[BugFix] Broadcast Join should not generate nondetermistic GRF (#44111)

Signed-off-by: satanson <[email protected]> (cherry picked from commit ecbc790) # Conflicts: # be/src/exec/exec_node.h # be/src/exec/vectorized/hash_join_node.cpp # fe/fe-core/src/main/java/com/starrocks/planner/JoinNode.java # fe/fe-core/src/main/java/com/starrocks/planner/PlanFragment.java # fe/fe-core/src/main/java/com/starrocks/planner/RuntimeFilterDescription.java # fe/fe-core/src/main/java/com/starrocks/planner/RuntimeFilterPushDownContext.java
StarRocks · May 7, 2024 · c7f3558 · c7f3558
1 parent 795be77
commit c7f3558
Show file tree

Hide file tree

Showing 22 changed files with 1,743 additions and 8 deletions.
diff --git a/be/src/exec/exec_node.cpp b/be/src/exec/exec_node.cpp
@@ -133,7 +133,7 @@ void ExecNode::push_down_join_runtime_filter(RuntimeState* state, vectorized::Ru
     if (_type != TPlanNodeType::AGGREGATION_NODE && _type != TPlanNodeType::ANALYTIC_EVAL_NODE) {
         push_down_join_runtime_filter_to_children(state, collector);
     }
-    _runtime_filter_collector.push_down(collector, _tuple_ids, _local_rf_waiting_set);
+    _runtime_filter_collector.push_down(state, id(), collector, _tuple_ids, _local_rf_waiting_set);
 }
 
 void ExecNode::push_down_join_runtime_filter_to_children(RuntimeState* state,

diff --git a/be/src/exec/exec_node.h b/be/src/exec/exec_node.h
@@ -233,6 +233,18 @@ class ExecNode {
     // Names of counters shared by all exec nodes
     static const std::string ROW_THROUGHPUT_COUNTER;
 
+<<<<<<< HEAD
+=======
+    static void may_add_chunk_accumulate_operator(OpFactories& ops, pipeline::PipelineBuilderContext* context, int id);
+
+    void set_children(std::vector<ExecNode*>&& children) { _children = std::move(children); }
+
+    const std::vector<ExecNode*>& children() const { return _children; }
+
+    [[nodiscard]] static Status create_vectorized_node(RuntimeState* state, ObjectPool* pool, const TPlanNode& tnode,
+                                                       const DescriptorTbl& descs, ExecNode** node);
+
+>>>>>>> ecbc7907bb ([BugFix] Broadcast Join should not generate nondetermistic GRF (#44111))
 protected:
     friend class DataSink;
 

diff --git a/be/src/exec/pipeline/fragment_executor.cpp b/be/src/exec/pipeline/fragment_executor.cpp
@@ -287,6 +287,41 @@ int FragmentExecutor::_calc_query_expired_seconds(const UnifiedExecPlanFragmentP
     return QueryContext::DEFAULT_EXPIRE_SECONDS;
 }
 
