src/vec/functions/function_case.h
@@ -17,16 +17,42 @@
 
 #pragma once
 
-#include <cstdint>
+#include <string.h>
 
+#include <algorithm>
+#include <cstdint>
+#include <memory>
+#include <optional>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
+#include "gutil/integral_types.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_complex.h"
-#include "vec/columns/column_map.h"
-#include "vec/columns/column_struct.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/columns_number.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/functions/function.h"
-#include "vec/functions/simple_function_factory.h"
 #include "vec/utils/template_helpers.hpp"
 
+namespace doris {
+class FunctionContext;
+
+namespace vectorized {
+class ColumnArray;
+class ColumnMap;
+class ColumnString;
+class ColumnStruct;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 template <bool has_case, bool has_else>
diff --git a/be/src/vec/functions/function_cast.h b/be/src/vec/functions/function_cast.h
index 4aeb3b373f..7af176395a 100644
--- a/be/src/vec/functions/function_cast.h
+++ b/be/src/vec/functions/function_cast.h
@@ -21,28 +21,76 @@
 #pragma once
 
 #include <fmt/format.h>
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <cmath>
+#include <functional>
+#include <iterator>
+#include <limits>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/status.h"
 #include "udf/udf.h"
+#include "util/jsonb_document.h"
+#include "util/jsonb_stream.h"
+#include "util/jsonb_writer.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
-#include "vec/columns/column_const.h"
 #include "vec/columns/column_nullable.h"
 #include "vec/columns/column_string.h"
+#include "vec/columns/column_struct.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_common.h"
+#include "vec/columns/columns_number.h"
 #include "vec/common/assert_cast.h"
-#include "vec/common/field_visitors.h"
 #include "vec/common/string_buffer.hpp"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/call_on_type_index.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_array.h"
+#include "vec/data_types/data_type_date.h"
+#include "vec/data_types/data_type_date_time.h"
 #include "vec/data_types/data_type_decimal.h"
-#include "vec/data_types/data_type_factory.hpp"
 #include "vec/data_types/data_type_jsonb.h"
+#include "vec/data_types/data_type_map.h"
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/data_types/data_type_string.h"
+#include "vec/data_types/data_type_struct.h"
+#include "vec/data_types/data_type_time_v2.h"
 #include "vec/functions/function.h"
 #include "vec/functions/function_helpers.h"
 #include "vec/io/io_helper.h"
 #include "vec/io/reader_buffer.h"
 #include "vec/runtime/vdatetime_value.h"
 
+class DateLUTImpl;
+
+namespace doris {
+namespace vectorized {
+template <typename T>
+class ColumnDecimal;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 /** Type conversion functions.
   * toType - conversion in "natural way";
@@ -971,6 +1019,7 @@ using FunctionToDateTimeV2 =
 
 template <typename DataType>
 struct FunctionTo;
+
 template <>
 struct FunctionTo<DataTypeUInt8> {
     using Type = FunctionToUInt8;
diff --git a/be/src/vec/functions/function_coalesce.cpp b/be/src/vec/functions/function_coalesce.cpp
index 9c66dff293..800e1de854 100644
--- a/be/src/vec/functions/function_coalesce.cpp
+++ b/be/src/vec/functions/function_coalesce.cpp
@@ -15,11 +15,42 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include "udf/udf.h"
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
+#include "util/bitmap_value.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_complex.h"
+#include "vec/columns/column_decimal.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
+#include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 #include "vec/utils/template_helpers.hpp"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 class FunctionCoalesce : public IFunction {
 public:
diff --git a/be/src/vec/functions/function_conv.cpp b/be/src/vec/functions/function_conv.cpp
index 68f2070bea..bd49398c8d 100644
--- a/be/src/vec/functions/function_conv.cpp
+++ b/be/src/vec/functions/function_conv.cpp
@@ -15,14 +15,42 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <stdint.h>
+#include <stdlib.h>
+
+#include <boost/iterator/iterator_facade.hpp>
+// IWYU pragma: no_include <bits/std_abs.h>
+#include <algorithm>
+#include <cmath> // IWYU pragma: keep
+#include <memory>
+#include <utility>
+
+#include "common/status.h"
 #include "exprs/math_functions.h"
+#include "util/string_parser.hpp"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/data_types/data_type_string.h"
+#include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
-#include "vec/utils/util.hpp"
+
+namespace doris {
+class FunctionContext;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/function_convert_tz.cpp b/be/src/vec/functions/function_convert_tz.cpp
index 00745d1c6d..6f83f1bf1a 100644
--- a/be/src/vec/functions/function_convert_tz.cpp
+++ b/be/src/vec/functions/function_convert_tz.cpp
@@ -17,7 +17,9 @@
 
 #include "vec/functions/function_convert_tz.h"
 
+#include "vec/data_types/data_type_date.h"
 #include "vec/functions/simple_function_factory.h"
+#include "vec/runtime/vdatetime_value.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/function_convert_tz.h b/be/src/vec/functions/function_convert_tz.h
index 4a1d66b459..ae5723aef1 100644
--- a/be/src/vec/functions/function_convert_tz.h
+++ b/be/src/vec/functions/function_convert_tz.h
@@ -17,13 +17,50 @@
 
 #pragma once
 
+#include <cctz/time_zone.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <map>
+#include <memory>
+#include <string>
+#include <type_traits>
+#include <utility>
+
+#include "common/status.h"
+#include "udf/udf.h"
+#include "util/binary_cast.hpp"
+#include "util/timezone_utils.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
 #include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_date_time.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_string.h"
-#include "vec/utils/util.hpp"
+#include "vec/data_types/data_type_time_v2.h"
+#include "vec/functions/function.h"
+
+namespace doris {
+namespace vectorized {
+class DataTypeDate;
+class VecDateTimeValue;
+struct DateTimeV2ValueType;
+struct DateV2ValueType;
+template <typename T>
+class DateV2Value;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/function_date_or_datetime_computation.h b/be/src/vec/functions/function_date_or_datetime_computation.h
index 9aaaa21708..80163db45e 100644
--- a/be/src/vec/functions/function_date_or_datetime_computation.h
+++ b/be/src/vec/functions/function_date_or_datetime_computation.h
@@ -17,19 +17,43 @@
 
 #pragma once
 
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <type_traits>
+#include <utility>
+
 #include "common/logging.h"
+#include "common/status.h"
 #include "fmt/format.h"
-#include "runtime/datetime_value.h"
 #include "runtime/runtime_state.h"
 #include "udf/udf.h"
 #include "util/binary_cast.hpp"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
+#include "vec/columns/column_nullable.h"
 #include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/common/typeid_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/field.h"
 #include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_date.h"
 #include "vec/data_types/data_type_date_time.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/data_types/data_type_time.h"
+#include "vec/data_types/data_type_time_v2.h"
 #include "vec/functions/function.h"
 #include "vec/functions/function_helpers.h"
 #include "vec/runtime/vdatetime_value.h"
diff --git a/be/src/vec/functions/function_date_or_datetime_computation_v2.cpp b/be/src/vec/functions/function_date_or_datetime_computation_v2.cpp
index f3984be2b4..cca56895ee 100644
--- a/be/src/vec/functions/function_date_or_datetime_computation_v2.cpp
+++ b/be/src/vec/functions/function_date_or_datetime_computation_v2.cpp
@@ -15,6 +15,7 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include "vec/data_types/data_type_time_v2.h"
 #include "vec/functions/function_date_or_datetime_computation.h"
 #include "vec/functions/simple_function_factory.h"
 
diff --git a/be/src/vec/functions/function_date_or_datetime_to_string.cpp b/be/src/vec/functions/function_date_or_datetime_to_string.cpp
index cf215cd14d..b8f1c32066 100644
--- a/be/src/vec/functions/function_date_or_datetime_to_string.cpp
+++ b/be/src/vec/functions/function_date_or_datetime_to_string.cpp
@@ -20,6 +20,7 @@
 
 #include "vec/functions/function_date_or_datetime_to_string.h"
 
+#include "vec/data_types/data_type_time_v2.h" // IWYU pragma: keep
 #include "vec/functions/simple_function_factory.h"
 
 namespace doris::vectorized {
@@ -41,4 +42,4 @@ void register_function_date_time_to_string(SimpleFunctionFactory& factory) {
     factory.register_function<FunctionDateTimeV2MonthName>();
 }
 
-} // namespace doris::vectorized
\ No newline at end of file
+} // namespace doris::vectorized
diff --git a/be/src/vec/functions/function_date_or_datetime_to_string.h b/be/src/vec/functions/function_date_or_datetime_to_string.h
index a48e934abd..6782d9236a 100644
--- a/be/src/vec/functions/function_date_or_datetime_to_string.h
+++ b/be/src/vec/functions/function_date_or_datetime_to_string.h
@@ -20,9 +20,36 @@
 
 #pragma once
 
-#include "vec/data_types/data_type_string.h"
+#include <stddef.h>
+
+#include <memory>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/functions/date_time_transforms.h"
 #include "vec/functions/function.h"
+#include "vec/utils/util.hpp"
+
+namespace doris {
+class FunctionContext;
+
+namespace vectorized {
+class DataTypeString;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/function_datetime_floor_ceil.cpp b/be/src/vec/functions/function_datetime_floor_ceil.cpp
index 70ba43784d..7a6773f415 100644
--- a/be/src/vec/functions/function_datetime_floor_ceil.cpp
+++ b/be/src/vec/functions/function_datetime_floor_ceil.cpp
@@ -15,11 +15,59 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <type_traits>
+#include <utility>
+
+#include "common/status.h"
+#include "util/binary_cast.hpp"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
+#include "vec/columns/column_nullable.h"
 #include "vec/columns/column_vector.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/common/typeid_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_date_time.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
+#include "vec/data_types/data_type_time_v2.h"
+#include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
+#include "vec/runtime/vdatetime_value.h"
+
+namespace doris {
+class FunctionContext;
+
+namespace vectorized {
+struct DayCeil;
+struct DayFloor;
+struct HourCeil;
+struct HourFloor;
+struct MinuteCeil;
+struct MinuteFloor;
+struct MonthCeil;
+struct MonthFloor;
+struct SecondCeil;
+struct SecondFloor;
+struct WeekCeil;
+struct WeekFloor;
+struct YearCeil;
+struct YearFloor;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/function_datetime_string_to_string.cpp b/be/src/vec/functions/function_datetime_string_to_string.cpp
index 70c82306db..200ddee5d6 100644
--- a/be/src/vec/functions/function_datetime_string_to_string.cpp
+++ b/be/src/vec/functions/function_datetime_string_to_string.cpp
@@ -17,7 +17,9 @@
 
 #include "vec/functions/function_datetime_string_to_string.h"
 
+#include "vec/data_types/data_type_time_v2.h" // IWYU pragma: keep
 #include "vec/functions/simple_function_factory.h"
+#include "vec/runtime/vdatetime_value.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/function_datetime_string_to_string.h b/be/src/vec/functions/function_datetime_string_to_string.h
index a5cde61268..b16da8abd6 100644
--- a/be/src/vec/functions/function_datetime_string_to_string.h
+++ b/be/src/vec/functions/function_datetime_string_to_string.h
@@ -17,17 +17,34 @@
 
 #pragma once
 
-#include "vec/columns/column_const.h"
+#include <stddef.h>
+
+#include <memory>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
 #include "vec/common/string_ref.h"
-#include "vec/data_types/data_type_date.h"
-#include "vec/data_types/data_type_date_time.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_string.h"
 #include "vec/functions/date_time_transforms.h"
 #include "vec/functions/function.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 template <typename Transform>
diff --git a/be/src/vec/functions/function_encryption.cpp b/be/src/vec/functions/function_encryption.cpp
index e0d9edfce1..9b93172475 100644
--- a/be/src/vec/functions/function_encryption.cpp
+++ b/be/src/vec/functions/function_encryption.cpp
@@ -15,9 +15,42 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <memory>
+#include <string>
+#include <string_view>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
 #include "util/encryption_util.h"
+#include "util/string_util.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/pod_array.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/data_types/data_type_string.h"
+#include "vec/functions/function.h"
 #include "vec/functions/function_string.h"
 #include "vec/functions/simple_function_factory.h"
+#include "vec/utils/util.hpp"
+
+namespace doris {
+class FunctionContext;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/function_fake.cpp b/be/src/vec/functions/function_fake.cpp
index 7e6cc25d04..4fd626d17b 100644
--- a/be/src/vec/functions/function_fake.cpp
+++ b/be/src/vec/functions/function_fake.cpp
@@ -17,9 +17,19 @@
 
 #include "vec/functions/function_fake.h"
 
+#include <glog/logging.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <ostream>
+#include <string>
+
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/data_types/data_type_array.h"
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
+#include "vec/data_types/data_type_string.h"
+#include "vec/exprs/table_function/table_function.h"
 #include "vec/functions/function_helpers.h"
 #include "vec/functions/simple_function_factory.h"
 
diff --git a/be/src/vec/functions/function_fake.h b/be/src/vec/functions/function_fake.h
index 0f2c83aed0..456672e7a2 100644
--- a/be/src/vec/functions/function_fake.h
+++ b/be/src/vec/functions/function_fake.h
@@ -17,9 +17,23 @@
 
 #pragma once
 
+#include <stddef.h>
+
+#include <memory>
+
 #include "common/status.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/functions/function.h"
 
+namespace doris {
+class FunctionContext;
+namespace vectorized {
+class Block;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 // FunctionFake is use for some function call expr only work at prepare/open phase, do not support execute().
 template <typename Impl>
diff --git a/be/src/vec/functions/function_grouping.h b/be/src/vec/functions/function_grouping.h
index ae96a50a76..bf8c90a224 100644
--- a/be/src/vec/functions/function_grouping.h
+++ b/be/src/vec/functions/function_grouping.h
@@ -17,9 +17,29 @@
 
 #pragma once
 
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+
+#include "common/status.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/functions/function.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 class FunctionGroupingBase : public IFunction {
diff --git a/be/src/vec/functions/function_hash.cpp b/be/src/vec/functions/function_hash.cpp
index a5cc8e9d5d..ae3680400b 100644
--- a/be/src/vec/functions/function_hash.cpp
+++ b/be/src/vec/functions/function_hash.cpp
@@ -20,7 +20,26 @@
 
 #include "vec/functions/function_hash.h"
 
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <string>
+
+#include "common/status.h"
 #include "util/hash_util.hpp"
+#include "util/murmur_hash3.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_const.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/bit_cast.h"
+#include "vec/core/field.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_number.h"
+#include "vec/functions/function_helpers.h"
 #include "vec/functions/function_variadic_arguments.h"
 #include "vec/functions/simple_function_factory.h"
 #include "vec/utils/template_helpers.hpp"
diff --git a/be/src/vec/functions/function_hash.h b/be/src/vec/functions/function_hash.h
index 7ea034b4a5..84dda9bd1a 100644
--- a/be/src/vec/functions/function_hash.h
+++ b/be/src/vec/functions/function_hash.h
@@ -20,16 +20,8 @@
 
 #pragma once
 
-#include <utility>
-
-#include "vec/columns/column_string.h"
-#include "vec/columns/column_vector.h"
-#include "vec/common/bit_cast.h"
 #include "vec/common/hash_table/hash.h"
-#include "vec/data_types/data_type.h"
-#include "vec/data_types/data_type_number.h"
-#include "vec/functions/function.h"
-#include "vec/functions/function_helpers.h"
+#include "vec/core/types.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/function_helpers.cpp b/be/src/vec/functions/function_helpers.cpp
index c77f3c5ab7..5fc2a11e87 100644
--- a/be/src/vec/functions/function_helpers.cpp
+++ b/be/src/vec/functions/function_helpers.cpp
@@ -20,8 +20,20 @@
 
 #include "vec/functions/function_helpers.h"
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+
+#include <algorithm>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <vector>
+
 #include "common/consts.h"
+#include "util/string_util.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/functions/function.h"
 
diff --git a/be/src/vec/functions/function_helpers.h b/be/src/vec/functions/function_helpers.h
index 85b4fcea48..8ac921d9a3 100644
--- a/be/src/vec/functions/function_helpers.h
+++ b/be/src/vec/functions/function_helpers.h
@@ -20,13 +20,19 @@
 
 #pragma once
 
+#include <stddef.h>
+
+#include <tuple>
+#include <type_traits>
+
 #include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
 #include "vec/common/assert_cast.h"
 #include "vec/common/typeid_cast.h"
 #include "vec/core/block.h"
-#include "vec/core/call_on_type_index.h"
 #include "vec/core/column_numbers.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/functions/function_hex.cpp b/be/src/vec/functions/function_hex.cpp
index 4fd69581c3..2f11072183 100644
--- a/be/src/vec/functions/function_hex.cpp
+++ b/be/src/vec/functions/function_hex.cpp
@@ -15,13 +15,40 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <string>
+#include <string_view>
+#include <utility>
+
+#include "common/status.h"
+#include "olap/hll.h"
 #include "util/simd/vstring_function.h" //place this header file at last to compile
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_complex.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_hll.h"
+#include "vec/data_types/data_type_number.h"
+#include "vec/data_types/data_type_string.h"
+#include "vec/functions/function.h"
 #include "vec/functions/function_string.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 template <typename Impl>
 class FunctionHexVariadic : public IFunction {
diff --git a/be/src/vec/functions/function_ifnull.h b/be/src/vec/functions/function_ifnull.h
index 3378afae84..5ecf4d153e 100644
--- a/be/src/vec/functions/function_ifnull.h
+++ b/be/src/vec/functions/function_ifnull.h
@@ -20,12 +20,30 @@
 
 #pragma once
 
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_nullable.h"
-#include "vec/data_types/get_least_supertype.h"
-#include "vec/functions/function_helpers.h"
-#include "vec/functions/function_string.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/data_types/data_type_number.h"
+#include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
-#include "vec/utils/util.hpp"
+
+namespace doris {
+class FunctionContext;
+} // namespace doris
 
 namespace doris::vectorized {
 class FunctionIfNull : public IFunction {
diff --git a/be/src/vec/functions/function_java_udf.cpp b/be/src/vec/functions/function_java_udf.cpp
index 0fadae2e61..fe96c92051 100644
--- a/be/src/vec/functions/function_java_udf.cpp
+++ b/be/src/vec/functions/function_java_udf.cpp
@@ -17,27 +17,26 @@
 
 #include "vec/functions/function_java_udf.h"
 
-#include <fmt/format.h>
-
 #include <memory>
-#include <sstream>
+#include <string>
+#include <utility>
+#include <vector>
 
