#13195 left some unresolved issues. One of them is that some BE unit tests fail.
This PR fixes this issue. Now, we can run the command ./run-be-ut.sh --run successfully on macOS.
mem tracker can be logically divided into 4 layers: 1)process 2)type 3)query/load/compation task etc. 4)exec node etc.
type includes
enum Type {
GLOBAL = 0, // Life cycle is the same as the process, e.g. Cache and default Orphan
QUERY = 1, // Count the memory consumption of all Query tasks.
LOAD = 2, // Count the memory consumption of all Load tasks.
COMPACTION = 3, // Count the memory consumption of all Base and Cumulative tasks.
SCHEMA_CHANGE = 4, // Count the memory consumption of all SchemaChange tasks.
CLONE = 5, // Count the memory consumption of all EngineCloneTask. Note: Memory that does not contain make/release snapshots.
BATCHLOAD = 6, // Count the memory consumption of all EngineBatchLoadTask.
CONSISTENCY = 7 // Count the memory consumption of all EngineChecksumTask.
}
Object pointers are no longer saved between each layer, and the values of process and each type are periodically aggregated.
other fix:
In [fix](memtracker) Fix transmit_tracker null pointer because phamp is not thread safe #13528, I tried to separate the memory that was manually abandoned in the query from the orphan mem tracker. But in the actual test, the accuracy of this part of the memory cannot be guaranteed, so put it back to the orphan mem tracker again.
Reuse compression ctx and buffer.
Use a global instance for every compression algorithm, and use a
thread saft buffer pool to reuse compression buffer, pool size is equal
to max parallel thread num in compression, and this will not be too large.
Test shows this feature increase 5% of data import and compaction.
Co-authored-by: yixiutt <yixiu@selectdb.com>
* [fix](threadpool) threadpool schedules does not work right on concurrent token
Assuming there is a concurrent thread token whose concurrency is 2, and the 1st
submit on the token is submitted to threadpool while the 2nd is not submitted due
to busy. The token's active_threads is 1, then thread pool does not schedule the
token.
The patch fixes the problem.
Refactor TaggableLogger
Refactor status handling in agent task:
Unify log format in TaskWorkerPool
Pass Status to the top caller, and replace some OLAPInternalError with more detailed error message Status
Premature return with the opposite condition to reduce indention
1. add some metrics for cpu monitor;
2. add metrics for process state monitor;
3. add metrics for memory monitor;
It is convenient for us to use grafana to filter through different conditions.
After the added, we can find the cpu metrics like this：
doris_be_cpu{device="cpu1",mode="guest_nice"} 0
doris_be_cpu{device="cpu1",mode="guest"} 0
doris_be_cpu{device="cpu1",mode="steal"} 0
doris_be_cpu{device="cpu1",mode="soft_irq"} 107168
doris_be_cpu{device="cpu1",mode="irq"} 0
doris_be_cpu{device="cpu1",mode="iowait"} 3726931
doris_be_cpu{device="cpu1",mode="idle"} 2358039214
doris_be_cpu{device="cpu1",mode="system"} 58699464
doris_be_cpu{device="cpu1",mode="nice"} 1700438
doris_be_cpu{device="cpu1",mode="user"} 54974091
we can find the memory metrics as follow：
doris_be_memory_pswpin 167785
doris_be_memory_pswpout 203724
doris_be_memory_pgpgin 22308762092
doris_be_memory_pgpgout 152101956232
we also can find the process metrics as follow:
doris_be_proc{mode="interrupt"} 421721020416
doris_be_proc{mode="ctxt_switch"} 2806640907317
doris_be_proc{mode="procs_running"} 8
doris_be_proc{mode="procs_blocked"} 3
ZSTD compression is fast with high compression ratio. It can be used to archive higher compression ratio
than default Lz4f codec for storing cost sensitive data such as logs.
Compared to Lz4f codec, we see zstd codec get 35% compressed size off, 30% faster at first time read without OS page
cache, 40% slower at second time read with OS page cache in the following comparison test.
test data: 25GB text log, 110 million rows
test table: test_table(ts varchar(30), log string)
test SQL: set enable_vectorized_engine=1; select sum(length(log)) from test_table
be.conf: disable_storage_page_cache = true
set this config to disable doris page cache to avoid all data cached in memory for test real decompression speed.
test result
master branch with lz4f codec result:
- compressed size 4.3G
- SQL first exec time(read data from disk + decompress + little computation) : 18.3s
- SQL second exec time(read data from OS pagecache + decompress + little computation) : 2.4s
this branch with zstd codec (hardcode enable it) result:
- compressed size: 2.8G
- SQL first exec time: 12.8s
- SQL second exec time: 3.4s
## Design:
For now, there are two categories of types in Doris, one is for scalar types (such as int, char and etc.) and the other is for composite types (array and etc.). For the sake of performance, we can cache type info of scalar types globally (unique objects) due to the limited number of scalar types. When we consider the composite types, normally, the type info is generated in runtime (we can also use some cache strategy to speed up). The memory thereby should be reclaimed when we create type info for composite types.
There are a lots of interfaces to get the type info of a specific type. I reorganized those as the following describes.
1. `const TypeInfo* get_scalar_type_info(FieldType field_type)`
The function is used to get the type info of scalar types. Due to the cache, the caller uses the result **WITHOUT** considering the problems about memory reclaim.