+static void collect_shuffle_hash_bucket_rf_ids(const ExecNode* node, std::unordered_set<int32_t>& filter_ids) {
+    for (const auto* child : node->children()) {
+        collect_shuffle_hash_bucket_rf_ids(child, filter_ids);
+    }
+    if (node->type() == TPlanNodeType::HASH_JOIN_NODE) {
+        const auto* join_node = down_cast<const HashJoinNode*>(node);
+        if (join_node->distribution_mode() == TJoinDistributionMode::SHUFFLE_HASH_BUCKET) {
+            for (const auto* rf : join_node->build_runtime_filters()) {
+                filter_ids.insert(rf->filter_id());
+            }
+        }
+    }
+}
+
+static std::unordered_set<int32_t> collect_broadcast_join_right_offsprings(
+        const ExecNode* node, BroadcastJoinRightOffsprings& broadcast_join_right_offsprings) {
+    std::vector<std::unordered_set<int32_t>> offsprings_per_child;
+    std::unordered_set<int32_t> offsprings;
+    offsprings_per_child.reserve(node->children().size());
+    for (const auto* child : node->children()) {
+        auto child_offspring = collect_broadcast_join_right_offsprings(child, broadcast_join_right_offsprings);
+        offsprings.insert(child_offspring.begin(), child_offspring.end());
+        offsprings_per_child.push_back(std::move(child_offspring));
+    }
+    offsprings.insert(node->id());
+    if (node->type() == TPlanNodeType::HASH_JOIN_NODE) {
+        const auto* join_node = down_cast<const HashJoinNode*>(node);
+        if (join_node->distribution_mode() == TJoinDistributionMode::BROADCAST &&
+            join_node->can_generate_global_runtime_filter()) {
+            broadcast_join_right_offsprings.insert(offsprings_per_child[1].begin(), offsprings_per_child[1].end());
+        }
+    }
+    return offsprings;
+}
+
 Status FragmentExecutor::_prepare_exec_plan(ExecEnv* exec_env, const UnifiedExecPlanFragmentParams& request) {
     auto* runtime_state = _fragment_ctx->runtime_state();
     auto* obj_pool = runtime_state->obj_pool();
@@ -307,6 +342,12 @@ Status FragmentExecutor::_prepare_exec_plan(ExecEnv* exec_env, const UnifiedExec
     // Set up plan
     RETURN_IF_ERROR(ExecNode::create_tree(runtime_state, obj_pool, fragment.plan, desc_tbl, &_fragment_ctx->plan()));
     ExecNode* plan = _fragment_ctx->plan();
+    std::unordered_set<int32_t> filter_ids;
+    collect_shuffle_hash_bucket_rf_ids(plan, filter_ids);
+    runtime_state->set_shuffle_hash_bucket_rf_ids(std::move(filter_ids));
+    BroadcastJoinRightOffsprings broadcast_join_right_offsprings_map;
+    collect_broadcast_join_right_offsprings(plan, broadcast_join_right_offsprings_map);
+    runtime_state->set_broadcast_join_right_offsprings(std::move(broadcast_join_right_offsprings_map));
     plan->push_down_join_runtime_filter_recursively(runtime_state);
     std::vector<TupleSlotMapping> empty_mappings;
     plan->push_down_tuple_slot_mappings(runtime_state, empty_mappings);

diff --git a/be/src/exec/vectorized/aggregate/aggregate_base_node.cpp b/be/src/exec/vectorized/aggregate/aggregate_base_node.cpp
@@ -51,7 +51,7 @@ Status AggregateBaseNode::close(RuntimeState* state) {
 void AggregateBaseNode::push_down_join_runtime_filter(RuntimeState* state,
                                                       vectorized::RuntimeFilterProbeCollector* collector) {
     // accept runtime filters from parent if possible.
-    _runtime_filter_collector.push_down(collector, _tuple_ids, _local_rf_waiting_set);
+    _runtime_filter_collector.push_down(state, id(), collector, _tuple_ids, _local_rf_waiting_set);
 
     // check to see if runtime filters can be rewritten
     auto& descriptors = _runtime_filter_collector.descriptors();

diff --git a/be/src/exec/vectorized/hash_join_node.cpp b/be/src/exec/vectorized/hash_join_node.cpp
@@ -942,4 +942,31 @@ Status HashJoinNode::_create_implicit_local_join_runtime_filters(RuntimeState* s
     return Status::OK();
 }
 