-#include "gen_cpp/Exprs_types.h"
 #include "gutil/strings/substitute.h"
-#include "runtime/exec_env.h"
+#include "jni.h"
+#include "jni_md.h"
 #include "runtime/user_function_cache.h"
 #include "util/jni-util.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
 #include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/string_ref.h"
 #include "vec/core/block.h"
-#include "vec/data_types/data_type_bitmap.h"
-#include "vec/data_types/data_type_date.h"
-#include "vec/data_types/data_type_date_time.h"
-#include "vec/data_types/data_type_decimal.h"
 #include "vec/data_types/data_type_nullable.h"
-#include "vec/data_types/data_type_number.h"
-#include "vec/data_types/data_type_string.h"
 
 const char* EXECUTOR_CLASS = "org/apache/doris/udf/UdfExecutor";
 const char* EXECUTOR_CTOR_SIGNATURE = "([B)V";
diff --git a/be/src/vec/functions/function_java_udf.h b/be/src/vec/functions/function_java_udf.h
index 9eb16c9e1c..b04097c24d 100644
--- a/be/src/vec/functions/function_java_udf.h
+++ b/be/src/vec/functions/function_java_udf.h
@@ -17,10 +17,24 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
 #include <jni.h>
+#include <stddef.h>
+#include <stdint.h>
 
-#include "gen_cpp/Exprs_types.h"
+#include <memory>
+#include <ostream>
+
+#include "common/logging.h"
+#include "common/status.h"
+#include "udf/udf.h"
 #include "util/jni-util.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/functions/function.h"
 
 namespace doris {
diff --git a/be/src/vec/functions/function_json.cpp b/be/src/vec/functions/function_json.cpp
index 3ee87c59b9..9e40b56aa3 100644
--- a/be/src/vec/functions/function_json.cpp
+++ b/be/src/vec/functions/function_json.cpp
@@ -15,30 +15,61 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <rapidjson/allocators.h>
 #include <rapidjson/document.h>
+#include <rapidjson/encodings.h>
+#include <rapidjson/rapidjson.h>
 #include <rapidjson/stringbuffer.h>
 #include <rapidjson/writer.h>
+#include <re2/re2.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
 
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
 #include <boost/token_functions.hpp>
+#include <boost/tokenizer.hpp>
+#include <memory>
+#include <string>
 #include <string_view>
+#include <type_traits>
+#include <utility>
 #include <vector>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/status.h"
 #include "exprs/json_functions.h"
+#include "util/jsonb_parser_simd.h"
 #include "util/string_parser.hpp"
 #include "util/string_util.h"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_nullable.h"
 #include "vec/columns/column_string.h"
 #include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
 #include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/data_types/data_type_string.h"
+#include "vec/functions/function.h"
 #include "vec/functions/function_string.h"
 #include "vec/functions/function_totype.h"
 #include "vec/functions/simple_function_factory.h"
 #include "vec/utils/template_helpers.hpp"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 static const re2::RE2 JSON_PATTERN("^([^\\\"\\[\\]]*)(?:\\[([0-9]+|\\*)\\])?");
 
diff --git a/be/src/vec/functions/function_jsonb.cpp b/be/src/vec/functions/function_jsonb.cpp
index 024c390072..8be6536b56 100644
--- a/be/src/vec/functions/function_jsonb.cpp
+++ b/be/src/vec/functions/function_jsonb.cpp
@@ -15,24 +15,50 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <boost/token_functions.hpp>
-#include <vector>
+#include <glog/logging.h>
+#include <simdjson/simdjson.h> // IWYU pragma: keep
+#include <stddef.h>
+#include <stdint.h>
 
-#include "common/compiler_util.h"
-#include "util/string_parser.hpp"
-#include "util/string_util.h"
+#include <memory>
+#include <ostream>
+#include <string>
+#include <string_view>
+#include <tuple>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/status.h"
+#include "udf/udf.h"
+#include "util/jsonb_document.h"
+#include "util/jsonb_error.h"
+#include "util/jsonb_parser_simd.h"
+#include "util/jsonb_stream.h"
+#include "util/jsonb_utils.h"
+#include "util/jsonb_writer.h"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column.h"
+#include "vec/columns/column_const.h"
 #include "vec/columns/column_nullable.h"
 #include "vec/columns/column_string.h"
 #include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
 #include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_jsonb.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/data_types/data_type_string.h"
+#include "vec/functions/function.h"
 #include "vec/functions/function_string.h"
-#include "vec/functions/function_totype.h"
 #include "vec/functions/simple_function_factory.h"
-#include "vec/utils/template_helpers.hpp"
+#include "vec/utils/util.hpp"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/function_map.cpp b/be/src/vec/functions/function_map.cpp
index a192db6c09..5206b4f078 100644
--- a/be/src/vec/functions/function_map.cpp
+++ b/be/src/vec/functions/function_map.cpp
@@ -15,11 +15,32 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <memory>
+#include <glog/logging.h>
+#include <stddef.h>
 
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <tuple>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
 #include "vec/columns/column_const.h"
 #include "vec/columns/column_map.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/typeid_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_array.h"
 #include "vec/data_types/data_type_map.h"
@@ -28,9 +49,12 @@
 #include "vec/data_types/get_least_supertype.h"
 #include "vec/functions/array/function_array_index.h"
 #include "vec/functions/function.h"
-#include "vec/functions/function_helpers.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 // construct a map
diff --git a/be/src/vec/functions/function_math_unary.h b/be/src/vec/functions/function_math_unary.h
index 455d1f1e02..3db8016047 100644
--- a/be/src/vec/functions/function_math_unary.h
+++ b/be/src/vec/functions/function_math_unary.h
@@ -22,6 +22,7 @@
 
 #include "vec/columns/column_decimal.h"
 #include "vec/columns/columns_number.h"
+#include "vec/core/call_on_type_index.h"
 #include "vec/data_types/data_type_decimal.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/functions/function.h"
@@ -164,4 +165,4 @@ struct UnaryFunctionPlain {
 
 #define UnaryFunctionVectorized UnaryFunctionPlain
 
-} // namespace doris::vectorized
\ No newline at end of file
+} // namespace doris::vectorized
diff --git a/be/src/vec/functions/function_nonnullable.cpp b/be/src/vec/functions/function_nonnullable.cpp
index e026be571e..ebebbcf36b 100644
--- a/be/src/vec/functions/function_nonnullable.cpp
+++ b/be/src/vec/functions/function_nonnullable.cpp
@@ -18,18 +18,26 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Functions/Ifnull.h
 // and modified by Doris
 
-#include "common/logging.h"
+#include <stddef.h>
+
+#include <memory>
+
 #include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
-#include "vec/data_types/data_type_array.h"
 #include "vec/data_types/data_type_nullable.h"
-#include "vec/data_types/get_least_supertype.h"
-#include "vec/functions/function_helpers.h"
-#include "vec/functions/function_string.h"
+#include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
-#include "vec/utils/util.hpp"
+
+namespace doris {
+class FunctionContext;
+} // namespace doris
 
 namespace doris::vectorized {
 class FunctionNonNullable : public IFunction {
diff --git a/be/src/vec/functions/function_quantile_state.cpp b/be/src/vec/functions/function_quantile_state.cpp
index 3cc3166ec8..7ab9c908d6 100644
--- a/be/src/vec/functions/function_quantile_state.cpp
+++ b/be/src/vec/functions/function_quantile_state.cpp
@@ -18,16 +18,48 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Functions/FunctionBitmap.h
 // and modified by Doris
 
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/status.h"
+#include "util/quantile_state.h"
 #include "util/string_parser.hpp"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column.h"
-#include "vec/columns/column_array.h"
+#include "vec/columns/column_complex.h"
+#include "vec/columns/column_const.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
-#include "vec/data_types/data_type_quantilestate.h"
-#include "vec/functions/function_always_not_nullable.h"
+#include "vec/data_types/data_type_quantilestate.h" // IWYU pragma: keep
+#include "vec/functions/function.h"
 #include "vec/functions/function_const.h"
-#include "vec/functions/function_string.h"
+#include "vec/functions/function_helpers.h"
 #include "vec/functions/simple_function_factory.h"
+#include "vec/utils/util.hpp"
+
+namespace doris {
+class FunctionContext;
+} // namespace doris
 
 namespace doris::vectorized {
 
@@ -274,4 +306,4 @@ void register_function_quantile_state(SimpleFunctionFactory& factory) {
     factory.register_function<FunctionToQuantileStateDouble>();
 }
 
-} // namespace doris::vectorized
\ No newline at end of file
+} // namespace doris::vectorized
diff --git a/be/src/vec/functions/function_regexp.cpp b/be/src/vec/functions/function_regexp.cpp
index c9061d3c61..730761594a 100644
--- a/be/src/vec/functions/function_regexp.cpp
+++ b/be/src/vec/functions/function_regexp.cpp
@@ -15,22 +15,40 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <glog/logging.h>
 #include <re2/re2.h>
+#include <re2/stringpiece.h>
+#include <stddef.h>
 
-#include <random>
+#include <algorithm>
+#include <memory>
+#include <string>
+#include <string_view>
+#include <utility>
+#include <vector>
 
+#include "common/status.h"
 #include "exprs/string_functions.h"
 #include "udf/udf.h"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
+#include "vec/columns/column_nullable.h"
 #include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
 #include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/core/types.h"
-#include "vec/data_types/data_type_number.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_string.h"
+#include "vec/functions/function.h"
 #include "vec/functions/function_string.h"
 #include "vec/functions/simple_function_factory.h"
-#include "vec/utils/util.hpp"
+
 namespace doris::vectorized {
 
 struct RegexpReplaceImpl {
diff --git a/be/src/vec/functions/function_rpc.cpp b/be/src/vec/functions/function_rpc.cpp
index 9a5f71b8c5..dbd92eeec2 100644
--- a/be/src/vec/functions/function_rpc.cpp
+++ b/be/src/vec/functions/function_rpc.cpp
@@ -17,14 +17,19 @@
 
 #include "vec/functions/function_rpc.h"
 
+#include <brpc/controller.h>
 #include <fmt/format.h>
+#include <gen_cpp/function_service.pb.h>
+#include <gen_cpp/types.pb.h>
 
+#include <algorithm>
 #include <memory>
+#include <utility>
 
-#include "gen_cpp/Exprs_types.h"
-#include "json2pb/json_to_pb.h"
-#include "json2pb/pb_to_json.h"
 #include "runtime/exec_env.h"
+#include "util/brpc_client_cache.h"
+#include "vec/columns/column.h"
+#include "vec/data_types/serde/data_type_serde.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/function_rpc.h b/be/src/vec/functions/function_rpc.h
index 0c7fda85b4..ec7d97677a 100644
--- a/be/src/vec/functions/function_rpc.h
+++ b/be/src/vec/functions/function_rpc.h
@@ -17,10 +17,30 @@
 
 #pragma once
 
-#include "gen_cpp/function_service.pb.h"
-#include "util/brpc_client_cache.h"
+#include <fmt/format.h>
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "common/status.h"
+#include "udf/udf.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/functions/function.h"
 
+namespace doris {
+class PFunctionCallRequest;
+class PFunctionService_Stub;
+class PValues;
+} // namespace doris
+
 namespace doris::vectorized {
 
 class RPCFnImpl {
diff --git a/be/src/vec/functions/function_running_difference.h b/be/src/vec/functions/function_running_difference.h
index cfdf77cda9..ba38f66fd0 100644
--- a/be/src/vec/functions/function_running_difference.h
+++ b/be/src/vec/functions/function_running_difference.h
@@ -17,26 +17,35 @@
 
 #pragma once
 
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <string>
+#include <utility>
+
 #include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column.h"
-#include "vec/columns/column_array.h"
-#include "vec/columns/column_decimal.h"
-#include "vec/columns/column_nullable.h"
-#include "vec/columns/column_string.h"
-#include "vec/columns/columns_number.h"
-#include "vec/common/assert_cast.h"
-#include "vec/common/typeid_cast.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
-#include "vec/data_types/data_type_date.h"
-#include "vec/data_types/data_type_date_time.h"
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/data_types/data_type_time.h"
-#include "vec/data_types/data_type_time_v2.h"
 #include "vec/data_types/number_traits.h"
 #include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 class FunctionRunningDifference : public IFunction {
diff --git a/be/src/vec/functions/function_string.cpp b/be/src/vec/functions/function_string.cpp
index c6599b2e3e..d87969fb76 100644
--- a/be/src/vec/functions/function_string.cpp
+++ b/be/src/vec/functions/function_string.cpp
@@ -17,10 +17,10 @@
 
 #include "vec/functions/function_string.h"
 
-#include <re2/re2.h>
+#include <ctype.h>
+#include <math.h>
+#include <re2/stringpiece.h>
 
-#include <cstddef>
-#include <cstdlib>
 #include <string_view>
 
 #include "runtime/string_search.hpp"
diff --git a/be/src/vec/functions/function_string.h b/be/src/vec/functions/function_string.h
index cb28fa7b82..8f9ba90ae8 100644
--- a/be/src/vec/functions/function_string.h
+++ b/be/src/vec/functions/function_string.h
@@ -17,34 +17,66 @@
 
 #pragma once
 
-#include <iconv.h>
-#include <stddef.h>
+#include <glog/logging.h>
+#include <limits.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/types.h>
 
 #include <algorithm>
+#include <array>
+#include <boost/iterator/iterator_facade.hpp>
+#include <cstddef>
 #include <memory>
+#include <ostream>
+#include <tuple>
+#include <utility>
+#include <vector>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/status.h"
+#include "gutil/strings/numbers.h"
+#include "gutil/strings/substitute.h"
+#include "runtime/decimalv2_value.h"
+#include "runtime/runtime_state.h"
+#include "runtime/string_search.hpp"
 #include "util/string_util.h"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
+#include "vec/columns/column_vector.h"
+#include "vec/common/int_exp.h"
+#include "vec/common/memcmp_small.h"
+#include "vec/common/memcpy_small.h"
+#include "vec/common/pod_array.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/common/typeid_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/io/io_helper.h"
 #ifndef USE_LIBCPP
 #include <memory_resource>
+
 #define PMR std::pmr
 #else
 #include <boost/container/pmr/monotonic_buffer_resource.hpp>
 #include <boost/container/pmr/vector.hpp>
+
 #define PMR boost::container::pmr
 #endif
 
-#include <fmt/core.h>
 #include <fmt/format.h>
-#include <fmt/ranges.h>
 
 #include <cstdint>
 #include <string>
 #include <string_view>
 
 #include "exprs/math_functions.h"
-#include "exprs/string_functions.h"
 #include "udf/udf.h"
 #include "util/md5.h"
 #include "util/simd/vstring_function.h"
@@ -315,6 +347,7 @@ struct Substr2Impl {
 
 template <bool Reverse>
 class FunctionMaskPartial;
+
 class FunctionMask : public IFunction {
 public:
     static constexpr auto name = "mask";
diff --git a/be/src/vec/functions/function_struct.cpp b/be/src/vec/functions/function_struct.cpp
index d8cac8406b..0343d3b616 100644
--- a/be/src/vec/functions/function_struct.cpp
+++ b/be/src/vec/functions/function_struct.cpp
@@ -15,16 +15,36 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include "vec/columns/column_const.h"
-#include "vec/columns/column_string.h"
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_nullable.h"
 #include "vec/columns/column_struct.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/typeid_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_struct.h"
 #include "vec/functions/function.h"
-#include "vec/functions/function_helpers.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 // construct a struct
diff --git a/be/src/vec/functions/function_timestamp.cpp b/be/src/vec/functions/function_timestamp.cpp
index 2122c2b09d..03a15779bf 100644
--- a/be/src/vec/functions/function_timestamp.cpp
+++ b/be/src/vec/functions/function_timestamp.cpp
@@ -15,17 +15,48 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <glog/logging.h>
+#include <limits.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <cstring>
+#include <memory>
+#include <tuple>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
+#include "runtime/define_primitive_type.h"
 #include "runtime/runtime_state.h"
+#include "runtime/types.h"
 #include "udf/udf.h"
+#include "util/binary_cast.hpp"
+#include "util/time_lut.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
 #include "vec/columns/column_nullable.h"
 #include "vec/columns/column_string.h"
 #include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/pod_array_fwd.h"
 #include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_date.h"
 #include "vec/data_types/data_type_date_time.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/data_types/data_type_string.h"
+#include "vec/data_types/data_type_time_v2.h"
+#include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 #include "vec/runtime/vdatetime_value.h"
 #include "vec/utils/util.hpp"
diff --git a/be/src/vec/functions/function_utility.cpp b/be/src/vec/functions/function_utility.cpp
index 886ed762c1..f6c22e5f2e 100644
--- a/be/src/vec/functions/function_utility.cpp
+++ b/be/src/vec/functions/function_utility.cpp
@@ -14,13 +14,42 @@
 // KIND, either express or implied.  See the License for the
 // specific language governing permissions and limitations
 // under the License.
-#include <thread>
+#include <stddef.h>
 
+#include <boost/iterator/iterator_facade.hpp>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <algorithm>
+#include <chrono> // IWYU pragma: keep
+#include <memory>
+#include <string>
+#include <thread>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/data_types/data_type_string.h"
+#include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 class FunctionSleep : public IFunction {
diff --git a/be/src/vec/functions/function_width_bucket.cpp b/be/src/vec/functions/function_width_bucket.cpp
index 0dbf554fe4..20f62b65c4 100644
--- a/be/src/vec/functions/function_width_bucket.cpp
+++ b/be/src/vec/functions/function_width_bucket.cpp
@@ -15,13 +15,33 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <stddef.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/functions/function.h"
-#include "vec/functions/function_helpers.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 class FunctionWidthBucket : public IFunction {
 public:
diff --git a/be/src/vec/functions/functions_geo.cpp b/be/src/vec/functions/functions_geo.cpp
index 2f91ef2403..1b3f34cf7d 100644
--- a/be/src/vec/functions/functions_geo.cpp
+++ b/be/src/vec/functions/functions_geo.cpp
@@ -17,9 +17,21 @@
 
 #include "vec/functions/functions_geo.h"
 
+#include <glog/logging.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <utility>
+
+#include "geo/geo_common.h"
 #include "geo/geo_types.h"
-#include "gutil/strings/substitute.h"
-#include "vec/columns/column_const.h"
+#include "vec/columns/column.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/field.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/data_types/data_type_number.h"
 #include "vec/functions/simple_function_factory.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/functions/functions_geo.h b/be/src/vec/functions/functions_geo.h
index 30266840c3..6f5158968d 100644
--- a/be/src/vec/functions/functions_geo.h
+++ b/be/src/vec/functions/functions_geo.h
@@ -17,13 +17,28 @@
 
 #pragma once
 
-#include "geo/geo_types.h"
+#include <stddef.h>
+
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "common/status.h"
 #include "udf/udf.h"
 #include "vec/core/column_numbers.h"
-#include "vec/data_types/data_type_number.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_string.h"
 #include "vec/functions/function.h"
 
+namespace doris {
+class GeoShape;
+
+namespace vectorized {
+class Block;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 struct StConstructState {
diff --git a/be/src/vec/functions/functions_logical.cpp b/be/src/vec/functions/functions_logical.cpp
index f5a1f0af09..af9deba563 100644
--- a/be/src/vec/functions/functions_logical.cpp
+++ b/be/src/vec/functions/functions_logical.cpp
@@ -20,20 +20,32 @@
 
 #include "vec/functions/functions_logical.h"
 
