diff --git a/be/src/exec/hash_join_node.cpp b/be/src/exec/hash_join_node.cpp
index c31e2ee462..d03dd3e49c 100644
--- a/be/src/exec/hash_join_node.cpp
+++ b/be/src/exec/hash_join_node.cpp
@@ -44,6 +44,9 @@ HashJoinNode::HashJoinNode(
     _match_all_build =
         (_join_op == TJoinOp::RIGHT_OUTER_JOIN || _join_op == TJoinOp::FULL_OUTER_JOIN);
     _is_push_down = tnode.hash_join_node.is_push_down;
+    _build_unique = _join_op == TJoinOp::LEFT_ANTI_JOIN|| _join_op == TJoinOp::RIGHT_ANTI_JOIN
+        || _join_op == TJoinOp::RIGHT_SEMI_JOIN || _join_op == TJoinOp::LEFT_SEMI_JOIN 
+        || _join_op == TJoinOp::NULL_AWARE_LEFT_ANTI_JOIN
 }
 
 HashJoinNode::~HashJoinNode() {
diff --git a/be/src/exec/hash_join_node.h b/be/src/exec/hash_join_node.h
index f4b703b7e8..eaa04879cc 100644
--- a/be/src/exec/hash_join_node.h
+++ b/be/src/exec/hash_join_node.h
@@ -88,6 +88,7 @@ private:
     bool _match_all_probe;  // output all rows coming from the probe input
     bool _match_one_build;  // match at most one build row to each probe row
     bool _match_all_build;  // output all rows coming from the build input
+    bool _build_unique;     // build a hash table without duplicated rows
 
     bool _matched_probe;  // if true, we have matched the current probe row
     bool _eos;  // if true, nothing left to return in get_next()
diff --git a/be/src/exec/hash_join_node_ir.cpp b/be/src/exec/hash_join_node_ir.cpp
index e7fec77962..518eb767fa 100644
--- a/be/src/exec/hash_join_node_ir.cpp
+++ b/be/src/exec/hash_join_node_ir.cpp
@@ -30,15 +30,14 @@ namespace doris {
 // we will not be able to replace the funcitons with codegen'd versions.
 // TODO: explicitly set the calling convention?
 // TODO: investigate using fastcc for all codegen internal functions?
-bool IR_NO_INLINE eval_other_join_conjuncts(
-        ExprContext* const* ctxs, int num_ctxs, TupleRow* row) {
+bool IR_NO_INLINE eval_other_join_conjuncts(ExprContext* const* ctxs, int num_ctxs, TupleRow* row) {
     return ExecNode::eval_conjuncts(ctxs, num_ctxs, row);
 }
 
 // CreateOutputRow, EvalOtherJoinConjuncts, and EvalConjuncts are replaced by
 // codegen.
 int HashJoinNode::process_probe_batch(RowBatch* out_batch, RowBatch* probe_batch,
-                                    int max_added_rows) {
+                                      int max_added_rows) {
     // This path does not handle full outer or right outer joins
     DCHECK(!_match_all_build);
 
@@ -63,8 +62,7 @@ int HashJoinNode::process_probe_batch(RowBatch* out_batch, RowBatch* probe_batch
             _hash_tbl_iterator.next<true>();
             create_output_row(out_row, _current_probe_row, matched_build_row);
 
-            if (!eval_other_join_conjuncts(
-                    other_conjunct_ctxs, num_other_conjunct_ctxs, out_row)) {
+            if (!eval_other_join_conjuncts(other_conjunct_ctxs, num_other_conjunct_ctxs, out_row)) {
                 continue;
             }
 
@@ -72,7 +70,7 @@ int HashJoinNode::process_probe_batch(RowBatch* out_batch, RowBatch* probe_batch
 
             // left_anti_join: equal match won't return
             if (_join_op == TJoinOp::LEFT_ANTI_JOIN) {
-                _hash_tbl_iterator= _hash_tbl->end();
+                _hash_tbl_iterator = _hash_tbl->end();
                 break;
             }
 
@@ -95,10 +93,10 @@ int HashJoinNode::process_probe_batch(RowBatch* out_batch, RowBatch* probe_batch
                 break;
             }
         }
-        
+
         // Handle left outer-join and left semi-join
-        if ((!_matched_probe && _match_all_probe) || 
-                ((!_matched_probe && _join_op == TJoinOp::LEFT_ANTI_JOIN))) {
+        if ((!_matched_probe && _match_all_probe) ||
+            ((!_matched_probe && _join_op == TJoinOp::LEFT_ANTI_JOIN))) {
             create_output_row(out_row, _current_probe_row, NULL);
             _matched_probe = true;
 
@@ -137,11 +135,19 @@ end:
     return rows_returned;
 }
 
+// when build table has too many duplicated rows, the collisions will be very serious,
+// so in some case will don't need to store duplicated value in hash table, we can build an unique one
 void HashJoinNode::process_build_batch(RowBatch* build_batch) {
     // insert build row into our hash table
-    for (int i = 0; i < build_batch->num_rows(); ++i) {
-        _hash_tbl->insert(build_batch->get_row(i));
+    if (_build_unique) {
+        for (int i = 0; i < build_batch->num_rows(); ++i) {
+            _hash_tbl->insert_unique(build_batch->get_row(i));
+        }
+    } else {
+        for (int i = 0; i < build_batch->num_rows(); ++i) {
+            _hash_tbl->insert(build_batch->get_row(i));
+        }
     }
 }
-}
 
+} // namespace doris