+<<<<<<< HEAD:be/src/exec/vectorized/hash_join_node.cpp
 } // namespace starrocks::vectorized
+=======
+bool HashJoinNode::can_generate_global_runtime_filter() const {
+    return std::any_of(_build_runtime_filters.begin(), _build_runtime_filters.end(),
+                       [](const RuntimeFilterBuildDescriptor* rf) { return rf->has_remote_targets(); });
+}
+
+void HashJoinNode::push_down_join_runtime_filter(RuntimeState* state, RuntimeFilterProbeCollector* collector) {
+    if (collector->empty()) return;
+    if (_join_type == TJoinOp::INNER_JOIN || _join_type == TJoinOp::LEFT_SEMI_JOIN ||
+        _join_type == TJoinOp::RIGHT_SEMI_JOIN) {
+        ExecNode::push_down_join_runtime_filter(state, collector);
+        return;
+    }
+    _runtime_filter_collector.push_down(state, id(), collector, _tuple_ids, _local_rf_waiting_set);
+}
+
+TJoinDistributionMode::type HashJoinNode::distribution_mode() const {
+    return _distribution_mode;
+}
+
+const std::list<RuntimeFilterBuildDescriptor*>& HashJoinNode::build_runtime_filters() const {
+    return _build_runtime_filters;
+}
+
+} // namespace starrocks
+>>>>>>> ecbc7907bb ([BugFix] Broadcast Join should not generate nondetermistic GRF (#44111)):be/src/exec/hash_join_node.cpp
diff --git a/be/src/exec/vectorized/hash_join_node.h b/be/src/exec/vectorized/hash_join_node.h
@@ -36,6 +36,10 @@ class HashJoinNode final : public ExecNode {
     Status get_next(RuntimeState* state, ChunkPtr* chunk, bool* eos) override;
     Status close(RuntimeState* state) override;
     pipeline::OpFactories decompose_to_pipeline(pipeline::PipelineBuilderContext* context) override;
+    bool can_generate_global_runtime_filter() const;
+    TJoinDistributionMode::type distribution_mode() const;
+    const std::list<RuntimeFilterBuildDescriptor*>& build_runtime_filters() const;
+    void push_down_join_runtime_filter(RuntimeState* state, RuntimeFilterProbeCollector* collector) override;
 
 private:
     static bool _has_null(const ColumnPtr& column);

diff --git a/be/src/exec/vectorized/project_node.cpp b/be/src/exec/vectorized/project_node.cpp
@@ -237,7 +237,7 @@ void ProjectNode::push_down_tuple_slot_mappings(RuntimeState* state,
 void ProjectNode::push_down_join_runtime_filter(RuntimeState* state,
                                                 vectorized::RuntimeFilterProbeCollector* collector) {
     // accept runtime filters from parent if possible.
-    _runtime_filter_collector.push_down(collector, _tuple_ids, _local_rf_waiting_set);
+    _runtime_filter_collector.push_down(state, id(), collector, _tuple_ids, _local_rf_waiting_set);
 
     // check to see if runtime filters can be rewritten
     auto& descriptors = _runtime_filter_collector.descriptors();

diff --git a/be/src/exprs/vectorized/runtime_filter_bank.cpp b/be/src/exprs/vectorized/runtime_filter_bank.cpp
@@ -514,7 +514,8 @@ void RuntimeFilterProbeCollector::update_selectivity(vectorized::Chunk* chunk,
     }
 }
 
-void RuntimeFilterProbeCollector::push_down(RuntimeFilterProbeCollector* parent, const std::vector<TupleId>& tuple_ids,
+void RuntimeFilterProbeCollector::push_down(const RuntimeState* state, TPlanNodeId target_plan_node_id,
+                                            RuntimeFilterProbeCollector* parent, const std::vector<TupleId>& tuple_ids,
                                             std::set<TPlanNodeId>& local_rf_waiting_set) {
     if (this == parent) return;
     auto iter = parent->_descriptors.begin();
@@ -524,7 +525,8 @@ void RuntimeFilterProbeCollector::push_down(RuntimeFilterProbeCollector* parent,
             ++iter;
             continue;
         }
-        if (desc->is_bound(tuple_ids)) {
+        if (desc->is_bound(tuple_ids) && !(state->broadcast_join_right_offsprings().contains(target_plan_node_id) &&
+                                           state->shuffle_hash_bucket_rf_ids().contains(desc->filter_id()))) {
             add_descriptor(desc);
             if (desc->is_local()) {
                 local_rf_waiting_set.insert(desc->build_plan_node_id());