-#include <algorithm>
+#include <glog/logging.h>
 
+#include <utility>
+#include <vector>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "gutil/integral_types.h"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
 #include "vec/columns/column_nullable.h"
 #include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
-#include "vec/common/field_visitors.h"
-#include "vec/common/typeid_cast.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
-#include "vec/functions/function_helpers.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 namespace {
diff --git a/be/src/vec/functions/functions_logical.h b/be/src/vec/functions/functions_logical.h
index 93e4f47599..e3c68183d3 100644
--- a/be/src/vec/functions/functions_logical.h
+++ b/be/src/vec/functions/functions_logical.h
@@ -20,12 +20,27 @@
 
 #pragma once
 
-#include <type_traits>
+#include <fmt/format.h>
+#include <stddef.h>
 
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+
+#include "common/status.h"
+#include "vec/core/column_numbers.h"
 #include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
 #include "vec/functions/function.h"
 
+namespace doris {
+class FunctionContext;
+
+namespace vectorized {
+class Block;
+} // namespace vectorized
+} // namespace doris
+
 /** Logical functions AND, OR and NOT support three-valued (or ternary) logic
   * https://en.wikibooks.org/wiki/Structured_Query_Language/NULLs_and_the_Three_Valued_Logic
   *
diff --git a/be/src/vec/functions/functions_multi_string_position.cpp b/be/src/vec/functions/functions_multi_string_position.cpp
index 5b8dd13b9b..062f34f995 100644
--- a/be/src/vec/functions/functions_multi_string_position.cpp
+++ b/be/src/vec/functions/functions_multi_string_position.cpp
@@ -18,22 +18,46 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Functions/FunctionsMultiStringPosition.h
 // and modified by Doris
 
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
 #include <cstdint>
 #include <iterator>
+#include <limits>
+#include <memory>
+#include <utility>
+#include <vector>
 
+#include "common/status.h"
 #include "function.h"
 #include "function_helpers.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
-#include "vec/columns/column_fixed_length_object.h"
+#include "vec/columns/column_const.h"
+#include "vec/columns/column_nullable.h"
 #include "vec/columns/column_string.h"
 #include "vec/columns/column_vector.h"
-#include "vec/common/pod_array.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/common/string_ref.h"
 #include "vec/common/string_searcher.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_array.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
-#include "vec/data_types/data_type_string.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 template <typename Impl>
diff --git a/be/src/vec/functions/functions_multi_string_search.cpp b/be/src/vec/functions/functions_multi_string_search.cpp
index 33f2211413..5c19fe6c8b 100644
--- a/be/src/vec/functions/functions_multi_string_search.cpp
+++ b/be/src/vec/functions/functions_multi_string_search.cpp
@@ -18,22 +18,44 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Functions/FunctionsMultiStringSearch.h
 // and modified by Doris
 
-#include <hs/hs.h>
+#include <hs/hs_common.h>
+#include <hs/hs_runtime.h>
+#include <stddef.h>
 
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <limits>
+#include <memory>
+#include <optional>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
 #include "function.h"
 #include "function_helpers.h"
 #include "regexps.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
 #include "vec/columns/column_const.h"
-#include "vec/columns/column_fixed_length_object.h"
+#include "vec/columns/column_nullable.h"
 #include "vec/columns/column_string.h"
 #include "vec/columns/column_vector.h"
-#include "vec/common/pod_array.h"
-#include "vec/data_types/data_type_array.h"
-#include "vec/data_types/data_type_number.h"
-#include "vec/data_types/data_type_string.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_number.h" // IWYU pragma: keep
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 template <typename Impl>
@@ -280,4 +302,4 @@ void register_function_multi_string_search(SimpleFunctionFactory& factory) {
     factory.register_function<FunctionMultiMatchAny>();
 }
 
-} // namespace doris::vectorized
\ No newline at end of file
+} // namespace doris::vectorized
diff --git a/be/src/vec/functions/hll_cardinality.cpp b/be/src/vec/functions/hll_cardinality.cpp
index 02a5982890..bd2a8a1363 100644
--- a/be/src/vec/functions/hll_cardinality.cpp
+++ b/be/src/vec/functions/hll_cardinality.cpp
@@ -15,12 +15,35 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include "udf/udf.h"
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
+#include "olap/hll.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_complex.h"
-#include "vec/data_types/number_traits.h"
-#include "vec/functions/function_always_not_nullable.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_number.h"
+#include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 struct HLLCardinality {
diff --git a/be/src/vec/functions/hll_empty.cpp b/be/src/vec/functions/hll_empty.cpp
index a62a4d81bf..8a4111be5f 100644
--- a/be/src/vec/functions/hll_empty.cpp
+++ b/be/src/vec/functions/hll_empty.cpp
@@ -15,12 +15,16 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <algorithm>
+#include <memory>
+#include <utility>
+
 #include "olap/hll.h"
 #include "vec/columns/column_complex.h"
 #include "vec/data_types/data_type_hll.h"
-#include "vec/data_types/data_type_string.h"
 #include "vec/functions/function_const.h"
 #include "vec/functions/simple_function_factory.h"
+
 namespace doris::vectorized {
 
 struct HLLEmptyImpl {
diff --git a/be/src/vec/functions/hll_hash.cpp b/be/src/vec/functions/hll_hash.cpp
index 715014dd5d..bfd4612f95 100644
--- a/be/src/vec/functions/hll_hash.cpp
+++ b/be/src/vec/functions/hll_hash.cpp
@@ -15,7 +15,15 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <stddef.h>
+#include <stdint.h>
+
 #include "olap/hll.h"
+#include "util/hash_util.hpp"
+#include "vec/columns/column_complex.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_hll.h"
 #include "vec/functions/function_always_not_nullable.h"
 #include "vec/functions/simple_function_factory.h"
diff --git a/be/src/vec/functions/if.cpp b/be/src/vec/functions/if.cpp
index a8eec9233e..a4e86f8e51 100644
--- a/be/src/vec/functions/if.cpp
+++ b/be/src/vec/functions/if.cpp
@@ -18,14 +18,49 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Functions/If.cpp
 // and modified by Doris
 
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <ostream>
+#include <type_traits>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_const.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/common/typeid_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/call_on_type_index.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/data_types/get_least_supertype.h"
-#include "vec/data_types/number_traits.h"
+#include "vec/functions/function.h"
 #include "vec/functions/function_helpers.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+
+namespace vectorized {
+namespace NumberTraits {
+struct Error;
+} // namespace NumberTraits
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 template <typename A, typename B, typename ResultType>
diff --git a/be/src/vec/functions/in.cpp b/be/src/vec/functions/in.cpp
index b0bc1d137f..a97387351c 100644
--- a/be/src/vec/functions/in.cpp
+++ b/be/src/vec/functions/in.cpp
@@ -16,7 +16,7 @@
 // under the License.
 // This file is copied from
 
-#include <vec/functions/in.h>
+#include "vec/functions/in.h"
 
 #include "vec/functions/simple_function_factory.h"
 
diff --git a/be/src/vec/functions/in.h b/be/src/vec/functions/in.h
index 2599dd3e4f..3cd0884605 100644
--- a/be/src/vec/functions/in.h
+++ b/be/src/vec/functions/in.h
@@ -18,18 +18,43 @@
 
 #pragma once
 
-#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stddef.h>
 
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
 #include "exprs/create_predicate_function.h"
+#include "exprs/hybrid_set.h"
+#include "runtime/define_primitive_type.h"
+#include "runtime/types.h"
+#include "udf/udf.h"
+#include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/functions/function.h"
 
+namespace doris {
+namespace vectorized {
+class ColumnString;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 
 struct InState {
diff --git a/be/src/vec/functions/int_div.cpp b/be/src/vec/functions/int_div.cpp
index 59b2d1c2dc..9696bebf72 100644
--- a/be/src/vec/functions/int_div.cpp
+++ b/be/src/vec/functions/int_div.cpp
@@ -20,6 +20,8 @@
 
 #include "vec/functions/int_div.h"
 
+#include <utility>
+
 #include "vec/functions/function_binary_arithmetic.h"
 #include "vec/functions/simple_function_factory.h"
 
diff --git a/be/src/vec/functions/int_div.h b/be/src/vec/functions/int_div.h
index 4bfd0f661b..605d6d8339 100644
--- a/be/src/vec/functions/int_div.h
+++ b/be/src/vec/functions/int_div.h
@@ -21,14 +21,11 @@
 #pragma once
 
 #include <libdivide.h>
+#include <string.h>
 
-#include <type_traits>
-
-#include "vec/columns/column_decimal.h"
-#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
 #include "vec/core/types.h"
 #include "vec/data_types/number_traits.h"
-#include "vec/functions/function_binary_arithmetic.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/is_not_null.cpp b/be/src/vec/functions/is_not_null.cpp
index 346dac1b2c..fecfeb0e80 100644
--- a/be/src/vec/functions/is_not_null.cpp
+++ b/be/src/vec/functions/is_not_null.cpp
@@ -18,10 +18,33 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Functions/IsNotNull.cpp
 // and modified by Doris
 
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/field.h"
 #include "vec/data_types/data_type_number.h"
+#include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 /// Implements the function isNotNull which returns true if a value
diff --git a/be/src/vec/functions/is_null.cpp b/be/src/vec/functions/is_null.cpp
index 743a271f1b..37c8da647d 100644
--- a/be/src/vec/functions/is_null.cpp
+++ b/be/src/vec/functions/is_null.cpp
@@ -18,10 +18,29 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Functions/IsNull.cpp
 // and modified by Doris
 
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <string>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/field.h"
 #include "vec/data_types/data_type_number.h"
+#include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 /// Implements the function is_null which returns true if a value
diff --git a/be/src/vec/functions/least_greast.cpp b/be/src/vec/functions/least_greast.cpp
index 3d841e2da7..4358a19bb3 100644
--- a/be/src/vec/functions/least_greast.cpp
+++ b/be/src/vec/functions/least_greast.cpp
@@ -15,18 +15,38 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <tuple>
+#include <type_traits>
+
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
+#include "vec/columns/column_decimal.h"
+#include "vec/columns/column_string.h"
 #include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/common/string_ref.h"
 #include "vec/core/accurate_comparison.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_number.h"
-#include "vec/functions/function_helpers.h"
 #include "vec/functions/function_multi_same_args.h"
 #include "vec/functions/simple_function_factory.h"
 #include "vec/utils/template_helpers.hpp"
-#include "vec/utils/util.hpp"
+
+namespace doris {
+class DecimalV2Value;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/like.cpp b/be/src/vec/functions/like.cpp
index d8fc1b0e6c..d75bea36c5 100644
--- a/be/src/vec/functions/like.cpp
+++ b/be/src/vec/functions/like.cpp
@@ -17,8 +17,24 @@
 
 #include "vec/functions/like.h"
 
+#include <fmt/format.h>
+#include <hs/hs_compile.h>
+#include <re2/stringpiece.h>
+
+#include <algorithm>
+#include <cstddef>
+#include <ostream>
+#include <utility>
+#include <vector>
+
+#include "common/logging.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_const.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
 #include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/functions/simple_function_factory.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/functions/like.h b/be/src/vec/functions/like.h
index fab8527e29..b248a4750c 100644
--- a/be/src/vec/functions/like.h
+++ b/be/src/vec/functions/like.h
@@ -17,22 +17,37 @@
 
 #pragma once
 
-#include <hs/hs.h>
+#include <hs/hs_common.h>
+#include <hs/hs_runtime.h>
+#include <re2/re2.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
 #include <functional>
 #include <memory>
+#include <string>
 
+#include "common/status.h"
+#include "runtime/define_primitive_type.h"
 #include "runtime/string_search.hpp"
-#include "vec/columns/column_const.h"
-#include "vec/columns/column_set.h"
+#include "udf/udf.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column_string.h"
 #include "vec/columns/columns_number.h"
 #include "vec/columns/predicate_column.h"
 #include "vec/common/string_ref.h"
-#include "vec/data_types/data_type_nullable.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_number.h"
-#include "vec/exprs/vexpr.h"
 #include "vec/functions/function.h"
-#include "vec/functions/simple_function_factory.h"
+
+namespace doris {
+namespace vectorized {
+class Block;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/functions/match.cpp b/be/src/vec/functions/match.cpp
index 2b90f44f6b..1e69699d1b 100644
--- a/be/src/vec/functions/match.cpp
+++ b/be/src/vec/functions/match.cpp
@@ -15,16 +15,33 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <limits>
-#include <type_traits>
+#include <stddef.h>
 
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <utility>
+
+#include "common/config.h"
 #include "common/consts.h"
 #include "common/logging.h"
-#include "vec/columns/column_string.h"
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_number.h"
-#include "vec/data_types/data_type_string.h"
+#include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 
 class FunctionMatchBase : public IFunction {
diff --git a/be/src/vec/functions/math.cpp b/be/src/vec/functions/math.cpp
index 58418fe084..34a126971a 100644
--- a/be/src/vec/functions/math.cpp
+++ b/be/src/vec/functions/math.cpp
@@ -15,7 +15,30 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <stdint.h>
+#include <string.h>
+
+#include <boost/iterator/iterator_facade.hpp>
+// IWYU pragma: no_include <bits/std_abs.h>
+#include <algorithm>
+#include <cmath>
+#include <memory>
+#include <string>
+#include <type_traits>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
 #include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_decimal.h"
+#include "vec/data_types/data_type_nullable.h"
+#include "vec/data_types/data_type_number.h"
+#include "vec/data_types/data_type_string.h"
 #include "vec/data_types/number_traits.h"
 #include "vec/functions/function_binary_arithmetic.h"
 #include "vec/functions/function_const.h"
@@ -27,6 +50,14 @@
 #include "vec/functions/round.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+namespace vectorized {
+struct LnImpl;
+struct Log10Impl;
+struct Log2Impl;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 struct AcosName {
     static constexpr auto name = "acos";
diff --git a/be/src/vec/functions/minus.cpp b/be/src/vec/functions/minus.cpp
index 7c72d0145c..122842b2d9 100644
--- a/be/src/vec/functions/minus.cpp
+++ b/be/src/vec/functions/minus.cpp
@@ -18,7 +18,11 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Functions/Minus.cpp
 // and modified by Doris
 
+#include <utility>
+
+#include "runtime/decimalv2_value.h"
 #include "vec/common/arithmetic_overflow.h"
+#include "vec/data_types/number_traits.h"
 #include "vec/functions/function_binary_arithmetic.h"
 #include "vec/functions/simple_function_factory.h"
 
diff --git a/be/src/vec/functions/modulo.cpp b/be/src/vec/functions/modulo.cpp
index 599790d0d4..66376b6601 100644
--- a/be/src/vec/functions/modulo.cpp
+++ b/be/src/vec/functions/modulo.cpp
@@ -18,11 +18,16 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Functions/Modulo.cpp
 // and modified by Doris
 
-#include <libdivide.h>
+#include <string.h>
+
+#include <cmath>
+#include <memory>
+#include <utility>
 
-#include "common/status.h"
 #include "runtime/decimalv2_value.h"
+#include "vec/columns/column_vector.h"
 #include "vec/core/types.h"
+#include "vec/data_types/number_traits.h"
 #include "vec/functions/function_binary_arithmetic.h"
 #include "vec/functions/simple_function_factory.h"
 
diff --git a/be/src/vec/functions/multiply.cpp b/be/src/vec/functions/multiply.cpp
index 0a350392e7..43bcdf1587 100644
--- a/be/src/vec/functions/multiply.cpp
+++ b/be/src/vec/functions/multiply.cpp
@@ -18,9 +18,16 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Functions/Multiply.cpp
 // and modified by Doris
 
+#include <stddef.h>
+
+#include <utility>
+
+#include "gutil/integral_types.h"
 #include "runtime/decimalv2_value.h"
-#include "vec/columns/column_decimal.h"
+#include "vec/columns/columns_number.h"
 #include "vec/common/arithmetic_overflow.h"
+#include "vec/core/types.h"
+#include "vec/data_types/number_traits.h"
 #include "vec/functions/function_binary_arithmetic.h"
 #include "vec/functions/simple_function_factory.h"
 
diff --git a/be/src/vec/functions/nullif.cpp b/be/src/vec/functions/nullif.cpp
index d2e1aaf25a..65f9922cad 100644
--- a/be/src/vec/functions/nullif.cpp
+++ b/be/src/vec/functions/nullif.cpp
@@ -18,16 +18,34 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Functions/Nullif.cpp
 // and modified by Doris
 
-#include "vec/columns/column_nullable.h"
+#include <glog/logging.h>
+#include <stddef.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <vector>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_const.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/columns_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 #include "vec/data_types/data_type_nothing.h"
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
-#include "vec/data_types/get_least_supertype.h"
-#include "vec/data_types/number_traits.h"
-#include "vec/functions/function_helpers.h"
-#include "vec/functions/function_string.h"
+#include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
-#include "vec/utils/util.hpp"
+
+namespace doris {
+class FunctionContext;
+} // namespace doris
 
 namespace doris::vectorized {
 class FunctionNullIf : public IFunction {
diff --git a/be/src/vec/functions/plus.cpp b/be/src/vec/functions/plus.cpp
index 9e3b8923c8..480db82a58 100644
--- a/be/src/vec/functions/plus.cpp
+++ b/be/src/vec/functions/plus.cpp
@@ -18,7 +18,11 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Functions/Plus.cpp
 // and modified by Doris
 
+#include <utility>
+
+#include "runtime/decimalv2_value.h"
 #include "vec/common/arithmetic_overflow.h"
+#include "vec/data_types/number_traits.h"
 #include "vec/functions/function_binary_arithmetic.h"
 #include "vec/functions/simple_function_factory.h"
 
diff --git a/be/src/vec/functions/random.cpp b/be/src/vec/functions/random.cpp
index 64f8920fb6..345a99c8bf 100644
--- a/be/src/vec/functions/random.cpp
+++ b/be/src/vec/functions/random.cpp
@@ -15,10 +15,28 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <random>
+#include <glog/logging.h>
 
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <cstdint>
+#include <cstdlib>
+#include <memory>
+#include <random>
+#include <utility>
+
+#include "common/status.h"
 #include "udf/udf.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_number.h"
+#include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/functions/round.h b/be/src/vec/functions/round.h
index cada723ace..8645963ec6 100644
--- a/be/src/vec/functions/round.h
+++ b/be/src/vec/functions/round.h
@@ -30,6 +30,7 @@
 
 #include "vec/columns/column.h"
 #include "vec/columns/column_decimal.h"
+#include "vec/core/call_on_type_index.h"
 #include "vec/data_types/data_type_decimal.h"
 #include "vec/data_types/data_type_number.h"
 
@@ -563,4 +564,4 @@ public:
     }
 };
 
