Merge pull request ClickHouse#88090 from shankar-iyer/fix_is_deleted_with_filter

Optimize ReplacingMergeTree is_deleted FINAL queries by adding a filter expression transform

Author: shankar-iyer

src/Processors/QueryPlan/ReadFromMergeTree.cpp

@@ -1400,6 +1400,18 @@ static std::pair<std::shared_ptr<ExpressionActions>, String> createExpressionFor
     return {std::make_shared<ExpressionActions>(std::move(actions)), sign_filter->getColumnName()};
 }
 
+static std::pair<std::shared_ptr<ExpressionActions>, String> createExpressionForIsDeleted(const String & is_deleted_column_name, const Block & header, const ContextPtr & context)
+{
+    ASTPtr is_deleted_identifier = std::make_shared<ASTIdentifier>(is_deleted_column_name);
+    ASTPtr is_deleted_filter = makeASTFunction("equals", is_deleted_identifier, std::make_shared<ASTLiteral>(Field(static_cast<Int8>(0))));
+
+    const auto & is_deleted_column = header.getByName(is_deleted_column_name);
+
+    auto syntax_result = TreeRewriter(context).analyze(is_deleted_filter, {{is_deleted_column.name, is_deleted_column.type}});
+    auto actions = ExpressionAnalyzer(is_deleted_filter, syntax_result, context).getActionsDAG(false);
+    return {std::make_shared<ExpressionActions>(std::move(actions)), is_deleted_filter->getColumnName()};
+}
+
 bool ReadFromMergeTree::doNotMergePartsAcrossPartitionsFinal() const
 {
     const auto & settings = context->getSettingsRef();
@@ -1493,7 +1505,7 @@ Pipe ReadFromMergeTree::spreadMarkRangesAmongStreamsFinal(
         bool no_merging_final = do_not_merge_across_partitions_select_final &&
             std::distance(parts_to_merge_ranges[range_index], parts_to_merge_ranges[range_index + 1]) == 1 &&
             parts_to_merge_ranges[range_index]->data_part->info.level > 0 &&
-            data.merging_params.is_deleted_column.empty() && !reader_settings.read_in_order;
+            !reader_settings.read_in_order;
 
         if (no_merging_final)
         {
@@ -1532,11 +1544,12 @@ Pipe ReadFromMergeTree::spreadMarkRangesAmongStreamsFinal(
                         info.use_uncompressed_cache);
                 };
 
-                /// Parts of non-zero level still may contain duplicate PK values to merge on FINAL if there's is_deleted column,
-                /// so we have to process all ranges. It would be more optimal to remove this flag and add an extra filtering step.
+                /// Parts of non-zero level still may contain duplicate PK values to merge on FINAL if there's is_deleted column.
+                /// Non-intersecting ranges will just go through extra filter added by createExpressionForIsDeleted() to filter
+                /// deleted rows.
                 bool split_parts_ranges_into_intersecting_and_non_intersecting_final
-                    = settings[Setting::split_parts_ranges_into_intersecting_and_non_intersecting_final]
-                    && data.merging_params.is_deleted_column.empty() && !reader_settings.read_in_order;
+                    = settings[Setting::split_parts_ranges_into_intersecting_and_non_intersecting_final] &&
+                          !reader_settings.read_in_order;
 
                 SplitPartsWithRangesByPrimaryKeyResult split_ranges_result = splitPartsWithRangesByPrimaryKey(
                     storage_snapshot->metadata->getPrimaryKey(),
@@ -1621,6 +1634,21 @@ Pipe ReadFromMergeTree::spreadMarkRangesAmongStreamsFinal(
                 return std::make_shared<FilterTransform>(header, expression, filter_name, true);
             });
         }
+        else if (!data.merging_params.is_deleted_column.empty())
+        {
+            auto columns_with_is_deleted = origin_column_names;
+            if (std::ranges::find(columns_with_is_deleted, data.merging_params.is_deleted_column) == columns_with_is_deleted.end())
+                columns_with_is_deleted.push_back(data.merging_params.is_deleted_column);
+
+            pipe = spreadMarkRangesAmongStreams(
+                std::move(non_intersecting_parts_by_primary_key), index_build_context, num_streams, columns_with_is_deleted);
+            auto [expression, filter_name] = createExpressionForIsDeleted(data.merging_params.is_deleted_column, pipe.getHeader(), context);
+
+            pipe.addSimpleTransform([&](const SharedHeader & header)
+            {
+                return std::make_shared<FilterTransform>(header, expression, filter_name, true);
+            });
+        }
         else
         {
             pipe = spreadMarkRangesAmongStreams(std::move(non_intersecting_parts_by_primary_key), num_streams, origin_column_names);