diff --git a/be/src/exprs/vectorized/runtime_filter_bank.h b/be/src/exprs/vectorized/runtime_filter_bank.h
@@ -192,8 +192,8 @@ class RuntimeFilterProbeCollector {
     void add_descriptor(RuntimeFilterProbeDescriptor* desc);
     // accept RuntimeFilterCollector from parent node
     // which means parent node to push down runtime filter.
-    void push_down(RuntimeFilterProbeCollector* parent, const std::vector<TupleId>& tuple_ids,
-                   std::set<TPlanNodeId>& rf_waiting_set);
+    void push_down(const RuntimeState* state, TPlanNodeId target_plan_node_id, RuntimeFilterProbeCollector* parent,
+                   const std::vector<TupleId>& tuple_ids, std::set<TPlanNodeId>& rf_waiting_set);
     std::map<int32_t, RuntimeFilterProbeDescriptor*>& descriptors() { return _descriptors; }
     const std::map<int32_t, RuntimeFilterProbeDescriptor*>& descriptors() const { return _descriptors; }
 

diff --git a/be/src/runtime/runtime_filter_worker.cpp b/be/src/runtime/runtime_filter_worker.cpp
@@ -88,6 +88,9 @@ void RuntimeFilterPort::publish_runtime_filters(std::list<vectorized::RuntimeFil
         auto* filter = rf_desc->runtime_filter();
 
         if (filter == nullptr || !rf_desc->has_remote_targets()) continue;
+        // Empty runtime filter generated by broadcast join can not be used as a global runtime, because it
+        // maybe shirt-circuited by empty probe side.
+        if (rf_desc->join_mode() == TRuntimeFilterBuildJoinMode::BORADCAST && filter->size() == 0) continue;
 
         auto directly_send_broadcast_grf = rf_desc->join_mode() == TRuntimeFilterBuildJoinMode::BORADCAST &&
                                            !rf_desc->broadcast_grf_senders().empty();

diff --git a/be/src/runtime/runtime_state.h b/be/src/runtime/runtime_state.h
@@ -60,7 +60,7 @@ class RowDescriptor;
 class RuntimeFilterPort;
 class QueryStatistics;
 class QueryStatisticsRecvr;
-
+using BroadcastJoinRightOffsprings = std::unordered_set<int32_t>;
 namespace pipeline {
 class QueryContext;
 }
@@ -345,6 +345,20 @@ class RuntimeState {
 
     bool use_page_cache();
 
+    void set_shuffle_hash_bucket_rf_ids(std::unordered_set<int32_t>&& filter_ids) {
+        this->_shuffle_hash_bucket_rf_ids = std::move(filter_ids);
+    }
+
+    const std::unordered_set<int32_t>& shuffle_hash_bucket_rf_ids() const { return this->_shuffle_hash_bucket_rf_ids; }
+
+    void set_broadcast_join_right_offsprings(BroadcastJoinRightOffsprings&& broadcast_join_right_offsprings) {
+        this->_broadcast_join_right_offsprings = std::move(broadcast_join_right_offsprings);
+    }
+
+    const BroadcastJoinRightOffsprings& broadcast_join_right_offsprings() const {
+        return this->_broadcast_join_right_offsprings;
+    }
+
 private:
     // Set per-query state.
     void _init(const TUniqueId& fragment_instance_id, const TQueryOptions& query_options,
@@ -464,6 +478,9 @@ class RuntimeState {
     pipeline::FragmentContext* _fragment_ctx = nullptr;
 
     bool _enable_pipeline_engine = false;
+
+    std::unordered_set<int32_t> _shuffle_hash_bucket_rf_ids;
+    BroadcastJoinRightOffsprings _broadcast_join_right_offsprings;
 };
 