-} // namespace doris::vectorized
\ No newline at end of file
+} // namespace doris::vectorized
diff --git a/be/src/vec/functions/time_of_function.cpp b/be/src/vec/functions/time_of_function.cpp
index 415e8030c6..8e2f5d73a9 100644
--- a/be/src/vec/functions/time_of_function.cpp
+++ b/be/src/vec/functions/time_of_function.cpp
@@ -15,7 +15,15 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <utility>
+
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
+#include "vec/data_types/data_type_time_v2.h" // IWYU pragma: keep
 #include "vec/functions/date_time_transforms.h"
 #include "vec/functions/function_date_or_datetime_to_something.h"
 #include "vec/functions/simple_function_factory.h"
@@ -117,4 +125,4 @@ void register_function_time_of_function(SimpleFunctionFactory& factory) {
     factory.register_alternative_function<FunctionWeekDayV2Old>();
     factory.register_alternative_function<FunctionDateTimeV2WeekDayOld>();
 }
-} // namespace doris::vectorized
\ No newline at end of file
+} // namespace doris::vectorized
diff --git a/be/src/vec/functions/to_time_function.cpp b/be/src/vec/functions/to_time_function.cpp
index 01130e123f..b4b4a12f13 100644
--- a/be/src/vec/functions/to_time_function.cpp
+++ b/be/src/vec/functions/to_time_function.cpp
@@ -16,7 +16,17 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <utility>
+
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/data_types/data_type_date_time.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
+#include "vec/data_types/data_type_time_v2.h"
 #include "vec/functions/date_time_transforms.h"
 #include "vec/functions/function_date_or_datetime_to_something.h"
 #include "vec/functions/simple_function_factory.h"
diff --git a/be/src/vec/functions/url/domain.h b/be/src/vec/functions/url/domain.h
index 98c2e053b5..54361134ef 100644
--- a/be/src/vec/functions/url/domain.h
+++ b/be/src/vec/functions/url/domain.h
@@ -21,10 +21,9 @@
 #pragma once
 
 // #include <base/find_symbols.h>
-#include <vec/common/string_utils/string_utils.h>
-
 #include <cstring>
 
+#include "vec/common/string_utils/string_utils.h"
 #include "vec/functions/url/protocol.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/functions/url/function_url.cpp b/be/src/vec/functions/url/function_url.cpp
index 501343c1f6..e25af6f7f2 100644
--- a/be/src/vec/functions/url/function_url.cpp
+++ b/be/src/vec/functions/url/function_url.cpp
@@ -18,6 +18,8 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/Functions/URL/domain.cpp
 // and modified by Doris
 
+#include "vec/columns/column.h"
+#include "vec/data_types/data_type.h"
 #include "vec/functions/function_string_to_string.h"
 #include "vec/functions/simple_function_factory.h"
 #include "vec/functions/url/domain.h"
diff --git a/be/src/vec/functions/uuid.cpp b/be/src/vec/functions/uuid.cpp
index acd1233cc1..2ada40d903 100644
--- a/be/src/vec/functions/uuid.cpp
+++ b/be/src/vec/functions/uuid.cpp
@@ -15,14 +15,29 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <boost/uuid/uuid.hpp>
-#include <boost/uuid/uuid_generators.hpp>
-#include <boost/uuid/uuid_io.hpp>
+#include <glog/logging.h>
+#include <stddef.h>
 
-#include "udf/udf.h"
+#include <boost/uuid/random_generator.hpp>
+#include <boost/uuid/uuid_io.hpp>
+#include <memory>
+#include <string>
+#include <utility>
+
+#include "common/status.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column_string.h"
+#include "vec/core/block.h"
+#include "vec/core/column_numbers.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_string.h"
+#include "vec/functions/function.h"
 #include "vec/functions/simple_function_factory.h"
 
+namespace doris {
+class FunctionContext;
+} // namespace doris
+
 namespace doris::vectorized {
 class Uuid : public IFunction {
 public:
diff --git a/be/src/vec/io/io_helper.h b/be/src/vec/io/io_helper.h
index 003124630f..7111913910 100644
--- a/be/src/vec/io/io_helper.h
+++ b/be/src/vec/io/io_helper.h
@@ -17,12 +17,12 @@
 
 #pragma once
 
+#include <gen_cpp/data.pb.h>
 #include <snappy/snappy.h>
 
 #include <iostream>
 
 #include "common/exception.h"
-#include "gen_cpp/data.pb.h"
 #include "util/binary_cast.hpp"
 #include "util/string_parser.hpp"
 #include "vec/common/arena.h"
diff --git a/be/src/vec/json/json_parser.cpp b/be/src/vec/json/json_parser.cpp
index 893634c19e..95dd0034c3 100644
--- a/be/src/vec/json/json_parser.cpp
+++ b/be/src/vec/json/json_parser.cpp
@@ -20,6 +20,14 @@
 
 #include "vec/json/json_parser.h"
 
+#include <assert.h>
+#include <fmt/format.h>
+#include <glog/logging.h>
+
+#include <algorithm>
+#include <string_view>
+
+#include "vec/json/path_in_data.h"
 #include "vec/json/simd_json_parser.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/json/json_parser.h b/be/src/vec/json/json_parser.h
index db728c2442..24d61b89b7 100644
--- a/be/src/vec/json/json_parser.h
+++ b/be/src/vec/json/json_parser.h
@@ -21,9 +21,19 @@
 #pragma once
 
 #include <parallel_hashmap/phmap.h>
-#include <vec/common/hash_table/hash_map.h>
-#include <vec/core/types.h>
-#include <vec/json/path_in_data.h>
+#include <stddef.h>
+
+#include <optional>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "vec/columns/column.h"
+#include "vec/common/string_ref.h"
+#include "vec/common/uint128.h"
+#include "vec/core/field.h"
+#include "vec/json/path_in_data.h"
+#include "vec/json/simd_json_parser.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/json/parse2column.cpp b/be/src/vec/json/parse2column.cpp
index 3970bb0156..7b12d3e7cd 100644
--- a/be/src/vec/json/parse2column.cpp
+++ b/be/src/vec/json/parse2column.cpp
@@ -15,17 +15,39 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <parallel_hashmap/phmap.h>
-#include <vec/columns/column_object.h>
-#include <vec/common/field_visitors.h>
-#include <vec/common/hash_table/hash_set.h>
-#include <vec/common/schema_util.h>
-#include <vec/common/string_ref.h>
-#include <vec/json/json_parser.h>
-#include <vec/json/parse2column.h>
-#include <vec/json/simd_json_parser.h>
+#include "vec/json/parse2column.h"
 
+#include <assert.h>
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <parallel_hashmap/phmap.h>
+#include <simdjson/simdjson.h> // IWYU pragma: keep
+#include <stddef.h>
+
+#include <algorithm>
+#include <memory>
+#include <mutex>
+#include <optional>
+#include <ostream>
 #include <stack>
+#include <string>
+#include <string_view>
+#include <utility>
+
+#include "common/config.h"
+#include "common/status.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_object.h"
+#include "vec/columns/column_string.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/field_visitors.h"
+#include "vec/common/schema_util.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/field.h"
+#include "vec/data_types/data_type.h"
+#include "vec/json/json_parser.h"
+#include "vec/json/path_in_data.h"
+#include "vec/json/simd_json_parser.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/json/parse2column.h b/be/src/vec/json/parse2column.h
index adf1c63e62..0bc13dd591 100644
--- a/be/src/vec/json/parse2column.h
+++ b/be/src/vec/json/parse2column.h
@@ -17,9 +17,21 @@
 #pragma once
 
 #include <common/status.h>
-#include <vec/columns/column.h>
-#include <vec/json/json_parser.h>
-#include <vec/json/simd_json_parser.h>
+
+#include <vector>
+
+#include "vec/columns/column.h"
+#include "vec/common/string_ref.h"
+
+namespace doris {
+namespace vectorized {
+class ColumnString;
+class SimdJSONParser;
+enum class ExtractType;
+template <typename ParserImpl>
+class JSONDataParser;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/json/path_in_data.cpp b/be/src/vec/json/path_in_data.cpp
index 0ee4ce8528..a05542e189 100644
--- a/be/src/vec/json/path_in_data.cpp
+++ b/be/src/vec/json/path_in_data.cpp
@@ -18,11 +18,12 @@
 // https://github.com/ClickHouse/ClickHouse/blob/master/src/DataTypes/Serializations/PathInData.cpp
 // and modified by Doris
 
-#include <vec/common/sip_hash.h>
-#include <vec/json/path_in_data.h>
+#include "vec/json/path_in_data.h"
+
+#include <assert.h>
+
+#include "vec/common/sip_hash.h"
 
-#include <boost/algorithm/string.hpp>
-#include <boost/algorithm/string/split.hpp>
 namespace doris::vectorized {
 PathInData::PathInData(std::string_view path_) : path(path_) {
     const char* begin = path.data();
diff --git a/be/src/vec/json/path_in_data.h b/be/src/vec/json/path_in_data.h
index 5118409e53..47da9c85cb 100644
--- a/be/src/vec/json/path_in_data.h
+++ b/be/src/vec/json/path_in_data.h
@@ -20,14 +20,19 @@
 
 #pragma once
 
-#include <bitset>
+#include <stddef.h>
+
+#include <algorithm>
+#include <string>
+#include <string_view>
 #include <vector>
 
+#include "vec/common/uint128.h"
 #include "vec/core/field.h"
 #include "vec/core/types.h"
+
 namespace doris::vectorized {
-class ReadBuffer;
-class WriteBuffer;
+
 /// Class that represents path in document, e.g. JSON.
 class PathInData {
 public:
diff --git a/be/src/vec/json/simd_json_parser.h b/be/src/vec/json/simd_json_parser.h
index 870ae6b947..9216f545a3 100644
--- a/be/src/vec/json/simd_json_parser.h
+++ b/be/src/vec/json/simd_json_parser.h
@@ -22,8 +22,9 @@
 
 #include <rapidjson/document.h>
 #include <simdjson.h>
-#include <vec/core/types.h>
-#include <vec/json/path_in_data.h>
+
+#include "vec/core/types.h"
+#include "vec/json/path_in_data.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/jsonb/serialize.cpp b/be/src/vec/jsonb/serialize.cpp
index 1d06b0873f..e1bf5f17b1 100644
--- a/be/src/vec/jsonb/serialize.cpp
+++ b/be/src/vec/jsonb/serialize.cpp
@@ -17,13 +17,35 @@
 
 #include "vec/jsonb/serialize.h"
 
+#include <assert.h>
+#include <glog/logging.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <ostream>
+#include <vector>
+
 #include "olap/hll.h"
+#include "olap/olap_common.h"
 #include "olap/tablet_schema.h"
-#include "runtime/jsonb_value.h"
+#include "runtime/define_primitive_type.h"
+#include "runtime/descriptors.h"
+#include "runtime/primitive_type.h"
+#include "runtime/types.h"
+#include "util/bitmap_value.h"
+#include "util/jsonb_document.h"
 #include "util/jsonb_stream.h"
 #include "util/jsonb_writer.h"
+#include "util/slice.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_string.h"
 #include "vec/common/arena.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/field.h"
 #include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
+#include "vec/runtime/vdatetime_value.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/jsonb/serialize.h b/be/src/vec/jsonb/serialize.h
index f3861325d1..0d9d13927d 100644
--- a/be/src/vec/jsonb/serialize.h
+++ b/be/src/vec/jsonb/serialize.h
@@ -16,9 +16,19 @@
 // under the License.
 
 #pragma once
-#include "vec/columns/column_string.h"
+#include <stddef.h>
+
 #include "vec/core/block.h"
 
+namespace doris {
+class TabletSchema;
+class TupleDescriptor;
+
+namespace vectorized {
+class ColumnString;
+} // namespace vectorized
+} // namespace doris
+
 namespace doris::vectorized {
 // use jsonb codec to store row format
 class JsonbSerializeUtil {
diff --git a/be/src/vec/olap/block_reader.cpp b/be/src/vec/olap/block_reader.cpp
index a9dbcd6226..5e9b270099 100644
--- a/be/src/vec/olap/block_reader.cpp
+++ b/be/src/vec/olap/block_reader.cpp
@@ -17,12 +17,39 @@
 
 #include "vec/olap/block_reader.h"
 
+#include <gen_cpp/olap_file.pb.h>
+#include <glog/logging.h>
+#include <stdint.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <memory>
+#include <ostream>
+#include <string>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/status.h"
+#include "exprs/function_filter.h"
 #include "olap/like_column_predicate.h"
 #include "olap/olap_common.h"
+#include "olap/olap_define.h"
+#include "olap/rowset/rowset.h"
+#include "olap/rowset/rowset_reader_context.h"
+#include "olap/tablet.h"
+#include "olap/tablet_schema.h"
 #include "vec/aggregate_functions/aggregate_function_reader.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/data_types/data_type_number.h"
 #include "vec/olap/vcollect_iterator.h"
 
+namespace doris {
+class ColumnPredicate;
+} // namespace doris
+
 namespace doris::vectorized {
 using namespace ErrorCode;
 
diff --git a/be/src/vec/olap/block_reader.h b/be/src/vec/olap/block_reader.h
index 4669622c65..db6dacfdc7 100644
--- a/be/src/vec/olap/block_reader.h
+++ b/be/src/vec/olap/block_reader.h
@@ -18,13 +18,25 @@
 #pragma once
 
 #include <parallel_hashmap/phmap.h>
+#include <stddef.h>
 
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
 #include "olap/reader.h"
 #include "olap/rowset/rowset_reader.h"
+#include "olap/utils.h"
 #include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
+#include "vec/core/block.h"
+#include "vec/data_types/data_type.h"
 #include "vec/olap/vcollect_iterator.h"
 
 namespace doris {
+class ColumnPredicate;
+class FunctionFilter;
+class RuntimeProfile;
 
 namespace vectorized {
 
diff --git a/be/src/vec/olap/olap_data_convertor.cpp b/be/src/vec/olap/olap_data_convertor.cpp
index f0d1cff777..a9eaae7e75 100644
--- a/be/src/vec/olap/olap_data_convertor.cpp
+++ b/be/src/vec/olap/olap_data_convertor.cpp
@@ -17,13 +17,31 @@
 
 #include "vec/olap/olap_data_convertor.h"
 
+#include <new>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/config.h"
+#include "olap/hll.h"
+#include "olap/olap_common.h"
 #include "olap/tablet_schema.h"
+#include "runtime/decimalv2_value.h"
+#include "util/bitmap_value.h"
+#include "util/quantile_state.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
 #include "vec/columns/column_complex.h"
+#include "vec/columns/column_decimal.h"
+#include "vec/columns/column_map.h"
 #include "vec/columns/column_struct.h"
 #include "vec/columns/column_vector.h"
+#include "vec/core/block.h"
 #include "vec/data_types/data_type_array.h"
+#include "vec/data_types/data_type_map.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_struct.h"
+#include "vec/runtime/vdatetime_value.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/olap/olap_data_convertor.h b/be/src/vec/olap/olap_data_convertor.h
index efe6f3e709..67847542b5 100644
--- a/be/src/vec/olap/olap_data_convertor.h
+++ b/be/src/vec/olap/olap_data_convertor.h
@@ -17,25 +17,47 @@
 
 #pragma once
 
-#include "olap/types.h"
-#include "vec/columns/column_map.h"
+#include <assert.h>
+#include <glog/logging.h>
+#include <stdint.h>
+#include <string.h>
+
+#include <algorithm>
+#include <memory>
+#include <ostream>
+#include <utility>
+#include <vector>
+
+#include "common/status.h"
+#include "gutil/integral_types.h"
+#include "olap/decimal12.h"
+#include "olap/uint24.h"
+#include "runtime/collection_value.h"
+#include "util/slice.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/pod_array_fwd.h"
+#include "vec/common/string_ref.h"
 #include "vec/core/column_with_type_and_name.h"
 #include "vec/core/types.h"
-#include "vec/data_types/data_type_map.h"
+#include "vec/data_types/data_type.h"
 
 namespace doris {
 
 class TabletSchema;
 class TabletColumn;
-class MemTracker;
-class Status;
 
 namespace vectorized {
 
 class Block;
 class ColumnArray;
 class DataTypeArray;
+class ColumnMap;
+class DataTypeMap;
+template <typename T>
+class ColumnDecimal;
 
 class IOlapColumnDataAccessor {
 public:
@@ -61,6 +83,7 @@ public:
 
 private:
     class OlapColumnDataConvertorBase;
+
     using OlapColumnDataConvertorBaseUPtr = std::unique_ptr<OlapColumnDataConvertorBase>;
     using OlapColumnDataConvertorBaseSPtr = std::shared_ptr<OlapColumnDataConvertorBase>;
 
diff --git a/be/src/vec/olap/vcollect_iterator.cpp b/be/src/vec/olap/vcollect_iterator.cpp
index f21dd25ab8..82bd608810 100644
--- a/be/src/vec/olap/vcollect_iterator.cpp
+++ b/be/src/vec/olap/vcollect_iterator.cpp
@@ -17,8 +17,33 @@
 
 #include "vec/olap/vcollect_iterator.h"
 
+#include <gen_cpp/olap_file.pb.h>
+#include <glog/logging.h>
+
+#include <algorithm>
+#include <iterator>
+#include <ostream>
+#include <set>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/status.h"
-#include "util/defer_op.h"
+#include "io/io_common.h"
+#include "olap/olap_common.h"
+#include "olap/olap_define.h"
+#include "olap/rowset/rowset.h"
+#include "olap/rowset/rowset_meta.h"
+#include "olap/tablet.h"
+#include "olap/tablet_schema.h"
+#include "runtime/query_fragments_ctx.h"
+#include "runtime/runtime_predicate.h"
+#include "runtime/runtime_state.h"
+#include "vec/columns/column.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/field.h"
+#include "vec/data_types/data_type.h"
+#include "vec/exprs/vexpr_context.h"
 
 namespace doris {
 using namespace ErrorCode;
diff --git a/be/src/vec/olap/vcollect_iterator.h b/be/src/vec/olap/vcollect_iterator.h
index cf1af2f73a..863173c932 100644
--- a/be/src/vec/olap/vcollect_iterator.h
+++ b/be/src/vec/olap/vcollect_iterator.h
@@ -17,7 +17,16 @@
 
 #pragma once
 
+#include <stddef.h>
+#include <stdint.h>
+
+#include <list>
+#include <memory>
+#include <vector>
+
 #include "common/status.h"
+#include "olap/rowset/rowset_reader_context.h"
+#include "olap/utils.h"
 #ifdef USE_LIBCPP
 #include <queue>
 #else
@@ -28,9 +37,14 @@
 #include "olap/rowset/rowset_reader.h"
 #include "vec/core/block.h"
 
+namespace __gnu_pbds {
+struct pairing_heap_tag;
+} // namespace __gnu_pbds
+
 namespace doris {
 
 class TabletSchema;
+class RuntimeProfile;
 
 namespace vectorized {
 
diff --git a/be/src/vec/olap/vertical_block_reader.cpp b/be/src/vec/olap/vertical_block_reader.cpp
index 097ddb7513..e153d2020e 100644
--- a/be/src/vec/olap/vertical_block_reader.cpp
+++ b/be/src/vec/olap/vertical_block_reader.cpp
@@ -17,12 +17,26 @@
 
 #include "vec/olap/vertical_block_reader.h"
 
