Merge pull request #51958 from ClickHouse/prewhere_and_pk_columns

Move conditions with columns from PK to the end of PREWHERE chain

src/Core/Settings.h

@@ -129,6 +129,7 @@ class IColumn;
     M(Bool, optimize_move_to_prewhere_if_final, false, "If query has `FINAL`, the optimization `move_to_prewhere` is not always correct and it is enabled only if both settings `optimize_move_to_prewhere` and `optimize_move_to_prewhere_if_final` are turned on", 0) \
     M(Bool, move_all_conditions_to_prewhere, true, "Move all viable conditions from WHERE to PREWHERE", 0) \
     M(Bool, enable_multiple_prewhere_read_steps, true, "Move more conditions from WHERE to PREWHERE and do reads from disk and filtering in multiple steps if there are multiple conditions combined with AND", 0) \
+    M(Bool, move_primary_key_columns_to_end_of_prewhere, true, "Move PREWHERE conditions containing primary key columns to the end of AND chain. It is likely that these conditions are taken into account during primary key analysis and thus will not contribute a lot to PREWHERE filtering.", 0) \
     \
     M(UInt64, alter_sync, 1, "Wait for actions to manipulate the partitions. 0 - do not wait, 1 - wait for execution only of itself, 2 - wait for everyone.", 0) ALIAS(replication_alter_partitions_sync) \
     M(Int64, replication_wait_for_inactive_replica_timeout, 120, "Wait for inactive replica to execute ALTER/OPTIMIZE. Time in seconds, 0 - do not wait, negative - wait for unlimited time.", 0) \

src/Storages/MergeTree/MergeTreeWhereOptimizer.cpp

@@ -22,6 +22,33 @@ namespace DB
 /// This is used to assume that condition is likely to have good selectivity.
 static constexpr auto threshold = 2;
 
+static NameToIndexMap fillNamesPositions(const Names & names)
+{
+    NameToIndexMap names_positions;
+
+    for (size_t position = 0; position < names.size(); ++position)
+    {
+        const auto & name = names[position];
+        names_positions[name] = position;
+    }
+
+    return names_positions;
+}
+
+/// Find minimal position of any of the column in primary key.
+static Int64 findMinPosition(const NameSet & condition_table_columns, const NameToIndexMap & primary_key_positions)
+{
+    Int64 min_position = std::numeric_limits<Int64>::max() - 1;
+
+    for (const auto & column : condition_table_columns)
+    {
+        auto it = primary_key_positions.find(column);
+        if (it != primary_key_positions.end())
+            min_position = std::min(min_position, static_cast<Int64>(it->second));
+    }
+
+    return min_position;
+}
 
 MergeTreeWhereOptimizer::MergeTreeWhereOptimizer(
     std::unordered_map<std::string, UInt64> column_sizes_,
@@ -35,6 +62,7 @@ MergeTreeWhereOptimizer::MergeTreeWhereOptimizer(
     , supported_columns{supported_columns_}
     , sorting_key_names{NameSet(
           metadata_snapshot->getSortingKey().column_names.begin(), metadata_snapshot->getSortingKey().column_names.end())}
+    , primary_key_names_positions(fillNamesPositions(metadata_snapshot->getPrimaryKey().column_names))
     , log{log_}
     , column_sizes{std::move(column_sizes_)}
 {
@@ -60,6 +88,7 @@ void MergeTreeWhereOptimizer::optimize(SelectQueryInfo & select_query_info, cons
     where_optimizer_context.context = context;
     where_optimizer_context.array_joined_names = determineArrayJoinedNames(select);
     where_optimizer_context.move_all_conditions_to_prewhere = context->getSettingsRef().move_all_conditions_to_prewhere;
+    where_optimizer_context.move_primary_key_columns_to_end_of_prewhere = context->getSettingsRef().move_primary_key_columns_to_end_of_prewhere;
     where_optimizer_context.is_final = select.final();
 
     RPNBuilderTreeContext tree_context(context, std::move(block_with_constants), {} /*prepared_sets*/);
@@ -89,6 +118,7 @@ std::optional<MergeTreeWhereOptimizer::FilterActionsOptimizeResult> MergeTreeWhe
     where_optimizer_context.context = context;
     where_optimizer_context.array_joined_names = {};
     where_optimizer_context.move_all_conditions_to_prewhere = context->getSettingsRef().move_all_conditions_to_prewhere;
+    where_optimizer_context.move_primary_key_columns_to_end_of_prewhere = context->getSettingsRef().move_primary_key_columns_to_end_of_prewhere;
     where_optimizer_context.is_final = is_final;
 