2. `const TypeInfo* get_collection_type_info(FieldType sub_type)`
The function is used to get the type info of array types with just **ONE** depth. Due to the cache, the caller uses the result **WITHOUT** considering the problems about memory reclaim.
3. `TypeInfoPtr get_type_info(segment_v2::ColumnMetaPB* column_meta_pb)`
4. `TypeInfoPtr get_type_info(const TabletColumn* col)`
These functions are used to get the type info of **BOTH** scalar types and composite types. The caller should be responsible to manage the resources returned.
#### About the new type `TypeInfoPtr`
`TypeInfoPtr` is an alias type to `unique_ptr` with a custom deleter.
1. For scalar types, the deleter does nothing.
2. For composite types, the deleter reclaim the memory.
By analyzing the callers of `get_type_info`, these classes should hold TypeInfoPtr:
1. `Field`
2. `ColumnReader`
3. `DefaultValueColumnIterator`
Other classes are either constructed by the foregoing classes or hold those, so they can just use the raw pointer of `TypeInfo` directly for the sake of performance.
1. `ScalarColumnWriter` - holds `Field`
1. `ZoneMapIndexWriter` - created by `ScalarColumnWriter`, use `type_info` from the field in `ScalarColumnWriter`
1. `IndexedColumnWriter` - created by `ZoneMapIndexWriter`, only uses scalar types.
2. `BitmapIndexWriter` - created by `ScalarColumnWriter`, uses `type_info` from the field in `ScalarColumnWriter`
1. `IndexedColumnWriter` - created by `BitmapIndexWriter`, uses `type_info` in `BitmapIndexWriter` and `BitmapIndexWriter` doesn't support `ArrayType`.
3. `BloomFilterIndexWriter` - created by `ScalarColumnWriter`, uses `type_info` from the field in `ScalarColumnWriter`
1. `IndexedColumnWriter` - created by `BloomFilterIndexWriter`, only uses scalar types.
2. `IndexedColumnReader` initializes `type_info` by the field type in meta (only scalar types).
3. `ColumnVectorBatch`
1. `ZoneMapIndexReader` creates `ColumnVectorBatch`, `ColumnVectorBatch` uses `type_info` in `IndexedColumnReader`
2. `BitmapIndexReader` supports scalar types only and it creates `ColumnVectorBatch`, `ColumnVectorBatch` uses `type_info` in `BitmapIndexReader`
3. `BloomFilterIndexWriter` supports scalar types only and it creates `ColumnVectorBatch`, `ColumnVectorBatch` uses `type_info` in `BloomFilterIndexWriter`
1. solved the previous delayed unit test file size is too large (1.7G+) and the unit test link time is too long problem problems
2. Unify all unit tests into one file to significantly reduce unit test execution time to less than 3 mins
3. temporarily disable stream_load_test.cpp, metrics_action_test.cpp, load_channel_mgr_test.cpp because it will re-implement part of the code and affect other tests
Add a new column-type to speed up the approximation of quantiles.
1. The new column-type is named `quantile_state` with fixed aggregation function `quantile_union`, which stores the intermediate results of pre-aggregated approximation calculations for quantiles.
2. support pre-aggregation of new column-type and quantile_state related functions.
Early Design Documentation: https://shimo.im/docs/DT6JXDRkdTvdyV3G
Implement a new way of memory statistics based on TCMalloc New/Delete Hook,
MemTracker and TLS, and it is expected that all memory new/delete/malloc/free
of the BE process can be counted.
Modify the implementation of MemTracker:
1. Simplify a lot of useless logic;
2. Added MemTrackerTaskPool, as the ancestor of all query and import trackers, This is used to track the local memory usage of all tasks executing;
3. Add cosume/release cache, trigger a cosume/release when the memory accumulation exceeds the parameter mem_tracker_consume_min_size_bytes;
4. Add a new memory leak detection mode (Experimental feature), throw an exception when the remaining statistical value is greater than the specified range when the MemTracker is destructed, and print the accurate statistical value in HTTP, the parameter memory_leak_detection
5. Added Virtual MemTracker, cosume/release will not sync to parent. It will be used when introducing TCMalloc Hook to record memory later, to record the specified memory independently;
6. Modify the GC logic, register the buffer cached in DiskIoMgr as a GC function, and add other GC functions later;
7. Change the global root node from Root MemTracker to Process MemTracker, and remove Process MemTracker in exec_env;
8. Modify the macro that detects whether the memory has reached the upper limit, modify the parameters and default behavior of creating MemTracker, modify the error message format in mem_limit_exceeded, extend and apply transfer_to, remove Metric in MemTracker, etc.;
Modify where MemTracker is used:
1. MemPool adds a constructor to create a temporary tracker to avoid a lot of redundant code;
2. Added trackers for global objects such as ChunkAllocator and StorageEngine;
3. Added more fine-grained trackers such as ExprContext;
4. RuntimeState removes FragmentMemTracker, that is, PlanFragmentExecutor mem_tracker, which was previously used for independent statistical scan process memory, and replaces it with _scanner_mem_tracker in OlapScanNode;
5. MemTracker is no longer recorded in ReservationTracker, and ReservationTracker will be removed later;
* [refactor] remove types_test
1. remove types_test, it will cause core dump in higher version GCC or
clang, because of memory align, some code will be vectorized in higher
GCC or clang
2. Change string type length to 2 GB instead of -1
3. modify inaccessible code