+#include <assert.h>
+#include <gen_cpp/olap_file.pb.h>
+
+#include <algorithm>
+#include <boost/iterator/iterator_facade.hpp>
+#include <ostream>
+
 #include "common/status.h"
-#include "olap/like_column_predicate.h"
 #include "olap/olap_common.h"
+#include "olap/olap_define.h"
+#include "olap/rowset/rowset.h"
+#include "olap/rowset/rowset_reader.h"
+#include "olap/rowset/rowset_reader_context.h"
+#include "olap/tablet_schema.h"
 #include "vec/aggregate_functions/aggregate_function_reader.h"
-#include "vec/olap/block_reader.h"
-#include "vec/olap/vcollect_iterator.h"
+#include "vec/columns/column_nullable.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/data_types/data_type_number.h"
 #include "vec/olap/vertical_merge_iterator.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/olap/vertical_block_reader.h b/be/src/vec/olap/vertical_block_reader.h
index fcd93330a1..2bc912193b 100644
--- a/be/src/vec/olap/vertical_block_reader.h
+++ b/be/src/vec/olap/vertical_block_reader.h
@@ -17,19 +17,34 @@
 
 #pragma once
 
+#include <glog/logging.h>
 #include <parallel_hashmap/phmap.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <memory>
+#include <utility>
+#include <vector>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/status.h"
+#include "olap/iterators.h"
 #include "olap/reader.h"
-#include "olap/rowset/rowset_reader.h"
+#include "olap/tablet.h"
+#include "olap/utils.h"
 #include "vec/aggregate_functions/aggregate_function.h"
-#include "vec/olap/vcollect_iterator.h"
-#include "vec/olap/vertical_merge_iterator.h"
+#include "vec/columns/column.h"
+#include "vec/core/block.h"
+#include "vec/data_types/data_type.h"
 
 #pragma once
 
 namespace doris {
+struct RowsetId;
 
 namespace vectorized {
+class RowSourcesBuffer;
 
 class VerticalBlockReader final : public TabletReader {
 public:
diff --git a/be/src/vec/olap/vertical_merge_iterator.cpp b/be/src/vec/olap/vertical_merge_iterator.cpp
index 1f2e682f3d..ea4df4b1aa 100644
--- a/be/src/vec/olap/vertical_merge_iterator.cpp
+++ b/be/src/vec/olap/vertical_merge_iterator.cpp
@@ -17,7 +17,23 @@
 
 #include "vec/olap/vertical_merge_iterator.h"
 
+#include <fcntl.h>
+#include <gen_cpp/olap_file.pb.h>
+#include <stdlib.h>
+
+#include <ostream>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/config.h"
+#include "common/logging.h"
+#include "olap/field.h"
 #include "olap/olap_common.h"
+#include "vec/columns/column.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
+#include "vec/data_types/data_type.h"
 
 namespace doris {
 using namespace ErrorCode;
diff --git a/be/src/vec/olap/vertical_merge_iterator.h b/be/src/vec/olap/vertical_merge_iterator.h
index 4fb3332892..120548ede4 100644
--- a/be/src/vec/olap/vertical_merge_iterator.h
+++ b/be/src/vec/olap/vertical_merge_iterator.h
@@ -15,16 +15,32 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <glog/logging.h>
+#include <stdint.h>
+#include <unistd.h>
+
+#include <algorithm>
+#include <list>
+#include <memory>
+#include <queue>
+#include <string>
+#include <utility>
+#include <vector>
+
 #include "common/status.h"
 #include "io/io_common.h"
 #include "olap/iterators.h"
+#include "olap/olap_common.h"
 #include "olap/schema.h"
+#include "olap/utils.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
 #include "vec/core/block.h"
 
 #pragma once
 
 namespace doris {
+enum KeysType : int;
 
 namespace vectorized {
 
diff --git a/be/src/vec/olap/vgeneric_iterators.cpp b/be/src/vec/olap/vgeneric_iterators.cpp
index b26cde0841..5d6caa434c 100644
--- a/be/src/vec/olap/vgeneric_iterators.cpp
+++ b/be/src/vec/olap/vgeneric_iterators.cpp
@@ -15,19 +15,28 @@
 // specific language governing permissions and limitations
 // under the License.
 
-#include <vec/olap/vgeneric_iterators.h>
+#include "vec/olap/vgeneric_iterators.h"
 
+#include <algorithm>
 #include <memory>
-#include <queue>
 #include <utility>
 
 #include "common/status.h"
+#include "olap/field.h"
 #include "olap/iterators.h"
+#include "olap/olap_common.h"
+#include "olap/rowset/segment_v2/column_reader.h"
 #include "olap/rowset/segment_v2/segment.h"
 #include "olap/schema.h"
+#include "olap/tablet_schema.h"
+#include "vec/columns/column.h"
 #include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/data_types/data_type.h"
 
 namespace doris {
+class RuntimeProfile;
+
 using namespace ErrorCode;
 
 namespace vectorized {
diff --git a/be/src/vec/olap/vgeneric_iterators.h b/be/src/vec/olap/vgeneric_iterators.h
index a59f65678c..b80598aa74 100644
--- a/be/src/vec/olap/vgeneric_iterators.h
+++ b/be/src/vec/olap/vgeneric_iterators.h
@@ -15,14 +15,33 @@
 // specific language governing permissions and limitations
 // under the License.
 
+#include <gen_cpp/PlanNodes_types.h>
+#include <glog/logging.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <list>
+#include <map>
+#include <memory>
+#include <queue>
+#include <utility>
+#include <vector>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "common/status.h"
 #include "olap/iterators.h"
-#include "olap/rowset/segment_v2/column_reader.h"
+#include "olap/schema.h"
+#include "olap/utils.h"
+#include "vec/core/block.h"
 
 namespace doris {
+class RuntimeProfile;
 
 namespace segment_v2 {
 class Segment;
-}
+class ColumnIterator;
+} // namespace segment_v2
 
 namespace vectorized {
 
diff --git a/be/src/vec/runtime/shared_hash_table_controller.cpp b/be/src/vec/runtime/shared_hash_table_controller.cpp
index 6241583200..9353bdbbec 100644
--- a/be/src/vec/runtime/shared_hash_table_controller.cpp
+++ b/be/src/vec/runtime/shared_hash_table_controller.cpp
@@ -17,7 +17,11 @@
 
 #include "shared_hash_table_controller.h"
 
+#include <glog/logging.h>
 #include <runtime/runtime_state.h>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <utility>
 
 namespace doris {
 namespace vectorized {
diff --git a/be/src/vec/runtime/shared_hash_table_controller.h b/be/src/vec/runtime/shared_hash_table_controller.h
index 4c579f1d91..0f2517c46c 100644
--- a/be/src/vec/runtime/shared_hash_table_controller.h
+++ b/be/src/vec/runtime/shared_hash_table_controller.h
@@ -17,18 +17,19 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+
 #include <condition_variable>
 #include <map>
 #include <memory>
 #include <mutex>
 #include <vector>
 
-#include "vec/core/block.h"
+#include "common/status.h"
 
 namespace doris {
 
 class RuntimeState;
-class TUniqueId;
 class MinMaxFuncBase;
 class HybridSetBase;
 class BloomFilterFuncBase;
@@ -36,7 +37,8 @@ class BitmapFilterFuncBase;
 
 namespace vectorized {
 
-class VExprContext;
+class Arena;
+class Block;
 
 struct SharedRuntimeFilterContext {
     std::shared_ptr<MinMaxFuncBase> minmax_func;
diff --git a/be/src/vec/runtime/vdata_stream_mgr.cpp b/be/src/vec/runtime/vdata_stream_mgr.cpp
index 09c9cd77c7..e0e2229daf 100644
--- a/be/src/vec/runtime/vdata_stream_mgr.cpp
+++ b/be/src/vec/runtime/vdata_stream_mgr.cpp
@@ -17,12 +17,17 @@
 
 #include "vec/runtime/vdata_stream_mgr.h"
 
-#include "gen_cpp/internal_service.pb.h"
-#include "runtime/descriptors.h"
-#include "runtime/primitive_type.h"
-#include "runtime/runtime_state.h"
-#include "util/doris_metrics.h"
-#include "util/runtime_profile.h"
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/internal_service.pb.h>
+#include <gen_cpp/types.pb.h>
+#include <stddef.h>
+
+#include <ostream>
+#include <string>
+#include <vector>
+
+#include "common/logging.h"
+#include "util/hash_util.hpp"
 #include "vec/runtime/vdata_stream_recvr.h"
 
 namespace doris {
diff --git a/be/src/vec/runtime/vdata_stream_mgr.h b/be/src/vec/runtime/vdata_stream_mgr.h
index bc7bbf02b4..ca0e7ab4b7 100644
--- a/be/src/vec/runtime/vdata_stream_mgr.h
+++ b/be/src/vec/runtime/vdata_stream_mgr.h
@@ -17,14 +17,17 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <stdint.h>
+
 #include <memory>
 #include <mutex>
 #include <set>
 #include <unordered_map>
+#include <utility>
 
 #include "common/global_types.h"
 #include "common/status.h"
-#include "gen_cpp/Types_types.h"
 
 namespace google {
 namespace protobuf {
@@ -35,7 +38,6 @@ class Closure;
 namespace doris {
 class RuntimeState;
 class RowDescriptor;
-class TUniqueId;
 class RuntimeProfile;
 class QueryStatisticsRecvr;
 class PTransmitDataParams;
diff --git a/be/src/vec/runtime/vdata_stream_recvr.cpp b/be/src/vec/runtime/vdata_stream_recvr.cpp
index 39d70ee4f1..7cc3f9cc71 100644
--- a/be/src/vec/runtime/vdata_stream_recvr.cpp
+++ b/be/src/vec/runtime/vdata_stream_recvr.cpp
@@ -17,8 +17,18 @@
 
 #include "vec/runtime/vdata_stream_recvr.h"
 
-#include "gen_cpp/data.pb.h"
+#include <fmt/format.h>
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/data.pb.h>
+
+#include <algorithm>
+#include <functional>
+#include <string>
+
+#include "common/logging.h"
 #include "runtime/memory/mem_tracker.h"
+#include "runtime/runtime_state.h"
 #include "runtime/thread_context.h"
 #include "util/uid_util.h"
 #include "vec/core/block.h"
diff --git a/be/src/vec/runtime/vdata_stream_recvr.h b/be/src/vec/runtime/vdata_stream_recvr.h
index dd179ca29e..17bceeed2a 100644
--- a/be/src/vec/runtime/vdata_stream_recvr.h
+++ b/be/src/vec/runtime/vdata_stream_recvr.h
@@ -17,34 +17,46 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <glog/logging.h>
+#include <google/protobuf/stubs/callback.h>
+#include <stddef.h>
+#include <stdint.h>
+
 #include <atomic>
 #include <condition_variable>
 #include <deque>
 #include <list>
+#include <memory>
+#include <mutex>
+#include <ostream>
 #include <thread>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
+#include <vector>
 
+#include "common/config.h"
 #include "common/global_types.h"
 #include "common/object_pool.h"
 #include "common/status.h"
-#include "gen_cpp/Types_types.h"
 #include "runtime/descriptors.h"
 #include "runtime/query_statistics.h"
 #include "util/runtime_profile.h"
+#include "util/stopwatch.hpp"
+#include "vec/columns/column.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/core/materialize_block.h"
 
-namespace google {
-namespace protobuf {
-class Closure;
-}
-} // namespace google
-
 namespace doris {
 class MemTracker;
-class RuntimeProfile;
 class PBlock;
+class MemTrackerLimiter;
+class PQueryStatistics;
+class RuntimeState;
 
 namespace vectorized {
-class Block;
 class VDataStreamMgr;
 class VSortedRunMerger;
 class VExprContext;
@@ -100,6 +112,7 @@ public:
 private:
     class SenderQueue;
     class PipSenderQueue;
+
     friend struct BlockSupplierSortCursorImpl;
 
     // DataStreamMgr instance used to create this recvr. (Not owned)
diff --git a/be/src/vec/runtime/vdatetime_value.cpp b/be/src/vec/runtime/vdatetime_value.cpp
index 148c67f95a..2a51cca8a0 100644
--- a/be/src/vec/runtime/vdatetime_value.cpp
+++ b/be/src/vec/runtime/vdatetime_value.cpp
@@ -17,16 +17,18 @@
 
 #include "vec/runtime/vdatetime_value.h"
 
+#include <cctz/civil_time.h>
+#include <cctz/time_zone.h>
 #include <ctype.h>
+#include <stdlib.h>
 #include <string.h>
 #include <time.h>
-
-#include <limits>
-#include <sstream>
-#include <valarray>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+// IWYU pragma: no_include <bits/std_abs.h>
+#include <cmath>
 
 #include "common/config.h"
-#include "runtime/datetime_value.h"
 #include "util/timezone_utils.h"
 
 namespace doris::vectorized {
diff --git a/be/src/vec/runtime/vdatetime_value.h b/be/src/vec/runtime/vdatetime_value.h
index 09c1d4dbe6..891d6fb8b1 100644
--- a/be/src/vec/runtime/vdatetime_value.h
+++ b/be/src/vec/runtime/vdatetime_value.h
@@ -17,19 +17,28 @@
 
 #pragma once
 
+#include <glog/logging.h>
 #include <re2/re2.h>
 #include <stdint.h>
+#include <string.h>
 
-#include <chrono>
-#include <climits>
+#include <algorithm>
 #include <cstddef>
 #include <iostream>
+#include <iterator>
+#include <string>
+#include <string_view>
+#include <tuple>
+#include <type_traits>
 
-#include "cctz/time_zone.h"
 #include "util/hash_util.hpp"
 #include "util/time_lut.h"
 #include "util/timezone_utils.h"
 
+namespace cctz {
+class time_zone;
+} // namespace cctz
+
 namespace doris {
 
 namespace vectorized {
diff --git a/be/src/vec/runtime/vfile_result_writer.cpp b/be/src/vec/runtime/vfile_result_writer.cpp
index 13eb1ab616..8820250eae 100644
--- a/be/src/vec/runtime/vfile_result_writer.cpp
+++ b/be/src/vec/runtime/vfile_result_writer.cpp
@@ -17,25 +17,50 @@
 
 #include "vec/runtime/vfile_result_writer.h"
 
+#include <gen_cpp/Data_types.h>
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <glog/logging.h>
+
+#include <ostream>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/consts.h"
 #include "common/status.h"
 #include "gutil/strings/numbers.h"
-#include "gutil/strings/substitute.h"
 #include "io/file_factory.h"
 #include "io/fs/file_writer.h"
 #include "io/fs/local_file_system.h"
 #include "runtime/buffer_control_block.h"
+#include "runtime/decimalv2_value.h"
+#include "runtime/define_primitive_type.h"
 #include "runtime/descriptors.h"
 #include "runtime/large_int_value.h"
+#include "runtime/primitive_type.h"
 #include "runtime/runtime_state.h"
+#include "runtime/types.h"
 #include "service/backend_options.h"
+#include "util/metrics.h"
 #include "util/mysql_global.h"
 #include "util/mysql_row_buffer.h"
+#include "util/types.h"
+#include "util/uid_util.h"
+#include "vec/columns/column.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
 #include "vec/common/string_ref.h"
 #include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/data_types/data_type.h"
 #include "vec/exprs/vexpr.h"
 #include "vec/exprs/vexpr_context.h"
+#include "vec/runtime/vdatetime_value.h"
 #include "vec/runtime/vorc_writer.h"
+#include "vec/runtime/vparquet_writer.h"
+#include "vec/sink/vresult_sink.h"
 
 namespace doris::vectorized {
 const size_t VFileResultWriter::OUTSTREAM_BUFFER_SIZE_BYTES = 1024 * 1024;
diff --git a/be/src/vec/runtime/vfile_result_writer.h b/be/src/vec/runtime/vfile_result_writer.h
index 6243244eb4..1f5cb05ba3 100644
--- a/be/src/vec/runtime/vfile_result_writer.h
+++ b/be/src/vec/runtime/vfile_result_writer.h
@@ -17,12 +17,34 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include <iosfwd>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "common/status.h"
 #include "io/fs/file_writer.h"
+#include "runtime/descriptors.h"
+#include "util/runtime_profile.h"
+#include "vec/core/block.h"
 #include "vec/runtime/vparquet_writer.h"
-#include "vec/sink/vresult_sink.h"
+#include "vec/sink/vresult_writer.h"
+
+namespace doris {
+class BufferControlBlock;
+class RuntimeState;
+
+namespace vectorized {
+class VExprContext;
+struct ResultFileOptions;
+} // namespace vectorized
+} // namespace doris
 
 namespace doris::vectorized {
-class VFileWriterWrapper;
 
 // write result to file
 class VFileResultWriter final : public VResultWriter {
diff --git a/be/src/vec/runtime/vorc_writer.cpp b/be/src/vec/runtime/vorc_writer.cpp
index d6809b026e..665ec27429 100644
--- a/be/src/vec/runtime/vorc_writer.cpp
+++ b/be/src/vec/runtime/vorc_writer.cpp
@@ -17,15 +17,36 @@
 
 #include "vec/runtime/vorc_writer.h"
 
+#include <glog/logging.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include <exception>
+#include <ostream>
+
 #include "io/fs/file_writer.h"
+#include "orc/Int128.hh"
+#include "orc/MemoryPool.hh"
+#include "orc/OrcFile.hh"
+#include "orc/Vector.hh"
+#include "runtime/define_primitive_type.h"
+#include "runtime/types.h"
+#include "util/binary_cast.hpp"
+#include "vec/columns/column.h"
 #include "vec/columns/column_complex.h"
+#include "vec/columns/column_decimal.h"
 #include "vec/columns/column_nullable.h"
 #include "vec/columns/column_string.h"
 #include "vec/columns/column_vector.h"
-#include "vec/data_types/data_type_decimal.h"
-#include "vec/data_types/data_type_nullable.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/pod_array.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
 #include "vec/exprs/vexpr.h"
-#include "vec/functions/function_helpers.h"
+#include "vec/exprs/vexpr_context.h"
+#include "vec/runtime/vdatetime_value.h"
 
 namespace doris::vectorized {
 VOrcOutputStream::VOrcOutputStream(doris::io::FileWriter* file_writer)
diff --git a/be/src/vec/runtime/vorc_writer.h b/be/src/vec/runtime/vorc_writer.h
index ef939ed5e8..cc66c65a6c 100644
--- a/be/src/vec/runtime/vorc_writer.h
+++ b/be/src/vec/runtime/vorc_writer.h
@@ -17,17 +17,31 @@
 
 #pragma once
 
+#include <stddef.h>
 #include <stdint.h>
 
-#include <map>
+#include <memory>
 #include <orc/OrcFile.hh>
 #include <string>
+#include <vector>
 
 #include "common/status.h"
-#include "io/fs/file_writer.h"
+#include "orc/Type.hh"
+#include "orc/Writer.hh"
 #include "vec/core/block.h"
-#include "vec/exprs/vexpr_context.h"
-#include "vec/runtime/vfile_result_writer.h"
+#include "vec/runtime/vparquet_writer.h"
+
+namespace doris {
+namespace io {
+class FileWriter;
+} // namespace io
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+namespace orc {
+struct ColumnVectorBatch;
+} // namespace orc
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/runtime/vparquet_writer.cpp b/be/src/vec/runtime/vparquet_writer.cpp
index 146d5f9cc7..ea25a1bc5c 100644
--- a/be/src/vec/runtime/vparquet_writer.cpp
+++ b/be/src/vec/runtime/vparquet_writer.cpp
@@ -17,21 +17,44 @@
 