tests/queries/0_stateless/02490_replacing_merge_tree_is_deleted_column_transform_opt.reference

@@ -0,0 +1,7 @@
+10000
+9950
+0
+10000
+17700
+17700
+17700

tests/queries/0_stateless/02490_replacing_merge_tree_is_deleted_column_transform_opt.sql

@@ -0,0 +1,72 @@
+-- Test for FINAL query on ReplacingMergeTree + is_deleted makes use of optimizations.
+
+DROP TABLE IF EXISTS tab;
+
+CREATE TABLE tab (
+     pkey String,
+     id Int32,
+     v Int32,
+     version UInt64,
+     is_deleted UInt8
+) Engine = ReplacingMergeTree(version,is_deleted)
+PARTITION BY pkey ORDER BY id
+SETTINGS index_granularity=512;
+
+-- insert 10000 rows in partition 'A' and delete half of them and merge the 2 parts
+INSERT INTO tab SELECT 'A', number, number, 1, 0 FROM numbers(10000);
+INSERT INTO tab SELECT 'A', number, number + 1, 2, IF(number % 2 = 0, 0, 1) FROM numbers(10000);
+
+OPTIMIZE TABLE tab SETTINGS mutations_sync = 2;
+
+SYSTEM STOP MERGES tab;
+
+-- insert 10000 rows in partition 'B' and delete half of them, but keep 2 parts
+INSERT INTO tab SELECT 'B', number+1000000, number, 1, 0 FROM numbers(10000);
+INSERT INTO tab SELECT 'B', number+1000000, number + 1, 2, IF(number % 2 = 0, 0, 1) FROM numbers(10000);
+
+SET do_not_merge_across_partitions_select_final=1;
+
+-- verify : 10000 rows expected
+SELECT count()
+FROM tab FINAL;
+
+-- add a filter : 9950 rows expected
+SELECT count()
+FROM tab FINAL
+WHERE id >= 100;
+
+-- only even id's are left - 0 rows expected
+SELECT count()
+FROM tab FINAL
+WHERE (id % 2) = 1; 
+
+-- 10000 rows expected
+SELECT count()
+FROM tab FINAL
+WHERE (id % 2) = 0; 
+
+-- create some more partitions
+INSERT INTO tab SELECT 'C', number+2000000, number, 1, 0 FROM numbers(100);
+
+-- insert and delete some rows to get intersecting/non-intersecting ranges in same partition
+INSERT INTO tab SELECT 'D', number+3000000, number, 1, 0 FROM numbers(10000);
+INSERT INTO tab SELECT 'D', number+3000000, number + 1, 1, IF(number % 2 = 0, 0, 1) FROM numbers(5000);
+
+INSERT INTO tab SELECT 'E', number+4000000, number, 1, 0 FROM numbers(100);
+
+-- Total 10000 (From A & B) + 100 (From C) + 7500 (From D) + 100 (From E) = 17700 rows
+SELECT count()
+FROM tab FINAL
+SETTINGS do_not_merge_across_partitions_select_final=0,split_intersecting_parts_ranges_into_layers_final=0;
+
+SELECT count()
+FROM tab FINAL
+SETTINGS do_not_merge_across_partitions_select_final=1,split_intersecting_parts_ranges_into_layers_final=1;
+
+SYSTEM START MERGES tab;
+OPTIMIZE TABLE tab FINAL SETTINGS mutations_sync = 2;
+
+SELECT count()
+FROM tab FINAL;
+
+DROP TABLE IF EXISTS tab;