Merge pull request #21761 from ClickHouse/aku/chunk-no-columns

fix incorrect number of rows for Chunks with no columns in PartialSor…

src/Interpreters/InterpreterSelectQuery.cpp

@@ -2038,7 +2038,13 @@ void InterpreterSelectQuery::executeWindow(QueryPlan & query_plan)
     for (size_t i = 0; i < windows_sorted.size(); ++i)
     {
         const auto & w = *windows_sorted[i];
-        if (i == 0 || !sortIsPrefix(w, *windows_sorted[i - 1]))
+
+        // We don't need to sort again if the input from previous window already
+        // has suitable sorting. Also don't create sort steps when there are no
+        // columns to sort by, because the sort nodes are confused by this. It
+        // happens in case of `over ()`.
+        if (!w.full_sort_description.empty()
+            && (i == 0 || !sortIsPrefix(w, *windows_sorted[i - 1])))
         {
             auto partial_sorting = std::make_unique<PartialSortingStep>(
                 query_plan.getCurrentDataStream(),

src/Processors/Transforms/PartialSortingTransform.cpp

@@ -10,6 +10,8 @@ PartialSortingTransform::PartialSortingTransform(
     : ISimpleTransform(header_, header_, false)
     , description(description_), limit(limit_)
 {
+    // Sorting by no columns doesn't make sense.
+    assert(!description.empty());
 }
 
 static ColumnRawPtrs extractColumns(const Block & block, const SortDescription & description)
@@ -91,6 +93,14 @@ size_t getFilterMask(const ColumnRawPtrs & lhs, const ColumnRawPtrs & rhs, size_
 
 void PartialSortingTransform::transform(Chunk & chunk)
 {
+    if (chunk.getNumRows())
+    {
+        // The following code works with Blocks and will lose the number of
+        // rows when there are no columns. We shouldn't get such block, because
+        // we have to sort by at least one column.
+        assert(chunk.getNumColumns());
+    }
+
     if (read_rows)
         read_rows->add(chunk.getNumRows());
 

src/Processors/Transforms/WindowTransform.cpp

@@ -881,12 +881,13 @@ void WindowTransform::appendChunk(Chunk & chunk)
         assert(chunk.hasRows());
         blocks.push_back({});
         auto & block = blocks.back();
+        // Use the number of rows from the Chunk, because it is correct even in
+        // the case where the Chunk has no columns. Not sure if this actually
+        // happens, because even in the case of `count() over ()` we have a dummy
+        // input column.
+        block.rows = chunk.getNumRows();
         block.input_columns = chunk.detachColumns();
 
-        // Even in case of `count() over ()` we should have a dummy input column.
-        // Not sure how reliable this is...
-        block.rows = block.input_columns[0]->size();
-
         for (auto & ws : workspaces)
         {
             // Aggregate functions can't work with constant columns, so we have to
@@ -1109,9 +1110,7 @@ IProcessor::Status WindowTransform::prepare()
         if (output.canPush())
         {
             // Output the ready block.
-//            fmt::print(stderr, "output block {}\n", next_output_block_number);
             const auto i = next_output_block_number - first_block_number;
-            ++next_output_block_number;
             auto & block = blocks[i];
             auto columns = block.input_columns;
             for (auto & res : block.output_columns)
@@ -1120,6 +1119,12 @@ IProcessor::Status WindowTransform::prepare()
             }
             output_data.chunk.setColumns(columns, block.rows);
 
+//            fmt::print(stderr, "output block {} as chunk '{}'\n",
+//                next_output_block_number,
+//                output_data.chunk.dumpStructure());
+
+            ++next_output_block_number;
+
             output.pushData(std::move(output_data));
         }
 

tests/queries/0_stateless/01591_window_functions.reference

@@ -993,3 +993,8 @@ order by number
 7	6	8
 8	7	9
 9	8	9
+-- In this case, we had a problem with PartialSortingTransform returning zero-row
+-- chunks for input chunks w/o columns.
+select count() over () from numbers(4) where number < 2;
+2
+2

tests/queries/0_stateless/01591_window_functions.sql

@@ -345,3 +345,7 @@ from numbers(10)
 window w as (order by number range between 1 preceding and 1 following)
 order by number
 ;
+
+-- In this case, we had a problem with PartialSortingTransform returning zero-row
+-- chunks for input chunks w/o columns.
+select count() over () from numbers(4) where number < 2;