 #include "vec/runtime/vparquet_writer.h"
 
-#include <arrow/array.h>
-#include <arrow/status.h>
+#include <arrow/io/type_fwd.h>
+#include <glog/logging.h>
+#include <math.h>
+#include <parquet/column_writer.h>
+#include <parquet/platform.h>
+#include <parquet/schema.h>
+#include <parquet/type_fwd.h>
 #include <time.h>
 
+#include <algorithm>
+#include <exception>
+#include <ostream>
+#include <string>
+
 #include "io/fs/file_writer.h"
+#include "runtime/decimalv2_value.h"
+#include "runtime/define_primitive_type.h"
+#include "runtime/types.h"
+#include "util/binary_cast.hpp"
 #include "util/mysql_global.h"
 #include "util/types.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_complex.h"
+#include "vec/columns/column_decimal.h"
 #include "vec/columns/column_nullable.h"
 #include "vec/columns/column_string.h"
 #include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_decimal.h"
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
 #include "vec/functions/function_helpers.h"
+#include "vec/runtime/vdatetime_value.h"
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/runtime/vparquet_writer.h b/be/src/vec/runtime/vparquet_writer.h
index 796fbbf983..419b4948ef 100644
--- a/be/src/vec/runtime/vparquet_writer.h
+++ b/be/src/vec/runtime/vparquet_writer.h
@@ -17,26 +17,34 @@
 
 #pragma once
 
-#include <arrow/api.h>
-#include <arrow/buffer.h>
-#include <arrow/io/api.h>
-#include <arrow/io/file.h>
 #include <arrow/io/interfaces.h>
+#include <arrow/result.h>
+#include <arrow/status.h>
 #include <gen_cpp/DataSinks_types.h>
-#include <parquet/api/reader.h>
-#include <parquet/api/writer.h>
-#include <parquet/arrow/reader.h>
-#include <parquet/arrow/writer.h>
-#include <parquet/exception.h>
+#include <parquet/file_writer.h>
+#include <parquet/properties.h>
+#include <parquet/types.h>
 #include <stdint.h>
 
-#include <map>
-#include <string>
+#include <memory>
+#include <vector>
 
 #include "common/status.h"
 #include "vec/core/block.h"
-#include "vec/exprs/vexpr_context.h"
-#include "vec/runtime/vfile_result_writer.h"
+
+namespace doris {
+namespace io {
+class FileWriter;
+} // namespace io
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
+namespace parquet {
+namespace schema {
+class GroupNode;
+} // namespace schema
+} // namespace parquet
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/runtime/vsorted_run_merger.cpp b/be/src/vec/runtime/vsorted_run_merger.cpp
index 17f87aac2e..0010327d01 100644
--- a/be/src/vec/runtime/vsorted_run_merger.cpp
+++ b/be/src/vec/runtime/vsorted_run_merger.cpp
@@ -17,12 +17,19 @@
 
 #include "vec/runtime/vsorted_run_merger.h"
 
+#include <utility>
 #include <vector>
 
-#include "runtime/descriptors.h"
-#include "util/debug_util.h"
-#include "util/defer_op.h"
 #include "util/runtime_profile.h"
+#include "util/stopwatch.hpp"
+#include "vec/columns/column.h"
+#include "vec/core/column_with_type_and_name.h"
+
+namespace doris {
+namespace vectorized {
+class VExprContext;
+} // namespace vectorized
+} // namespace doris
 
 using std::vector;
 
diff --git a/be/src/vec/runtime/vsorted_run_merger.h b/be/src/vec/runtime/vsorted_run_merger.h
index dd0d8b1d35..e48552eaba 100644
--- a/be/src/vec/runtime/vsorted_run_merger.h
+++ b/be/src/vec/runtime/vsorted_run_merger.h
@@ -17,16 +17,23 @@
 
 #pragma once
 
-#include <queue>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <queue>
+#include <vector>
+
+#include "common/status.h"
+#include "util/runtime_profile.h"
+#include "vec/core/block.h"
 #include "vec/core/sort_cursor.h"
+#include "vec/core/sort_description.h"
 
 namespace doris {
 
-class RuntimeProfile;
-
 namespace vectorized {
-class Block;
+class VExprContext;
+
 // VSortedRunMerger is used to merge multiple sorted runs of blocks. A run is a sorted
 // sequence of blocks, which are fetched from a BlockSupplier function object.
 // Merging is implemented using a binary min-heap that maintains the run with the next
diff --git a/be/src/vec/sink/vdata_stream_sender.cpp b/be/src/vec/sink/vdata_stream_sender.cpp
index 47204ba8c4..f11364e9ac 100644
--- a/be/src/vec/sink/vdata_stream_sender.cpp
+++ b/be/src/vec/sink/vdata_stream_sender.cpp
@@ -17,19 +17,38 @@
 
 #include "vec/sink/vdata_stream_sender.h"
 
+#include <butil/iobuf_inl.h>
 #include <fmt/format.h>
-#include <fmt/ranges.h>
+#include <fmt/ranges.h> // IWYU pragma: keep
+#include <gen_cpp/DataSinks_types.h>
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/data.pb.h>
+#include <gen_cpp/internal_service.pb.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+#include <stddef.h>
 
+#include <algorithm>
+#include <functional>
+#include <map>
 #include <random>
 
+#include "common/object_pool.h"
 #include "common/status.h"
+#include "runtime/descriptors.h"
 #include "runtime/exec_env.h"
 #include "runtime/memory/mem_tracker.h"
+#include "runtime/query_statistics.h"
 #include "runtime/runtime_state.h"
 #include "runtime/thread_context.h"
+#include "runtime/types.h"
 #include "util/brpc_client_cache.h"
 #include "util/proto_util.h"
+#include "util/telemetry/telemetry.h"
+#include "vec/columns/column.h"
 #include "vec/common/sip_hash.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/exprs/vexpr.h"
 #include "vec/runtime/vdata_stream_mgr.h"
 #include "vec/runtime/vdata_stream_recvr.h"
 
diff --git a/be/src/vec/sink/vdata_stream_sender.h b/be/src/vec/sink/vdata_stream_sender.h
index d0357cdbcc..2e78d2dbf9 100644
--- a/be/src/vec/sink/vdata_stream_sender.h
+++ b/be/src/vec/sink/vdata_stream_sender.h
@@ -17,36 +17,54 @@
 
 #pragma once
 
-#include <memory>
+#include <brpc/controller.h>
+#include <butil/errno.h>
+#include <fmt/format.h>
+#include <gen_cpp/Partitions_types.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/data.pb.h>
+#include <gen_cpp/internal_service.pb.h>
+#include <gen_cpp/types.pb.h>
+#include <stdint.h>
 
+#include <atomic>
+#include <memory>
+#include <ostream>
+#include <string>
+#include <vector>
+
+#include "common/config.h"
 #include "common/global_types.h"
+#include "common/logging.h"
+#include "common/status.h"
 #include "exec/data_sink.h"
-#include "gen_cpp/PaloInternalService_types.h"
-#include "gen_cpp/data.pb.h"
-#include "gen_cpp/internal_service.pb.h"
 #include "pipeline/exec/exchange_sink_buffer.h"
-#include "runtime/descriptors.h"
 #include "service/backend_options.h"
 #include "util/ref_count_closure.h"
+#include "util/runtime_profile.h"
 #include "util/uid_util.h"
-#include "vec/exprs/vexpr.h"
+#include "vec/core/block.h"
 #include "vec/exprs/vexpr_context.h"
-#include "vec/runtime/vdata_stream_mgr.h"
 #include "vec/runtime/vdata_stream_recvr.h"
 
 namespace doris {
 class ObjectPool;
 class RuntimeState;
-class RuntimeProfile;
-class BufferControlBlock;
 class MemTracker;
+class RowDescriptor;
+class TDataSink;
+class TDataStreamSink;
+class TPlanFragmentDestination;
+
+namespace segment_v2 {
+enum CompressionTypePB : int;
+} // namespace segment_v2
 
 namespace pipeline {
 class ExchangeSinkOperator;
 }
 
 namespace vectorized {
-class VExprContext;
 class Channel;
 
 template <typename T>
diff --git a/be/src/vec/sink/vjdbc_table_sink.cpp b/be/src/vec/sink/vjdbc_table_sink.cpp
index 973342da73..671c736e12 100644
--- a/be/src/vec/sink/vjdbc_table_sink.cpp
+++ b/be/src/vec/sink/vjdbc_table_sink.cpp
@@ -18,13 +18,23 @@
 #include "vec/sink/vjdbc_table_sink.h"
 
 #include <gen_cpp/DataSinks_types.h>
+#include <gen_cpp/Descriptors_types.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+#include <stdint.h>
 
-#include <sstream>
-
+#include "exec/data_sink.h"
+#include "runtime/runtime_state.h"
+#include "util/telemetry/telemetry.h"
+#include "vec/core/block.h"
 #include "vec/core/materialize_block.h"
+#include "vec/exprs/vexpr_context.h"
 #include "vec/sink/vtable_sink.h"
 
 namespace doris {
+class ObjectPool;
+class RowDescriptor;
+class TExpr;
+
 namespace vectorized {
 
 VJdbcTableSink::VJdbcTableSink(ObjectPool* pool, const RowDescriptor& row_desc,
diff --git a/be/src/vec/sink/vjdbc_table_sink.h b/be/src/vec/sink/vjdbc_table_sink.h
index c313cc8214..f091d7e138 100644
--- a/be/src/vec/sink/vjdbc_table_sink.h
+++ b/be/src/vec/sink/vjdbc_table_sink.h
@@ -15,12 +15,22 @@
 // specific language governing permissions and limitations
 // under the License.
 #pragma once
+#include <memory>
+#include <vector>
+
 #include "common/status.h"
 #include "vec/exec/vjdbc_connector.h"
 #include "vec/sink/vtable_sink.h"
 
 namespace doris {
+class ObjectPool;
+class RowDescriptor;
+class RuntimeState;
+class TDataSink;
+class TExpr;
+
 namespace vectorized {
+class Block;
 
 // This class is a sinker, which put input data to jdbc table
 class VJdbcTableSink : public VTableSink {
diff --git a/be/src/vec/sink/vmemory_scratch_sink.cpp b/be/src/vec/sink/vmemory_scratch_sink.cpp
index 79fdbb2425..480ad10e13 100644
--- a/be/src/vec/sink/vmemory_scratch_sink.cpp
+++ b/be/src/vec/sink/vmemory_scratch_sink.cpp
@@ -17,19 +17,31 @@
 
 #include "vec/sink/vmemory_scratch_sink.h"
 
-#include <arrow/memory_pool.h>
-#include <arrow/record_batch.h>
+#include <arrow/type_fwd.h>
+#include <gen_cpp/Types_types.h>
 
 #include <sstream>
 
-#include "gen_cpp/Types_types.h"
+#include "common/object_pool.h"
 #include "runtime/exec_env.h"
-#include "runtime/primitive_type.h"
+#include "runtime/record_batch_queue.h"
 #include "runtime/runtime_state.h"
 #include "util/arrow/block_convertor.h"
 #include "util/arrow/row_batch.h"
+#include "util/runtime_profile.h"
+#include "vec/core/block.h"
 #include "vec/exprs/vexpr.h"
 
+namespace arrow {
+class RecordBatch;
+} // namespace arrow
+
+namespace doris {
+class RowDescriptor;
+class TExpr;
+class TMemoryScratchSink;
+} // namespace doris
+
 namespace doris::vectorized {
 
 MemoryScratchSink::MemoryScratchSink(const RowDescriptor& row_desc,
diff --git a/be/src/vec/sink/vmemory_scratch_sink.h b/be/src/vec/sink/vmemory_scratch_sink.h
index 85a070d534..47a9984b8f 100644
--- a/be/src/vec/sink/vmemory_scratch_sink.h
+++ b/be/src/vec/sink/vmemory_scratch_sink.h
@@ -17,32 +17,31 @@
 
 #pragma once
 
+#include <memory>
+#include <vector>
+
 #include "common/status.h"
 #include "exec/data_sink.h"
-#include "gen_cpp/DorisExternalService_types.h"
-#include "gen_cpp/PlanNodes_types.h"
 #include "runtime/result_queue_mgr.h"
-#include "util/blocking_queue.hpp"
 
 namespace arrow {
 
-class MemoryPool;
-class RecordBatch;
 class Schema;
 
 } // namespace arrow
 
 namespace doris {
 
-class ObjectPool;
 class ObjectPool;
 class RuntimeState;
 class RuntimeProfile;
-class BufferControlBlock;
-class ResultWriter;
+class RowDescriptor;
+class TExpr;
+class TMemoryScratchSink;
 
 namespace vectorized {
 class VExprContext;
+class Block;
 
 // used to push data to blocking queue
 class MemoryScratchSink final : public DataSink {
diff --git a/be/src/vec/sink/vmysql_result_writer.cpp b/be/src/vec/sink/vmysql_result_writer.cpp
index ca09c16663..fcb133db10 100644
--- a/be/src/vec/sink/vmysql_result_writer.cpp
+++ b/be/src/vec/sink/vmysql_result_writer.cpp
@@ -17,19 +17,51 @@
 
 #include "vec/sink/vmysql_result_writer.h"
 
+#include <gen_cpp/Data_types.h>
+#include <gen_cpp/Metrics_types.h>
+#include <glog/logging.h>
+#include <stdint.h>
+#include <string.h>
+#include <sys/types.h>
+
+#include <ostream>
+#include <string>
+#include <utility>
+
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "gutil/integral_types.h"
 #include "olap/hll.h"
 #include "runtime/buffer_control_block.h"
-#include "runtime/jsonb_value.h"
+#include "runtime/decimalv2_value.h"
 #include "runtime/large_int_value.h"
+#include "runtime/primitive_type.h"
 #include "runtime/runtime_state.h"
+#include "runtime/types.h"
+#include "util/binary_cast.hpp"
+#include "util/bitmap_value.h"
+#include "util/jsonb_utils.h"
+#include "util/quantile_state.h"
+#include "vec/aggregate_functions/aggregate_function.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
 #include "vec/columns/column_complex.h"
+#include "vec/columns/column_decimal.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/columns/column_struct.h"
 #include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
 #include "vec/common/assert_cast.h"
+#include "vec/common/pod_array.h"
+#include "vec/common/string_ref.h"
+#include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
+#include "vec/core/field.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_array.h"
 #include "vec/data_types/data_type_decimal.h"
 #include "vec/data_types/data_type_map.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_struct.h"
 #include "vec/exprs/vexpr.h"
 #include "vec/exprs/vexpr_context.h"
diff --git a/be/src/vec/sink/vmysql_result_writer.h b/be/src/vec/sink/vmysql_result_writer.h
index 07a44fa0af..572f7313b2 100644
--- a/be/src/vec/sink/vmysql_result_writer.h
+++ b/be/src/vec/sink/vmysql_result_writer.h
@@ -16,18 +16,26 @@
 // under the License.
 
 #pragma once
-#include "runtime/primitive_type.h"
+#include <gen_cpp/PaloInternalService_types.h>
+#include <stddef.h>
+
+#include <memory>
+#include <vector>
+
+#include "common/status.h"
+#include "runtime/define_primitive_type.h"
 #include "util/mysql_row_buffer.h"
 #include "util/runtime_profile.h"
-#include "vec/core/block.h"
+#include "vec/data_types/data_type.h"
 #include "vec/sink/vresult_writer.h"
 
 namespace doris {
 class BufferControlBlock;
-class TFetchDataResult;
+class RuntimeState;
 
 namespace vectorized {
 class VExprContext;
+class Block;
 
 template <bool is_binary_format = false>
 class VMysqlResultWriter final : public VResultWriter {
diff --git a/be/src/vec/sink/vmysql_table_sink.cpp b/be/src/vec/sink/vmysql_table_sink.cpp
index ee9bc8095b..e2ade983db 100644
--- a/be/src/vec/sink/vmysql_table_sink.cpp
+++ b/be/src/vec/sink/vmysql_table_sink.cpp
@@ -17,13 +17,21 @@
 
 #include "vec/sink/vmysql_table_sink.h"
 
-#include <sstream>
+#include <gen_cpp/DataSinks_types.h>
+#include <opentelemetry/nostd/shared_ptr.h>
 
 #include "runtime/runtime_state.h"
+#include "util/telemetry/telemetry.h"
 #include "vec/sink/vtable_sink.h"
 
 namespace doris {
+class ObjectPool;
+class RowDescriptor;
+class TExpr;
+
 namespace vectorized {
+class Block;
+
 VMysqlTableSink::VMysqlTableSink(ObjectPool* pool, const RowDescriptor& row_desc,
                                  const std::vector<TExpr>& t_exprs)
         : VTableSink(pool, row_desc, t_exprs) {
diff --git a/be/src/vec/sink/vmysql_table_sink.h b/be/src/vec/sink/vmysql_table_sink.h
index 9d41fe8a84..00ce4346fe 100644
--- a/be/src/vec/sink/vmysql_table_sink.h
+++ b/be/src/vec/sink/vmysql_table_sink.h
@@ -15,6 +15,7 @@
 // specific language governing permissions and limitations
 // under the License.
 #pragma once
+#include <memory>
 #include <vector>
 
 #include "common/status.h"
@@ -22,7 +23,14 @@
 #include "vec/sink/vtable_sink.h"
 
 namespace doris {
+class ObjectPool;
+class RowDescriptor;
+class RuntimeState;
+class TDataSink;
+class TExpr;
+
 namespace vectorized {
+class Block;
 
 // This class is a sinker, which put input data to mysql table
 class VMysqlTableSink : public VTableSink {
diff --git a/be/src/vec/sink/vmysql_table_writer.cpp b/be/src/vec/sink/vmysql_table_writer.cpp
index 03a38e302b..661b1c7f00 100644
--- a/be/src/vec/sink/vmysql_table_writer.cpp
+++ b/be/src/vec/sink/vmysql_table_writer.cpp
@@ -17,17 +17,36 @@
 
 #include "vec/sink/vmysql_table_writer.h"
 
+#include <glog/logging.h>
 #include <mysql/mysql.h>
+#include <stdint.h>
 
+#include <memory>
 #include <sstream>
 
-#include "util/types.h"
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
+#include "runtime/decimalv2_value.h"
+#include "runtime/define_primitive_type.h"
+#include "runtime/types.h"
+#include "util/binary_cast.hpp"
+#include "vec/columns/column.h"
+#include "vec/columns/column_decimal.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
+#include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/pod_array.h"
+#include "vec/common/string_ref.h"
 #include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/core/materialize_block.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type.h"
 #include "vec/exprs/vexpr.h"
 #include "vec/exprs/vexpr_context.h"
+#include "vec/runtime/vdatetime_value.h"
 
 namespace doris {
 namespace vectorized {
diff --git a/be/src/vec/sink/vmysql_table_writer.h b/be/src/vec/sink/vmysql_table_writer.h
index 5c494262c2..54a8b470c9 100644
--- a/be/src/vec/sink/vmysql_table_writer.h
+++ b/be/src/vec/sink/vmysql_table_writer.h
@@ -19,6 +19,7 @@
 