 #define LIMIT_EXCEEDED(tracker, state, msg)                                                                         \

diff --git a/fe/fe-core/src/main/java/com/starrocks/analysis/JoinOperator.java b/fe/fe-core/src/main/java/com/starrocks/analysis/JoinOperator.java
@@ -95,6 +95,10 @@ public boolean isLeftAntiJoin() {
         return this == LEFT_ANTI_JOIN || this == NULL_AWARE_LEFT_ANTI_JOIN;
     }
 
+    public boolean isNullAwareLeftAntiJoin() {
+        return this == NULL_AWARE_LEFT_ANTI_JOIN;
+    }
+
     public boolean isRightSemiJoin() {
         return this == RIGHT_SEMI_JOIN;
     }
@@ -140,6 +144,10 @@ public static Set<JoinOperator> semiAntiJoinSet() {
     public static Set<JoinOperator> innerCrossJoinSet() {
         return Sets.newHashSet(INNER_JOIN, CROSS_JOIN);
     }
+
+    public boolean canGenerateRuntimeFilter() {
+        return !(isLeftOuterJoin() || isFullOuterJoin() || isLeftAntiJoin());
+    }
 }
 
 
diff --git a/fe/fe-core/src/main/java/com/starrocks/planner/JoinNode.java b/fe/fe-core/src/main/java/com/starrocks/planner/JoinNode.java
@@ -23,9 +23,15 @@
 
 import com.google.common.base.MoreObjects;
 import com.google.common.base.Preconditions;
+import com.google.common.collect.ImmutableList;
 import com.google.common.collect.Lists;
 import com.starrocks.analysis.Analyzer;
 import com.starrocks.analysis.BinaryPredicate;
+<<<<<<< HEAD
+=======
+import com.starrocks.analysis.BinaryType;
+import com.starrocks.analysis.DescriptorTable;
+>>>>>>> ecbc7907bb ([BugFix] Broadcast Join should not generate nondetermistic GRF (#44111))
 import com.starrocks.analysis.Expr;
 import com.starrocks.analysis.JoinOperator;
 import com.starrocks.analysis.SlotId;
@@ -158,7 +164,12 @@ public List<Expr> getProbePartitionByExprs() {
     }
 
     @Override
+<<<<<<< HEAD
     public void buildRuntimeFilters(IdGenerator<RuntimeFilterId> runtimeFilterIdIdGenerator) {
+=======
+    public void buildRuntimeFilters(IdGenerator<RuntimeFilterId> runtimeFilterIdIdGenerator, DescriptorTable descTbl,
+                                    ExecGroupSets execGroupSets) {
+>>>>>>> ecbc7907bb ([BugFix] Broadcast Join should not generate nondetermistic GRF (#44111))
         SessionVariable sessionVariable = ConnectContext.get().getSessionVariable();
         JoinOperator joinOp = getJoinOp();
         PlanNode inner = getChild(1);
@@ -271,13 +282,79 @@ public boolean pushDownRuntimeFiltersForChild(RuntimeFilterDescription descripti
                 partitionByExprs, candidatesOfSlotExprsForChild(partitionByExprs, childIdx), childIdx, false);
     }
 
