Merge pull request #51734 from ClickHouse/backport/23.6/51610

Backport #51610 to 23.6: Fix for moving 'IN' conditions to PREWHERE

src/Processors/QueryPlan/Optimizations/Optimizations.h

@@ -14,6 +14,9 @@ namespace QueryPlanOptimizations
 void optimizeTreeFirstPass(const QueryPlanOptimizationSettings & settings, QueryPlan::Node & root, QueryPlan::Nodes & nodes);
 /// Second pass is used to apply read-in-order and attach a predicate to PK.
 void optimizeTreeSecondPass(const QueryPlanOptimizationSettings & optimization_settings, QueryPlan::Node & root, QueryPlan::Nodes & nodes);
+/// Third pass is used to apply filters such as key conditions and skip indexes to the storages that support them.
+/// After that it add CreateSetsStep for the subqueries that has not be used in the filters.
+void optimizeTreeThirdPass(QueryPlan::Node & root, QueryPlan::Nodes & nodes);
 
 /// Optimization (first pass) is a function applied to QueryPlan::Node.
 /// It can read and update subtree of specified node.

src/Processors/QueryPlan/Optimizations/optimizePrimaryKeyCondition.cpp

@@ -28,8 +28,6 @@ void optimizePrimaryKeyCondition(const Stack & stack)
         else
             break;
     }
-
-    source_step_with_filter->onAddFilterFinish();
 }
 
 }

src/Processors/QueryPlan/Optimizations/optimizeTree.cpp

@@ -167,7 +167,6 @@ void optimizeTreeSecondPass(const QueryPlanOptimizationSettings & optimization_s
         optimizePrewhere(stack, nodes);
         optimizePrimaryKeyCondition(stack);
         enableMemoryBoundMerging(*stack.back().node, nodes);
-        addPlansForSets(*stack.back().node, nodes);
 
         stack.pop_back();
     }
@@ -178,5 +177,35 @@ void optimizeTreeSecondPass(const QueryPlanOptimizationSettings & optimization_s
             "No projection is used when optimize_use_projections = 1 and force_optimize_projection = 1");
 }
 
+void optimizeTreeThirdPass(QueryPlan::Node & root, QueryPlan::Nodes & nodes)
+{
+    Stack stack;
+    stack.push_back({.node = &root});
+
+    while (!stack.empty())
+    {
+        /// NOTE: frame cannot be safely used after stack was modified.
+        auto & frame = stack.back();
+
+        /// Traverse all children first.
+        if (frame.next_child < frame.node->children.size())
+        {
+            auto next_frame = Frame{.node = frame.node->children[frame.next_child]};
+            ++frame.next_child;
+            stack.push_back(next_frame);
+            continue;
+        }
+
+        if (auto * source_step_with_filter = dynamic_cast<SourceStepWithFilter *>(frame.node->step.get()))
+        {
+            source_step_with_filter->applyFilters();
+        }
+
+        addPlansForSets(*frame.node, nodes);
+
+        stack.pop_back();
+    }
+}
+
 }
 }

src/Processors/QueryPlan/QueryPlan.cpp

@@ -482,6 +482,7 @@ void QueryPlan::optimize(const QueryPlanOptimizationSettings & optimization_sett
 
     QueryPlanOptimizations::optimizeTreeFirstPass(optimization_settings, *root, nodes);
     QueryPlanOptimizations::optimizeTreeSecondPass(optimization_settings, *root, nodes);
+    QueryPlanOptimizations::optimizeTreeThirdPass(*root, nodes);
 
     updateDataStreams(*root);
 }

src/Processors/QueryPlan/ReadFromMergeTree.cpp

@@ -1299,13 +1299,10 @@ static void buildIndexes(
     indexes->skip_indexes = std::move(skip_indexes);
 }
 
-void ReadFromMergeTree::onAddFilterFinish()
+void ReadFromMergeTree::applyFilters()
 {
-    if (!filter_nodes.nodes.empty())
-    {
-        auto filter_actions_dag = buildFilterDAG(context, prewhere_info, filter_nodes, query_info);
-        buildIndexes(indexes, filter_actions_dag, data, context, query_info, metadata_for_reading);
-    }
+    auto filter_actions_dag = buildFilterDAG(context, prewhere_info, filter_nodes, query_info);
+    buildIndexes(indexes, filter_actions_dag, data, context, query_info, metadata_for_reading);
 }
 
 MergeTreeDataSelectAnalysisResultPtr ReadFromMergeTree::selectRangesToRead(

src/Processors/QueryPlan/ReadFromMergeTree.h

@@ -226,7 +226,7 @@ class ReadFromMergeTree final : public SourceStepWithFilter
     size_t getNumStreams() const { return requested_num_streams; }
     bool isParallelReadingEnabled() const { return read_task_callback != std::nullopt; }
 
-    void onAddFilterFinish() override;
+    void applyFilters() override;
 
 private:
     static MergeTreeDataSelectAnalysisResultPtr selectRangesToReadImpl(

src/Processors/QueryPlan/SourceStepWithFilter.h

@@ -37,7 +37,8 @@ class SourceStepWithFilter : public ISourceStep
         filter_dags.push_back(std::move(filter_dag));
     }
 
-    virtual void onAddFilterFinish() {}
+    /// Apply filters that can optimize reading from storage.
+    virtual void applyFilters() {}
 
 protected:
     std::vector<ActionsDAGPtr> filter_dags;

src/Storages/MergeTree/MergeTreeWhereOptimizer.cpp

@@ -110,6 +110,9 @@ static void collectColumns(const RPNBuilderTreeNode & node, const NameSet & colu
     if (node.isConstant())
         return;
 
