Backport #97493 to 25.12: Make convertToFullIfNeeded recursive for compound column types

robot-clickhouse · robot-clickhouse · commit 2d5892af6388 · 2026-02-22T16:17:27.000Z
diff --git a/src/Columns/IColumn.h b/src/Columns/IColumn.h
@@ -130,7 +130,32 @@ class IColumn : public COW<IColumn>
 
     [[nodiscard]] virtual Ptr convertToFullIfNeeded() const
     {
-        return convertToFullColumnIfConst()->convertToFullColumnIfReplicated()->convertToFullColumnIfSparse()->convertToFullColumnIfLowCardinality();
+        Ptr converted = convertToFullColumnIfConst()
+            ->convertToFullColumnIfReplicated()
+            ->convertToFullColumnIfSparse()
+            ->convertToFullColumnIfLowCardinality();
+
+        Columns new_subcolumns;
+        bool any_changed = false;
+
+        converted->forEachSubcolumn([&](const WrappedPtr & subcolumn)
+        {
+            auto new_sub = subcolumn->convertToFullIfNeeded();
+            any_changed |= (new_sub.get() != subcolumn.get());
+            new_subcolumns.push_back(std::move(new_sub));
+        });
+
+        if (!any_changed)
+            return converted;
+
+        auto mutable_column = IColumn::mutate(std::move(converted));
+        size_t i = 0;
+        mutable_column->forEachMutableSubcolumn([&](WrappedPtr & subcolumn)
+        {
+            subcolumn = std::move(new_subcolumns[i++]);
+        });
+
+        return std::move(mutable_column);
     }
 
     /// Creates empty column with the same type.
diff --git a/tests/queries/0_stateless/03926_convert_to_full_recursive_tuple_sparse.reference b/tests/queries/0_stateless/03926_convert_to_full_recursive_tuple_sparse.reference
@@ -0,0 +1 @@
+200
diff --git a/tests/queries/0_stateless/03926_convert_to_full_recursive_tuple_sparse.sql b/tests/queries/0_stateless/03926_convert_to_full_recursive_tuple_sparse.sql
@@ -0,0 +1,40 @@
+-- Tags: no-random-merge-tree-settings
+
+-- Reproduces assertion failure in Set::appendSetElements when ColumnTuple has
+-- inner sparse columns from one MergeTree part and non-sparse from another.
+-- The fix: IColumn::convertToFullIfNeeded now recursively converts subcolumns.
+--
+-- The bug path: Set::appendSetElements is called only when fill_set_elements is
+-- true, which happens when KeyCondition calls buildOrderedSetInplace for index
+-- evaluation. So the IN column must be part of the ORDER BY key.
+--
+-- The assertion fires when the non-sparse chunk is read first (into set_elements)
+-- and the sparse chunk is inserted second, because ColumnVector::insertRangeFrom
+-- asserts typeid equality with the source.
+
+SET optimize_on_insert = 0;
+
+DROP TABLE IF EXISTS t_sparse_tuple;
+
+-- ORDER BY val is essential: it makes KeyCondition call buildOrderedSetInplace
+-- for the IN expression, which triggers fillSetElements and appendSetElements.
+CREATE TABLE t_sparse_tuple (key UInt64, val Tuple(a UInt64, b UInt64))
+ENGINE = MergeTree ORDER BY val
+SETTINGS ratio_of_defaults_for_sparse_serialization = 0.5;
+
+SYSTEM STOP MERGES t_sparse_tuple;
+
+-- Part 1 (all_1_1_0): non-sparse inner columns (no defaults in second element).
+-- This part is read first into set_elements, so set_elements uses ColumnVector.
+INSERT INTO t_sparse_tuple SELECT number, (number + 1, number + 1) FROM numbers(100);
+
+-- Part 2 (all_2_2_0): second tuple element is all zeros → sparse serialization.
+-- When this chunk is inserted into the Set, insertRangeFrom sees ColumnSparse
+-- source vs ColumnVector destination, triggering the assertion.
+INSERT INTO t_sparse_tuple SELECT number + 200, (number + 200, 0) FROM numbers(100);
+
+-- Building a Set from a subquery that reads both parts triggers the bug.
+-- val must be in ORDER BY for KeyCondition to use buildOrderedSetInplace.
+SELECT count() FROM t_sparse_tuple WHERE val IN (SELECT val FROM t_sparse_tuple);
+
+DROP TABLE t_sparse_tuple;