 #include <fmt/format.h>
 #include <mysql/mysql.h>
+#include <stddef.h>
 
 #include <string>
 #include <vector>
@@ -41,6 +42,7 @@ struct MysqlConnInfo {
 
 class VExprContext;
 class Block;
+
 class VMysqlTableWriter {
 public:
     VMysqlTableWriter(const std::vector<vectorized::VExprContext*>& output_exprs);
diff --git a/be/src/vec/sink/vodbc_table_sink.cpp b/be/src/vec/sink/vodbc_table_sink.cpp
index b2012cb911..1cdf1a6d24 100644
--- a/be/src/vec/sink/vodbc_table_sink.cpp
+++ b/be/src/vec/sink/vodbc_table_sink.cpp
@@ -17,14 +17,23 @@
 
 #include "vec/sink/vodbc_table_sink.h"
 
-#include <sstream>
+#include <gen_cpp/DataSinks_types.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+#include <stdint.h>
 
+#include "exec/data_sink.h"
 #include "runtime/runtime_state.h"
-#include "util/runtime_profile.h"
+#include "util/telemetry/telemetry.h"
+#include "vec/core/block.h"
 #include "vec/core/materialize_block.h"
+#include "vec/exprs/vexpr_context.h"
 #include "vec/sink/vtable_sink.h"
 
 namespace doris {
+class ObjectPool;
+class RowDescriptor;
+class TExpr;
+
 namespace vectorized {
 
 VOdbcTableSink::VOdbcTableSink(ObjectPool* pool, const RowDescriptor& row_desc,
diff --git a/be/src/vec/sink/vodbc_table_sink.h b/be/src/vec/sink/vodbc_table_sink.h
index 7218edf10f..5a9f1af7d8 100644
--- a/be/src/vec/sink/vodbc_table_sink.h
+++ b/be/src/vec/sink/vodbc_table_sink.h
@@ -16,12 +16,22 @@
 // under the License.
 #pragma once
 
+#include <memory>
+#include <vector>
+
 #include "common/status.h"
 #include "exec/odbc_connector.h"
 #include "vec/sink/vtable_sink.h"
 
 namespace doris {
+class ObjectPool;
+class RowDescriptor;
+class RuntimeState;
+class TDataSink;
+class TExpr;
+
 namespace vectorized {
+class Block;
 
 // This class is a sinker, which put input data to odbc table
 class VOdbcTableSink : public VTableSink {
diff --git a/be/src/vec/sink/vresult_file_sink.cpp b/be/src/vec/sink/vresult_file_sink.cpp
index 07d0c45927..f96649712b 100644
--- a/be/src/vec/sink/vresult_file_sink.cpp
+++ b/be/src/vec/sink/vresult_file_sink.cpp
@@ -17,14 +17,32 @@
 
 #include "vec/sink/vresult_file_sink.h"
 
+#include <gen_cpp/DataSinks_types.h>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <glog/logging.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+#include <time.h>
+
+#include <new>
+#include <ostream>
+
 #include "common/config.h"
+#include "common/object_pool.h"
 #include "runtime/buffer_control_block.h"
 #include "runtime/exec_env.h"
 #include "runtime/result_buffer_mgr.h"
 #include "runtime/runtime_state.h"
+#include "util/runtime_profile.h"
+#include "util/telemetry/telemetry.h"
 #include "util/uid_util.h"
 #include "vec/exprs/vexpr.h"
 #include "vec/runtime/vfile_result_writer.h"
+#include "vec/sink/vresult_writer.h"
+
+namespace doris {
+class QueryStatistics;
+class TExpr;
+} // namespace doris
 
 namespace doris::vectorized {
 
diff --git a/be/src/vec/sink/vresult_file_sink.h b/be/src/vec/sink/vresult_file_sink.h
index 038d825acc..ed62b3c60a 100644
--- a/be/src/vec/sink/vresult_file_sink.h
+++ b/be/src/vec/sink/vresult_file_sink.h
@@ -17,12 +17,33 @@
 
 #pragma once
 
+#include <gen_cpp/Types_types.h>
+
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "common/status.h"
+#include "exec/data_sink.h"
+#include "runtime/descriptors.h"
+#include "vec/core/block.h"
 #include "vec/sink/vdata_stream_sender.h"
 #include "vec/sink/vresult_sink.h"
 
 namespace doris {
+class BufferControlBlock;
+class ObjectPool;
+class QueryStatistics;
+class RuntimeProfile;
+class RuntimeState;
+class TDataSink;
+class TExpr;
+class TPlanFragmentDestination;
+class TResultFileSink;
+
 namespace vectorized {
 class VResultWriter;
+class VExprContext;
 
 class VResultFileSink : public DataSink {
 public:
diff --git a/be/src/vec/sink/vresult_sink.cpp b/be/src/vec/sink/vresult_sink.cpp
index bd49319b8f..15d41f1f02 100644
--- a/be/src/vec/sink/vresult_sink.cpp
+++ b/be/src/vec/sink/vresult_sink.cpp
@@ -17,15 +17,31 @@
 
 #include "vec/sink/vresult_sink.h"
 
+#include <fmt/format.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+#include <time.h>
+
+#include <new>
+
+#include "common/config.h"
+#include "common/object_pool.h"
 #include "runtime/buffer_control_block.h"
 #include "runtime/exec_env.h"
 #include "runtime/result_buffer_mgr.h"
 #include "runtime/runtime_state.h"
+#include "util/runtime_profile.h"
+#include "util/telemetry/telemetry.h"
 #include "vec/exprs/vexpr.h"
 #include "vec/sink/vmysql_result_writer.h"
+#include "vec/sink/vresult_writer.h"
 
 namespace doris {
+class QueryStatistics;
+class RowDescriptor;
+class TExpr;
+
 namespace vectorized {
+class Block;
 
 VResultSink::VResultSink(const RowDescriptor& row_desc, const std::vector<TExpr>& t_output_expr,
                          const TResultSink& sink, int buffer_size)
diff --git a/be/src/vec/sink/vresult_sink.h b/be/src/vec/sink/vresult_sink.h
index 9c02877571..50a7b0141d 100644
--- a/be/src/vec/sink/vresult_sink.h
+++ b/be/src/vec/sink/vresult_sink.h
@@ -16,22 +16,34 @@
 // under the License.
 
 #pragma once
+#include <gen_cpp/DataSinks_types.h>
+#include <gen_cpp/PlanNodes_types.h>
+#include <gen_cpp/Types_types.h>
+#include <stddef.h>
+
+#include <map>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "common/status.h"
 #include "exec/data_sink.h"
-#include "vec/sink/vresult_writer.h"
 
 namespace doris {
-class ObjectPool;
 class RuntimeState;
 class RuntimeProfile;
 class BufferControlBlock;
-class ResultWriter;
-class MemTracker;
-struct ResultFileOptions;
+class QueryStatistics;
+class RowDescriptor;
+class TExpr;
+
 namespace pipeline {
 class ResultSinkOperator;
 }
 namespace vectorized {
 class VExprContext;
+class Block;
+class VResultWriter;
 
 struct ResultFileOptions {
     // [[deprecated]]
diff --git a/be/src/vec/sink/vresult_writer.h b/be/src/vec/sink/vresult_writer.h
index 5bd590f14d..66c15e9cd6 100644
--- a/be/src/vec/sink/vresult_writer.h
+++ b/be/src/vec/sink/vresult_writer.h
@@ -21,6 +21,9 @@
 
 namespace doris {
 namespace vectorized {
+
+class Block;
+
 class VResultWriter : public ResultWriter {
 public:
     VResultWriter() : ResultWriter() {}
@@ -30,4 +33,4 @@ public:
     virtual bool can_sink() { return true; }
 };
 } // namespace vectorized
-} // namespace doris
\ No newline at end of file
+} // namespace doris
diff --git a/be/src/vec/sink/vtable_sink.cpp b/be/src/vec/sink/vtable_sink.cpp
index a09fb2cb5f..1a34c20e91 100644
--- a/be/src/vec/sink/vtable_sink.cpp
+++ b/be/src/vec/sink/vtable_sink.cpp
@@ -17,13 +17,22 @@
 
 #include "vec/sink/vtable_sink.h"
 
+#include <gen_cpp/Types_types.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+
 #include <sstream>
 
+#include "common/object_pool.h"
 #include "runtime/runtime_state.h"
+#include "util/runtime_profile.h"
+#include "util/telemetry/telemetry.h"
 #include "vec/exprs/vexpr.h"
 
 namespace doris {
+class TDataSink;
+
 namespace vectorized {
+class Block;
 
 VTableSink::VTableSink(ObjectPool* pool, const RowDescriptor& row_desc,
                        const std::vector<TExpr>& t_exprs)
diff --git a/be/src/vec/sink/vtable_sink.h b/be/src/vec/sink/vtable_sink.h
index 339df25f5c..325d44566c 100644
--- a/be/src/vec/sink/vtable_sink.h
+++ b/be/src/vec/sink/vtable_sink.h
@@ -16,9 +16,11 @@
 // under the License.
 #pragma once
 
+#include <string>
+#include <vector>
+
 #include "common/status.h"
 #include "exec/data_sink.h"
-#include "vec/exprs/vexpr_context.h"
 
 namespace doris {
 
@@ -26,7 +28,12 @@ class RowDescriptor;
 class TExpr;
 class RuntimeState;
 class RuntimeProfile;
+class ObjectPool;
+class TDataSink;
+
 namespace vectorized {
+class Block;
+class VExprContext;
 
 class VTableSink : public DataSink {
 public:
diff --git a/be/src/vec/sink/vtablet_sink.cpp b/be/src/vec/sink/vtablet_sink.cpp
index d1bcf7fcca..6abc3fb616 100644
--- a/be/src/vec/sink/vtablet_sink.cpp
+++ b/be/src/vec/sink/vtablet_sink.cpp
@@ -17,42 +17,76 @@
 
 #include "vec/sink/vtablet_sink.h"
 
+#include <brpc/http_header.h>
+#include <brpc/http_method.h>
+#include <brpc/uri.h>
+#include <bthread/bthread.h>
+#include <butil/iobuf_inl.h>
 #include <fmt/format.h>
+#include <gen_cpp/DataSinks_types.h>
+#include <gen_cpp/Descriptors_types.h>
+#include <gen_cpp/Metrics_types.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/data.pb.h>
+#include <gen_cpp/internal_service.pb.h>
+#include <opentelemetry/nostd/shared_ptr.h>
+#include <sys/param.h>
+#include <sys/types.h>
 
+#include <algorithm>
+#include <iterator>
 #include <mutex>
-#include <sstream>
 #include <string>
 #include <unordered_map>
 
+// IWYU pragma: no_include <opentelemetry/common/threadlocal.h>
+#include "common/compiler_util.h" // IWYU pragma: keep
 #include "common/logging.h"
+#include "common/object_pool.h"
 #include "common/status.h"
 #include "exec/tablet_info.h"
-#include "olap/hll.h"
+#include "runtime/define_primitive_type.h"
+#include "runtime/descriptors.h"
 #include "runtime/exec_env.h"
 #include "runtime/runtime_state.h"
 #include "runtime/thread_context.h"
 #include "service/backend_options.h"
+#include "util/binary_cast.hpp"
 #include "util/brpc_client_cache.h"
 #include "util/debug/sanitizer_scopes.h"
 #include "util/defer_op.h"
 #include "util/doris_metrics.h"
+#include "util/network_util.h"
 #include "util/proto_util.h"
+#include "util/ref_count_closure.h"
+#include "util/telemetry/telemetry.h"
+#include "util/thread.h"
 #include "util/threadpool.h"
 #include "util/time.h"
 #include "util/uid_util.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
 #include "vec/columns/column_const.h"
+#include "vec/columns/column_decimal.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
 #include "vec/columns/column_struct.h"
+#include "vec/columns/column_vector.h"
 #include "vec/columns/columns_number.h"
+#include "vec/common/assert_cast.h"
+#include "vec/common/pod_array.h"
+#include "vec/common/string_ref.h"
 #include "vec/core/block.h"
+#include "vec/core/column_with_type_and_name.h"
 #include "vec/core/types.h"
 #include "vec/data_types/data_type_decimal.h"
+#include "vec/data_types/data_type_nullable.h"
 #include "vec/exprs/vexpr.h"
 #include "vec/exprs/vexpr_context.h"
 
 namespace doris {
+class TExpr;
+
 namespace stream_load {
 
 IndexChannel::~IndexChannel() {
diff --git a/be/src/vec/sink/vtablet_sink.h b/be/src/vec/sink/vtablet_sink.h
index 6f4492a37b..61d64a2bdb 100644
--- a/be/src/vec/sink/vtablet_sink.h
+++ b/be/src/vec/sink/vtablet_sink.h
@@ -17,33 +17,67 @@
 
 #pragma once
 #include <brpc/controller.h>
+#include <bthread/types.h>
+#include <butil/errno.h>
+#include <fmt/format.h>
+#include <gen_cpp/PaloInternalService_types.h>
+#include <gen_cpp/Types_types.h>
+#include <gen_cpp/internal_service.pb.h>
+#include <gen_cpp/types.pb.h>
+#include <glog/logging.h>
+#include <google/protobuf/stubs/callback.h>
+#include <stddef.h>
+#include <stdint.h>
 
+#include <atomic>
+// IWYU pragma: no_include <bits/chrono.h>
+#include <chrono> // IWYU pragma: keep
+#include <functional>
+#include <initializer_list>
+#include <map>
 #include <memory>
+#include <mutex>
+#include <ostream>
 #include <queue>
+#include <set>
 #include <string>
+#include <thread>
 #include <unordered_map>
 #include <unordered_set>
 #include <utility>
 #include <vector>
 
-#include "common/object_pool.h"
+#include "common/config.h"
 #include "common/status.h"
 #include "exec/data_sink.h"
 #include "exec/tablet_info.h"
-#include "gen_cpp/Types_types.h"
-#include "gen_cpp/internal_service.pb.h"
+#include "gutil/ref_counted.h"
+#include "runtime/decimalv2_value.h"
+#include "runtime/exec_env.h"
+#include "runtime/memory/mem_tracker.h"
 #include "runtime/thread_context.h"
+#include "runtime/types.h"
 #include "util/bitmap.h"
 #include "util/countdown_latch.h"
-#include "util/ref_count_closure.h"
+#include "util/runtime_profile.h"
 #include "util/spinlock.h"
-#include "util/thread.h"
+#include "util/stopwatch.hpp"
 #include "vec/columns/column.h"
-#include "vec/columns/columns_number.h"
+#include "vec/common/allocator.h"
 #include "vec/core/block.h"
-#include "vec/exprs/vexpr_context.h"
+#include "vec/data_types/data_type.h"
 
 namespace doris {
+class ObjectPool;
+class RowDescriptor;
+class RuntimeState;
+class TDataSink;
+class TExpr;
+class Thread;
+class ThreadPoolToken;
+class TupleDescriptor;
+template <typename T>
+class RefCountClosure;
 
 namespace vectorized {
 class VExprContext;
diff --git a/be/src/vec/utils/arrow_column_to_doris_column.cpp b/be/src/vec/utils/arrow_column_to_doris_column.cpp
index c0d64df211..af94664f1b 100644
--- a/be/src/vec/utils/arrow_column_to_doris_column.cpp
+++ b/be/src/vec/utils/arrow_column_to_doris_column.cpp
@@ -17,17 +17,34 @@
 
 #include "vec/utils/arrow_column_to_doris_column.h"
 
-#include <arrow/array.h>
-#include <arrow/record_batch.h>
-#include <arrow/status.h>
+#include <arrow/array/array_base.h>
+#include <arrow/array/array_decimal.h>
+#include <arrow/array/array_primitive.h>
+#include <arrow/array/data.h>
+#include <arrow/buffer.h>
+#include <cctz/time_zone.h>
+#include <fmt/format.h>
+#include <glog/logging.h>
+#include <stdint.h>
+
+#include <memory>
+#include <utility>
+#include <vector>
 
 #include "arrow/array/array_binary.h"
 #include "arrow/array/array_nested.h"
 #include "arrow/type.h"
-#include "arrow/type_fwd.h"
 #include "gutil/casts.h"
+#include "util/binary_cast.hpp"
+#include "util/timezone_utils.h"
+#include "vec/columns/column.h"
 #include "vec/columns/column_array.h"
+#include "vec/columns/column_decimal.h"
 #include "vec/columns/column_nullable.h"
+#include "vec/columns/column_string.h"
+#include "vec/columns/column_vector.h"
+#include "vec/common/assert_cast.h"
+#include "vec/core/types.h"
 #include "vec/data_types/data_type_array.h"
 #include "vec/data_types/data_type_decimal.h"
 #include "vec/data_types/data_type_nullable.h"
diff --git a/be/src/vec/utils/arrow_column_to_doris_column.h b/be/src/vec/utils/arrow_column_to_doris_column.h
index 73704c95f5..285030bb81 100644
--- a/be/src/vec/utils/arrow_column_to_doris_column.h
+++ b/be/src/vec/utils/arrow_column_to_doris_column.h
@@ -17,15 +17,21 @@
 
 #pragma once
 
-#include <arrow/type.h>
 #include <arrow/type_fwd.h>
+#include <stddef.h>
 
-#include <iostream>
-#include <memory>
+#include <string>
 
 #include "common/status.h"
-#include "runtime/types.h"
-#include "vec/core/column_with_type_and_name.h"
+#include "runtime/define_primitive_type.h"
+#include "vec/data_types/data_type.h"
+
+namespace arrow {
+class Array;
+} // namespace arrow
+namespace cctz {
+class time_zone;
+} // namespace cctz
 
 // This files contains some utilities to convert Doris internal
 // data format from Apache Arrow format.
diff --git a/be/src/vec/utils/util.hpp b/be/src/vec/utils/util.hpp
index cb01c36eb3..97ab412160 100644
--- a/be/src/vec/utils/util.hpp
+++ b/be/src/vec/utils/util.hpp
@@ -25,6 +25,7 @@
 #include "vec/columns/column_nullable.h"
 #include "vec/core/block.h"
 #include "vec/exprs/vexpr.h"
+#include "vec/exprs/vexpr_context.h"
 
 namespace doris::vectorized {
 class VectorizedUtils {
diff --git a/be/test/http/http_client_test.cpp b/be/test/http/http_client_test.cpp
index 8dceb8d494..f2c443cd85 100644
--- a/be/test/http/http_client_test.cpp
+++ b/be/test/http/http_client_test.cpp
@@ -24,7 +24,9 @@
 #include "http/ev_http_server.h"
 #include "http/http_channel.h"
 #include "http/http_handler.h"
+#include "http/http_headers.h"
 #include "http/http_request.h"
+#include "http/utils.h"
 
 namespace doris {
 
diff --git a/be/test/io/cache/remote_file_cache_test.cpp b/be/test/io/cache/remote_file_cache_test.cpp
index e2f43657a6..89b296c13d 100644
--- a/be/test/io/cache/remote_file_cache_test.cpp
+++ b/be/test/io/cache/remote_file_cache_test.cpp
@@ -25,6 +25,7 @@
 