+    if (node.isSubqueryOrSet())
+        return;
+
     if (!node.isFunction())
     {
         auto column_name = node.getColumnName();

src/Storages/MergeTree/RPNBuilder.cpp

@@ -181,6 +181,21 @@ bool RPNBuilderTreeNode::isConstant() const
     }
 }
 
+bool RPNBuilderTreeNode::isSubqueryOrSet() const
+{
+    if (ast_node)
+    {
+        return
+            typeid_cast<const ASTSubquery *>(ast_node) ||
+            typeid_cast<const ASTTableIdentifier *>(ast_node);
+    }
+    else
+    {
+        const auto * node_without_alias = getNodeWithoutAlias(dag_node);
+        return node_without_alias->result_type->getTypeId() == TypeIndex::Set;
+    }
+}
+
 ColumnWithTypeAndName RPNBuilderTreeNode::getConstantColumn() const
 {
     if (!isConstant())

src/Storages/MergeTree/RPNBuilder.h

@@ -98,6 +98,8 @@ class RPNBuilderTreeNode
     /// Is node constant
     bool isConstant() const;
 
+    bool isSubqueryOrSet() const;
+
     /** Get constant as constant column.
       * Node must be constant before calling these method, otherwise logical exception is thrown.
       */

src/Storages/System/StorageSystemZooKeeper.cpp

@@ -184,7 +184,7 @@ class ReadFromSystemZooKeeper final : public SourceStepWithFilter
 
     void initializePipeline(QueryPipelineBuilder & pipeline, const BuildQueryPipelineSettings & settings) override;
 
-    void onAddFilterFinish() override;
+    void applyFilters() override;
 
 private:
     void fillData(MutableColumns & res_columns);
@@ -421,7 +421,7 @@ static Paths extractPath(const ActionsDAG::NodeRawConstPtrs & filter_nodes, Cont
 }
 
 
-void ReadFromSystemZooKeeper::onAddFilterFinish()
+void ReadFromSystemZooKeeper::applyFilters()
 {
     paths = extractPath(getFilterNodes().nodes, context, context->getSettingsRef().allow_unrestricted_reads_from_keeper);
 }

tests/queries/0_stateless/02809_prewhere_and_in.reference

@@ -0,0 +1,8 @@
+PREWHERE a IN
+PREWHERE a IN
+PREWHERE a IN
+PREWHERE a IN
+PREWHERE b NOT IN
+PREWHERE b NOT IN
+PREWHERE b NOT IN
+PREWHERE b NOT IN

tests/queries/0_stateless/02809_prewhere_and_in.sql

@@ -0,0 +1,56 @@
+DROP TABLE IF EXISTS t_02809;
+
+CREATE TABLE t_02809(a Int64, b Int64, s String)
+ENGINE=MergeTree order by tuple()
+AS SELECT number, number%10, toString(arrayMap(i-> cityHash64(i*number), range(50))) FROM numbers(10000);
+
+CREATE TABLE t_02809_set(c Int64)
+ENGINE=Set()
+AS SELECT * FROM numbers(10);
+
+CREATE TABLE t_02809_aux(c Int64)
+ENGINE=Memory()
+AS SELECT * FROM numbers(10);
+
+
+SET optimize_move_to_prewhere=1;
+
+-- Queries with 'IN'
+SELECT substring(explain, 1, 13) FROM (EXPLAIN SYNTAX
+     SELECT * FROM t_02809 WHERE a IN (SELECT * FROM system.one)
+) WHERE explain LIKE '%WHERE%';
+
+SELECT substring(explain, 1, 13) FROM (EXPLAIN SYNTAX
+     SELECT * FROM t_02809 WHERE a IN (1,2,3)
+) WHERE explain LIKE '%WHERE%';
+
+SELECT substring(explain, 1, 13) FROM (EXPLAIN SYNTAX
+     SELECT * FROM t_02809 WHERE a IN t_02809_set
+) WHERE explain LIKE '%WHERE%';
+
+SELECT substring(explain, 1, 13) FROM (EXPLAIN SYNTAX
+     SELECT * FROM t_02809 WHERE a IN t_02809_aux
+) WHERE explain LIKE '%WHERE%';
+
+
+-- Queries with 'NOT IN'
+SELECT substring(explain, 1, 17) FROM (EXPLAIN SYNTAX
+     SELECT * FROM t_02809 WHERE b NOT IN (SELECT * FROM system.one)
+) WHERE explain LIKE '%WHERE%';
+
+SELECT substring(explain, 1, 17) FROM (EXPLAIN SYNTAX
+     SELECT * FROM t_02809 WHERE b NOT IN (1,2,3)
+) WHERE explain LIKE '%WHERE%';
+
+SELECT substring(explain, 1, 17) FROM (EXPLAIN SYNTAX
+     SELECT * FROM t_02809 WHERE b NOT IN t_02809_set
+) WHERE explain LIKE '%WHERE%';
+
+SELECT substring(explain, 1, 17) FROM (EXPLAIN SYNTAX
+     SELECT * FROM t_02809 WHERE b NOT IN t_02809_aux
+) WHERE explain LIKE '%WHERE%';
+
+
+DROP TABLE t_02809;
+DROP TABLE t_02809_set;
+DROP TABLE t_02809_aux;