     RPNBuilderTreeContext tree_context(context);
@@ -234,6 +264,14 @@ void MergeTreeWhereOptimizer::analyzeImpl(Conditions & res, const RPNBuilderTree
         if (cond.viable)
             cond.good = isConditionGood(node, table_columns);
 
+        if (where_optimizer_context.move_primary_key_columns_to_end_of_prewhere)
+        {
+            /// Consider all conditions good with this setting enabled.
+            cond.good = cond.viable;
+            /// Find min position in PK of any column that is used in this condition.
+            cond.min_position_in_primary_key = findMinPosition(cond.table_columns, primary_key_names_positions);
+        }
+
         res.emplace_back(std::move(cond));
     }
 }

src/Storages/MergeTree/MergeTreeWhereOptimizer.h

@@ -72,9 +72,14 @@ class MergeTreeWhereOptimizer : private boost::noncopyable
         /// Does the condition presumably have good selectivity?
         bool good = false;
 
+        /// Does the condition contain primary key column?
+        /// If so, it is better to move it further to the end of PREWHERE chain depending on minimal position in PK of any
+        /// column in this condition because this condition have bigger chances to be already satisfied by PK analysis.
+        Int64 min_position_in_primary_key = std::numeric_limits<Int64>::max() - 1;
+
         auto tuple() const
         {
-            return std::make_tuple(!viable, !good, columns_size, table_columns.size());
+            return std::make_tuple(!viable, !good, -min_position_in_primary_key, columns_size, table_columns.size());
         }
 
         /// Is condition a better candidate for moving to PREWHERE?
@@ -91,6 +96,7 @@ class MergeTreeWhereOptimizer : private boost::noncopyable
         ContextPtr context;
         NameSet array_joined_names;
         bool move_all_conditions_to_prewhere = false;
+        bool move_primary_key_columns_to_end_of_prewhere = false;
         bool is_final = false;
     };
 
@@ -141,6 +147,7 @@ class MergeTreeWhereOptimizer : private boost::noncopyable
     const Names queried_columns;
     const std::optional<NameSet> supported_columns;
     const NameSet sorting_key_names;
+    const NameToIndexMap primary_key_names_positions;
     Poco::Logger * log;
     std::unordered_map<std::string, UInt64> column_sizes;
     UInt64 total_size_of_queried_columns = 0;

tests/queries/0_stateless/02156_storage_merge_prewhere.reference

@@ -1,6 +1,6 @@
 SELECT count()
 FROM t_02156_merge1
-PREWHERE (k = 3) AND notEmpty(v)
+PREWHERE notEmpty(v) AND (k = 3)
 2
 SELECT count()
 FROM t_02156_merge2

tests/queries/0_stateless/02156_storage_merge_prewhere.sql

@@ -1,4 +1,5 @@
 SET optimize_move_to_prewhere = 1;
+SET enable_multiple_prewhere_read_steps = 1;
 
 DROP TABLE IF EXISTS t_02156_mt1;
 DROP TABLE IF EXISTS t_02156_mt2;
@@ -8,8 +9,8 @@ DROP TABLE IF EXISTS t_02156_merge1;
 DROP TABLE IF EXISTS t_02156_merge2;
 DROP TABLE IF EXISTS t_02156_merge3;
 
-CREATE TABLE t_02156_mt1 (k UInt32, v String) ENGINE = MergeTree ORDER BY k;
-CREATE TABLE t_02156_mt2 (k UInt32, v String) ENGINE = MergeTree ORDER BY k;
+CREATE TABLE t_02156_mt1 (k UInt32, v String) ENGINE = MergeTree ORDER BY k SETTINGS min_bytes_for_wide_part=0;
+CREATE TABLE t_02156_mt2 (k UInt32, v String) ENGINE = MergeTree ORDER BY k SETTINGS min_bytes_for_wide_part=0;
 CREATE TABLE t_02156_log (k UInt32, v String) ENGINE = Log;
 
 CREATE TABLE t_02156_dist (k UInt32, v String) ENGINE = Distributed(test_shard_localhost, currentDatabase(), t_02156_mt1);