+    private Optional<Boolean> pushDownRuntimeFilterBilaterally(RuntimeFilterPushDownContext context,
+                                                                      Expr probeExpr,
+                                                                      List<Expr> partitionByExprs) {
+        if (joinOp.isCrossJoin() || joinOp.isNullAwareLeftAntiJoin() || eqJoinConjuncts.isEmpty()) {
+            return Optional.empty();
+        }
+
+        if (!(probeExpr instanceof SlotRef)) {
+            return Optional.empty();
+        }
+        SlotRef probeSlotRefExpr = probeExpr.cast();
+        int slotId = probeSlotRefExpr.getSlotId().asInt();
+        boolean probeExprIsNotJoinColumn = eqJoinConjuncts.stream()
+                .filter(conj -> conj.getOp().equals(BinaryType.EQ))
+                .noneMatch(conj -> conj.getUsedSlotIds().contains(slotId));
+
+        if (probeExprIsNotJoinColumn) {
+            return Optional.empty();
+        }
+
+        // for join types except null-aware-left-anti-join and cross join
+        // runtime-filer probe expr uses join column, it can always be push down to both side of the join.
+        boolean hasPushedDown = pushDownRuntimeFiltersForChild(context, probeExpr, partitionByExprs, 0);
+        hasPushedDown |= pushDownRuntimeFiltersForChild(context, probeExpr, partitionByExprs, 1);
+        return Optional.of(hasPushedDown);
+    }
+
+
+    private Optional<Boolean> pushDownRuntimeFilterUnilaterally(RuntimeFilterPushDownContext context,
+                                                                Expr probeExpr,
+                                                                List<Expr> partitionByExprs) {
+        List<Integer> sides = ImmutableList.of();
+        if (joinOp.isLeftAntiJoin() || joinOp.isLeftOuterJoin()) {
+            sides = ImmutableList.of(0);
+        } else if (joinOp.isRightAntiJoin() || joinOp.isRightOuterJoin()) {
+            sides = ImmutableList.of(1);
+        } else if (joinOp.isInnerJoin() || joinOp.isSemiJoin() || joinOp.isCrossJoin()) {
+            sides = ImmutableList.of(0, 1);
+        }
+
+        boolean result = false;
+        Optional<List<List<Expr>>> optCandidatePartitionByExprs =
+                canPushDownRuntimeFilterCrossExchange(partitionByExprs);
+        if (optCandidatePartitionByExprs.isEmpty()) {
+            return Optional.of(false);
+        }
+        List<List<Expr>> candidatePartitionByExprs = optCandidatePartitionByExprs.get();
+        for (Integer side : sides) {
+            if (candidatePartitionByExprs.isEmpty()) {
+                result = getChild(side).pushDownRuntimeFilters(context, probeExpr, Lists.newArrayList());
+            } else {
+                for (List<Expr> partByExprs : candidatePartitionByExprs) {
+                    result = getChild(side).pushDownRuntimeFilters(context, probeExpr, partByExprs);
+                    if (result) {
+                        break;
+                    }
+                }
+            }
+            if (result) {
+                break;
+            }
+        }
+        return Optional.of(result);
+    }
+
     @Override
     public boolean pushDownRuntimeFilters(RuntimeFilterDescription description, Expr probeExpr, List<Expr> partitionByExprs) {
         if (!canPushDownRuntimeFilter()) {
             return false;
         }
 
         if (probeExpr.isBoundByTupleIds(getTupleIds())) {
+<<<<<<< HEAD
             boolean hasPushedDown = false;
             // If probeExpr is SlotRef(a) and an equalJoinConjunct SlotRef(a)=SlotRef(b) exists in SemiJoin
             // or InnerJoin, then the rf also can be pushed down to both sides of HashJoin because SlotRef(a) and
@@ -290,6 +367,15 @@ public boolean pushDownRuntimeFilters(RuntimeFilterDescription description, Expr
             // fall back to PlanNode.pushDownRuntimeFilters for HJ if rf cannot be pushed down via equivalent
             // equalJoinConjuncts
             if (hasPushedDown || super.pushDownRuntimeFilters(description, probeExpr, partitionByExprs)) {
+=======
+
+            Optional<Boolean> pushDownResult = pushDownRuntimeFilterBilaterally(context, probeExpr, partitionByExprs);
+            if (pushDownResult.isEmpty()) {
+                pushDownResult = pushDownRuntimeFilterUnilaterally(context, probeExpr, partitionByExprs);
+            }
+
+            if (pushDownResult.isPresent() && pushDownResult.get()) {
+>>>>>>> ecbc7907bb ([BugFix] Broadcast Join should not generate nondetermistic GRF (#44111))
                 return true;
             }