 #include "gen_cpp/olap_file.pb.h"
 #include "gtest/gtest.h"
+#include "io/fs/file_writer.h"
 #include "io/fs/local_file_system.h"
 #include "io/fs/s3_common.h"
 #include "io/fs/s3_file_system.h"
diff --git a/be/test/olap/delete_handler_test.cpp b/be/test/olap/delete_handler_test.cpp
index 88669e1e83..ec4d62f7aa 100644
--- a/be/test/olap/delete_handler_test.cpp
+++ b/be/test/olap/delete_handler_test.cpp
@@ -17,6 +17,7 @@
 
 #include "olap/delete_handler.h"
 
+#include <gen_cpp/Descriptors_types.h>
 #include <gtest/gtest.h>
 #include <unistd.h>
 
@@ -31,8 +32,10 @@
 #include "olap/olap_define.h"
 #include "olap/options.h"
 #include "olap/push_handler.h"
+#include "olap/row_cursor.h"
 #include "olap/rowset/beta_rowset.h"
 #include "olap/storage_engine.h"
+#include "olap/tablet_manager.h"
 #include "olap/utils.h"
 #include "util/cpu_info.h"
 
diff --git a/be/test/olap/delta_writer_test.cpp b/be/test/olap/delta_writer_test.cpp
index 99dec485d3..857e33a13c 100644
--- a/be/test/olap/delta_writer_test.cpp
+++ b/be/test/olap/delta_writer_test.cpp
@@ -33,9 +33,12 @@
 #include "olap/rowset/beta_rowset.h"
 #include "olap/storage_engine.h"
 #include "olap/tablet.h"
+#include "olap/tablet_manager.h"
 #include "olap/tablet_meta_manager.h"
+#include "olap/txn_manager.h"
 #include "olap/utils.h"
 #include "runtime/descriptor_helper.h"
+#include "runtime/descriptors.h"
 #include "runtime/exec_env.h"
 
 namespace doris {
diff --git a/be/test/olap/engine_storage_migration_task_test.cpp b/be/test/olap/engine_storage_migration_task_test.cpp
index 5987a533f1..3aa138ddcb 100644
--- a/be/test/olap/engine_storage_migration_task_test.cpp
+++ b/be/test/olap/engine_storage_migration_task_test.cpp
@@ -36,6 +36,7 @@
 #include "olap/tablet_meta_manager.h"
 #include "olap/utils.h"
 #include "runtime/descriptor_helper.h"
+#include "runtime/descriptors.h"
 #include "runtime/exec_env.h"
 
 namespace doris {
diff --git a/be/test/olap/ordered_data_compaction_test.cpp b/be/test/olap/ordered_data_compaction_test.cpp
index df5bf117c1..fa73e4e901 100644
--- a/be/test/olap/ordered_data_compaction_test.cpp
+++ b/be/test/olap/ordered_data_compaction_test.cpp
@@ -30,6 +30,7 @@
 #include "olap/rowset/rowset_writer.h"
 #include "olap/rowset/rowset_writer_context.h"
 #include "olap/schema.h"
+#include "olap/storage_engine.h"
 #include "olap/tablet_schema.h"
 #include "olap/tablet_schema_helper.h"
 #include "vec/olap/vertical_block_reader.h"
diff --git a/be/test/olap/remote_rowset_gc_test.cpp b/be/test/olap/remote_rowset_gc_test.cpp
index f6e0948e82..72b8f3a10a 100644
--- a/be/test/olap/remote_rowset_gc_test.cpp
+++ b/be/test/olap/remote_rowset_gc_test.cpp
@@ -29,7 +29,10 @@
 #include "olap/storage_engine.h"
 #include "olap/storage_policy.h"
 #include "olap/tablet.h"
+#include "olap/tablet_manager.h"
+#include "olap/txn_manager.h"
 #include "runtime/descriptor_helper.h"
+#include "runtime/descriptors.h"
 #include "util/s3_util.h"
 
 namespace doris {
diff --git a/be/test/olap/rowset/segment_v2/encoding_info_test.cpp b/be/test/olap/rowset/segment_v2/encoding_info_test.cpp
index 530c619695..cb4bc2502a 100644
--- a/be/test/olap/rowset/segment_v2/encoding_info_test.cpp
+++ b/be/test/olap/rowset/segment_v2/encoding_info_test.cpp
@@ -17,6 +17,7 @@
 
 #include "olap/rowset/segment_v2/encoding_info.h"
 
+#include <gen_cpp/segment_v2.pb.h>
 #include <gtest/gtest.h>
 
 #include <iostream>
diff --git a/be/test/olap/tablet_cooldown_test.cpp b/be/test/olap/tablet_cooldown_test.cpp
index d4a07c8dfd..4a8371e1e2 100644
--- a/be/test/olap/tablet_cooldown_test.cpp
+++ b/be/test/olap/tablet_cooldown_test.cpp
@@ -33,7 +33,10 @@
 #include "olap/storage_engine.h"
 #include "olap/storage_policy.h"
 #include "olap/tablet.h"
+#include "olap/tablet_manager.h"
+#include "olap/txn_manager.h"
 #include "runtime/descriptor_helper.h"
+#include "runtime/descriptors.h"
 #include "util/s3_util.h"
 
 namespace doris {
diff --git a/be/test/olap/tablet_meta_manager_test.cpp b/be/test/olap/tablet_meta_manager_test.cpp
index da170b5972..f4864d5940 100644
--- a/be/test/olap/tablet_meta_manager_test.cpp
+++ b/be/test/olap/tablet_meta_manager_test.cpp
@@ -25,6 +25,8 @@
 #include <sstream>
 #include <string>
 
+#include "olap/data_dir.h"
+
 using std::string;
 
 namespace doris {
diff --git a/be/test/olap/tablet_mgr_test.cpp b/be/test/olap/tablet_mgr_test.cpp
index 22d277ebe3..67818ad638 100644
--- a/be/test/olap/tablet_mgr_test.cpp
+++ b/be/test/olap/tablet_mgr_test.cpp
@@ -24,6 +24,7 @@
 #include "gtest/gtest.h"
 #include "io/fs/local_file_system.h"
 #include "olap/storage_engine.h"
+#include "olap/tablet_manager.h"
 #include "olap/tablet_meta_manager.h"
 #include "olap/txn_manager.h"
 
diff --git a/be/test/olap/tablet_schema_helper.cpp b/be/test/olap/tablet_schema_helper.cpp
index a3ddf2e71f..f675ca9d0f 100644
--- a/be/test/olap/tablet_schema_helper.cpp
+++ b/be/test/olap/tablet_schema_helper.cpp
@@ -17,6 +17,8 @@
 
 #include "olap/tablet_schema_helper.h"
 
+#include "vec/common/arena.h"
+
 namespace doris {
 
 TabletColumn create_int_key(int32_t id, bool is_nullable, bool is_bf_column,
diff --git a/be/test/olap/timestamped_version_tracker_test.cpp b/be/test/olap/timestamped_version_tracker_test.cpp
index 578349e794..c5b5c20462 100644
--- a/be/test/olap/timestamped_version_tracker_test.cpp
+++ b/be/test/olap/timestamped_version_tracker_test.cpp
@@ -17,6 +17,8 @@
 
 #include <cctz/time_zone.h>
 #include <gtest/gtest.h>
+#include <rapidjson/prettywriter.h>
+#include <rapidjson/stringbuffer.h>
 
 #include <filesystem>
 #include <fstream>
diff --git a/be/test/runtime/external_scan_context_mgr_test.cpp b/be/test/runtime/external_scan_context_mgr_test.cpp
index 29a045234b..2576fa88a3 100644
--- a/be/test/runtime/external_scan_context_mgr_test.cpp
+++ b/be/test/runtime/external_scan_context_mgr_test.cpp
@@ -23,6 +23,7 @@
 
 #include "common/config.h"
 #include "common/status.h"
+#include "runtime/exec_env.h"
 #include "runtime/fragment_mgr.h"
 #include "runtime/result_queue_mgr.h"
 
diff --git a/be/test/runtime/fragment_mgr_test.cpp b/be/test/runtime/fragment_mgr_test.cpp
index 9a3967c031..323628c44d 100644
--- a/be/test/runtime/fragment_mgr_test.cpp
+++ b/be/test/runtime/fragment_mgr_test.cpp
@@ -21,6 +21,7 @@
 
 #include "common/config.h"
 #include "exec/data_sink.h"
+#include "runtime/exec_env.h"
 #include "runtime/plan_fragment_executor.h"
 
 namespace doris {
diff --git a/be/test/runtime/result_queue_mgr_test.cpp b/be/test/runtime/result_queue_mgr_test.cpp
index 3403df0b8e..da7c5a0426 100644
--- a/be/test/runtime/result_queue_mgr_test.cpp
+++ b/be/test/runtime/result_queue_mgr_test.cpp
@@ -26,6 +26,7 @@
 #include <memory>
 
 #include "gen_cpp/DorisExternalService_types.h"
+#include "runtime/record_batch_queue.h"
 #include "testutil/test_util.h"
 #include "util/blocking_queue.hpp"
 
diff --git a/be/test/runtime/routine_load_task_executor_test.cpp b/be/test/runtime/routine_load_task_executor_test.cpp
index 4cc3122714..517ecaf383 100644
--- a/be/test/runtime/routine_load_task_executor_test.cpp
+++ b/be/test/runtime/routine_load_task_executor_test.cpp
@@ -18,6 +18,7 @@
 #include "runtime/routine_load/routine_load_task_executor.h"
 
 #include <gtest/gtest.h>
+#include <librdkafka/rdkafkacpp.h>
 
 #include "gen_cpp/BackendService_types.h"
 #include "gen_cpp/FrontendService_types.h"
diff --git a/be/test/testutil/desc_tbl_builder.cc b/be/test/testutil/desc_tbl_builder.cc
index a03a87b0d6..d27e464b16 100644
--- a/be/test/testutil/desc_tbl_builder.cc
+++ b/be/test/testutil/desc_tbl_builder.cc
@@ -21,6 +21,7 @@
 
 #include <vector>
 
+#include "common/object_pool.h"
 #include "runtime/descriptors.h"
 #include "util/bit_util.h"
 
diff --git a/be/test/testutil/run_all_tests.cpp b/be/test/testutil/run_all_tests.cpp
index 561ee34b2c..306971784d 100644
--- a/be/test/testutil/run_all_tests.cpp
+++ b/be/test/testutil/run_all_tests.cpp
@@ -21,6 +21,7 @@
 #include "olap/page_cache.h"
 #include "olap/rowset/segment_v2/inverted_index_cache.h"
 #include "olap/segment_loader.h"
+#include "olap/tablet_schema_cache.h"
 #include "runtime/exec_env.h"
 #include "runtime/memory/mem_tracker_limiter.h"
 #include "service/backend_options.h"
diff --git a/be/test/util/block_compression_test.cpp b/be/test/util/block_compression_test.cpp
index 867ee65d20..123b846cb3 100644
--- a/be/test/util/block_compression_test.cpp
+++ b/be/test/util/block_compression_test.cpp
@@ -17,6 +17,7 @@
 
 #include "util/block_compression.h"
 
+#include <gen_cpp/segment_v2.pb.h>
 #include <gtest/gtest.h>
 
 #include <iostream>
diff --git a/be/test/util/brpc_client_cache_test.cpp b/be/test/util/brpc_client_cache_test.cpp
index 13a6a09790..9c4efd5bf9 100644
--- a/be/test/util/brpc_client_cache_test.cpp
+++ b/be/test/util/brpc_client_cache_test.cpp
@@ -17,6 +17,8 @@
 
 #include "util/brpc_client_cache.h"
 
+#include <gen_cpp/function_service.pb.h>
+#include <gen_cpp/internal_service.pb.h>
 #include <gtest/gtest.h>
 
 namespace doris {
diff --git a/be/test/util/date_func_test.cpp b/be/test/util/date_func_test.cpp
index bfbe6b6c53..b39a371a15 100644
--- a/be/test/util/date_func_test.cpp
+++ b/be/test/util/date_func_test.cpp
@@ -21,6 +21,8 @@
 
 #include <iostream>
 
+#include "olap/uint24.h"
+
 namespace doris {
 
 class DateFuncTest : public testing::Test {
diff --git a/be/test/vec/aggregate_functions/agg_replace_test.cpp b/be/test/vec/aggregate_functions/agg_replace_test.cpp
index aa1e0cc8fa..e74df64aca 100644
--- a/be/test/vec/aggregate_functions/agg_replace_test.cpp
+++ b/be/test/vec/aggregate_functions/agg_replace_test.cpp
@@ -20,9 +20,12 @@
 #include "common/logging.h"
 #include "vec/aggregate_functions/aggregate_function.h"
 #include "vec/aggregate_functions/aggregate_function_reader.h"
+#include "vec/aggregate_functions/aggregate_function_reader_first_last.h"
 #include "vec/aggregate_functions/aggregate_function_simple_factory.h"
+#include "vec/columns/column_array.h"
 #include "vec/columns/column_complex.h"
 #include "vec/columns/column_vector.h"
+#include "vec/common/arena.h"
 #include "vec/data_types/data_type_array.h"
 #include "vec/data_types/data_type_bitmap.h"
 #include "vec/data_types/data_type_date.h"
diff --git a/be/test/vec/core/block_spill_test.cpp b/be/test/vec/core/block_spill_test.cpp
index 94e12931f2..8cf600f852 100644
--- a/be/test/vec/core/block_spill_test.cpp
+++ b/be/test/vec/core/block_spill_test.cpp
@@ -19,6 +19,7 @@
 
 #include "io/fs/local_file_system.h"
 #include "runtime/block_spill_manager.h"
+#include "runtime/exec_env.h"
 #include "runtime/runtime_state.h"
 #include "vec/columns/column_array.h"
 #include "vec/columns/column_decimal.h"
diff --git a/be/test/vec/core/block_test.cpp b/be/test/vec/core/block_test.cpp
index 4195be2a27..47da35cf04 100644
--- a/be/test/vec/core/block_test.cpp
+++ b/be/test/vec/core/block_test.cpp
@@ -41,6 +41,7 @@
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/data_types/data_type_string.h"
+#include "vec/data_types/data_type_time_v2.h"
 #include "vec/runtime/vdatetime_value.h"
 
 namespace doris {
diff --git a/be/test/vec/exec/parquet/parquet_reader_test.cpp b/be/test/vec/exec/parquet/parquet_reader_test.cpp
index ec4eb12fed..e85f976ff0 100644
--- a/be/test/vec/exec/parquet/parquet_reader_test.cpp
+++ b/be/test/vec/exec/parquet/parquet_reader_test.cpp
@@ -18,7 +18,9 @@
 #include <glog/logging.h>
 #include <gtest/gtest.h>
 
+#include "common/object_pool.h"
 #include "io/fs/local_file_system.h"
+#include "runtime/descriptors.h"
 #include "runtime/runtime_state.h"
 #include "util/runtime_profile.h"
 #include "vec/data_types/data_type_factory.hpp"
diff --git a/be/test/vec/exec/parquet/parquet_thrift_test.cpp b/be/test/vec/exec/parquet/parquet_thrift_test.cpp
index 3db30dfc2b..498d9cc60d 100644
--- a/be/test/vec/exec/parquet/parquet_thrift_test.cpp
+++ b/be/test/vec/exec/parquet/parquet_thrift_test.cpp
@@ -22,6 +22,7 @@
 
 #include <string>
 
+#include "common/object_pool.h"
 #include "exec/schema_scanner.h"
 #include "io/fs/buffered_reader.h"
 #include "io/fs/local_file_system.h"
diff --git a/be/test/vec/exec/vtablet_sink_test.cpp b/be/test/vec/exec/vtablet_sink_test.cpp
index 1182e1cb8e..2700ed3861 100644
--- a/be/test/vec/exec/vtablet_sink_test.cpp
+++ b/be/test/vec/exec/vtablet_sink_test.cpp
@@ -27,6 +27,7 @@
 #include "gen_cpp/internal_service.pb.h"
 #include "runtime/decimalv2_value.h"
 #include "runtime/descriptor_helper.h"
+#include "runtime/descriptors.h"
 #include "runtime/exec_env.h"
 #include "runtime/result_queue_mgr.h"
 #include "runtime/runtime_state.h"
diff --git a/be/test/vec/function/function_test_util.h b/be/test/vec/function/function_test_util.h
index 9f351c3dd1..2903e0fa1c 100644
--- a/be/test/vec/function/function_test_util.h
+++ b/be/test/vec/function/function_test_util.h
@@ -25,15 +25,18 @@
 #include "testutil/any_type.h"
 #include "testutil/function_utils.h"
 #include "udf/udf.h"
+#include "util/jsonb_utils.h"
 #include "vec/columns/column.h"
 #include "vec/columns/column_const.h"
 #include "vec/core/columns_with_type_and_name.h"
+#include "vec/data_types/data_type_date.h"
 #include "vec/data_types/data_type_date_time.h"
 #include "vec/data_types/data_type_decimal.h"
 #include "vec/data_types/data_type_jsonb.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/data_types/data_type_string.h"
 #include "vec/data_types/data_type_time.h"
+#include "vec/data_types/data_type_time_v2.h"
 #include "vec/exprs/table_function/table_function.h"
 #include "vec/functions/simple_function_factory.h"
 
diff --git a/be/test/vec/jsonb/serialize_test.cpp b/be/test/vec/jsonb/serialize_test.cpp
index c6f17fa9f0..b39523248b 100644
--- a/be/test/vec/jsonb/serialize_test.cpp
+++ b/be/test/vec/jsonb/serialize_test.cpp
@@ -16,10 +16,12 @@
 // under the License.
 #include "vec/jsonb/serialize.h"
 
+#include <gen_cpp/Descriptors_types.h>
 #include <gtest/gtest.h>
 
 #include "gen_cpp/descriptors.pb.h"
 #include "olap/tablet_schema.h"
+#include "runtime/descriptors.h"
 #include "vec/columns/column_array.h"
 #include "vec/columns/column_decimal.h"
 #include "vec/columns/column_nullable.h"
@@ -34,6 +36,7 @@
 #include "vec/data_types/data_type_nullable.h"
 #include "vec/data_types/data_type_number.h"
 #include "vec/data_types/data_type_string.h"
+#include "vec/data_types/data_type_time_v2.h"
 #include "vec/runtime/vdatetime_value.h"
 
 namespace doris::vectorized {
diff --git a/be/test/vec/olap/vertical_compaction_test.cpp b/be/test/vec/olap/vertical_compaction_test.cpp
index a2268b3f1a..a5e112948d 100644
--- a/be/test/vec/olap/vertical_compaction_test.cpp
+++ b/be/test/vec/olap/vertical_compaction_test.cpp
@@ -22,6 +22,7 @@
 
 #include "io/fs/local_file_system.h"
 #include "olap/merger.h"
+#include "olap/rowid_conversion.h"
 #include "olap/rowset/beta_rowset.h"
 #include "olap/rowset/rowset.h"
 #include "olap/rowset/rowset_factory.h"