Read chunk row count from catalog tables

src/backend/columnar/cstore_reader.c

@@ -178,18 +178,8 @@ ColumnarReadNextRow(TableReadState *readState, Datum *columnValues, bool *column
 
 	if (chunkIndex != readState->deserializedChunkIndex)
 	{
-		uint32 chunkRowCount = 0;
+		uint32 chunkRowCount =
-
+			readState->stripeBuffers->selectedChunkRowCount[chunkIndex];
-		uint32 stripeRowCount = stripeMetadata->rowCount;
-		uint32 lastChunkIndex = stripeRowCount / stripeMetadata->chunkRowCount;
-		if (chunkIndex == lastChunkIndex)
-		{
-			chunkRowCount = stripeRowCount % stripeMetadata->chunkRowCount;
-		}
-		else
-		{
-			chunkRowCount = stripeMetadata->chunkRowCount;
-		}
 
 		oldContext = MemoryContextSwitchTo(readState->stripeReadContext);
 
@@ -356,6 +346,14 @@ LoadFilteredStripeBuffers(Relation relation, StripeMetadata *stripeMetadata,
 		SelectedChunkSkipList(stripeSkipList, projectedColumnMask,
 							  selectedChunkMask);
 
+	uint32 selectedChunkCount = selectedChunkSkipList->chunkCount;
+	uint32 *selectedChunkRowCount = palloc0(selectedChunkCount * sizeof(uint32));
+	for (int chunkIndex = 0; chunkIndex < selectedChunkCount; chunkIndex++)
+	{
+		selectedChunkRowCount[chunkIndex] =
+			selectedChunkSkipList->chunkSkipNodeArray[0][chunkIndex].rowCount;
+	}
+
 	/* load column data for projected columns */
 	ColumnBuffers **columnBuffersArray = palloc0(columnCount * sizeof(ColumnBuffers *));
 
@@ -381,6 +379,8 @@ LoadFilteredStripeBuffers(Relation relation, StripeMetadata *stripeMetadata,
 	stripeBuffers->columnCount = columnCount;
 	stripeBuffers->rowCount = StripeSkipListRowCount(selectedChunkSkipList);
 	stripeBuffers->columnBuffersArray = columnBuffersArray;
+	stripeBuffers->selectedChunks = selectedChunkCount;
+	stripeBuffers->selectedChunkRowCount = selectedChunkRowCount;
 
 	return stripeBuffers;
 }

src/include/columnar/columnar.h

@@ -201,6 +201,14 @@ typedef struct StripeBuffers
 	uint32 columnCount;
 	uint32 rowCount;
 	ColumnBuffers **columnBuffersArray;
+
+	/*
+	 * We might skip reading some chunks because they're refuted by the
+	 * WHERE clause. We keep number of selected chunks and number of rows
+	 * in each of them.
+	 */
+	uint32 selectedChunks;
+	uint32 *selectedChunkRowCount;
 } StripeBuffers;
 
 

src/test/regress/input/am_chunk_filtering.source

@@ -76,4 +76,12 @@ CREATE TABLE simple_chunk_filtering(i int) USING COLUMNAR;
 INSERT INTO simple_chunk_filtering SELECT generate_series(0,234567);
 EXPLAIN (analyze on, costs off, timing off, summary off)
   SELECT * FROM simple_chunk_filtering WHERE i > 123456;
+
+-- https://github.com/citusdata/citus/issues/4555
+TRUNCATE simple_chunk_filtering;
+INSERT INTO simple_chunk_filtering SELECT generate_series(0,200000);
+COPY (SELECT * FROM simple_chunk_filtering WHERE i > 180000) TO '/dev/null';
+EXPLAIN (analyze on, costs off, timing off, summary off)
+  SELECT * FROM simple_chunk_filtering WHERE i > 180000;
+
 DROP TABLE simple_chunk_filtering;

src/test/regress/output/am_chunk_filtering.source

@@ -130,4 +130,18 @@ EXPLAIN (analyze on, costs off, timing off, summary off)
    Columnar Chunks Removed by Filter: 12
 (4 rows)
 
+-- https://github.com/citusdata/citus/issues/4555
+TRUNCATE simple_chunk_filtering;
+INSERT INTO simple_chunk_filtering SELECT generate_series(0,200000);
+COPY (SELECT * FROM simple_chunk_filtering WHERE i > 180000) TO '/dev/null';
+EXPLAIN (analyze on, costs off, timing off, summary off)
+  SELECT * FROM simple_chunk_filtering WHERE i > 180000;
+                                    QUERY PLAN
+---------------------------------------------------------------------
+ Custom Scan (ColumnarScan) on simple_chunk_filtering (actual rows=20000 loops=1)
+   Filter: (i > 180000)
+   Rows Removed by Filter: 1
+   Columnar Chunks Removed by Filter: 18
+(4 rows)
+
 DROP TABLE simple_chunk_filtering;