Try row-by-row decompression if we consider startup cost

.unreleased/row-by-row

@@ -0,0 +1 @@
+Implements: #9732 Speed up some queries with small LIMIT by switching to row-by-row query execution pipeline

tsl/src/nodes/columnar_scan/columnar_scan.c

@@ -1490,6 +1490,10 @@ build_on_single_compressed_path(PlannerInfo *root, const Chunk *chunk, RelOptInf
 	 */
 	if (sort_info->use_compressed_sort)
 	{
+		ColumnarScanPath *columnar_scan_with_compressed_sort = NULL;
+		Path dummy_sort_path; /* dummy for result of cost_sort */
+		Path *compressed_path_for_cost = NULL;
+
 		if (pathkeys_contained_in(sort_info->required_compressed_pathkeys,
 								  compressed_path->pathkeys))
 		{
@@ -1502,6 +1506,9 @@ build_on_single_compressed_path(PlannerInfo *root, const Chunk *chunk, RelOptInf
 			path->needs_sequence_num = sort_info->needs_sequence_num;
 			path->required_compressed_pathkeys = sort_info->required_compressed_pathkeys;
 			path->custom_path.path.pathkeys = sort_info->decompressed_sort_pathkeys;
+
+			columnar_scan_with_compressed_sort = path;
+			compressed_path_for_cost = linitial(path->custom_path.custom_paths);
 		}
 		else
 		{
@@ -1523,9 +1530,8 @@ build_on_single_compressed_path(PlannerInfo *root, const Chunk *chunk, RelOptInf
 			 * creation. Examples of this in: create_merge_append_path &
 			 * create_merge_append_plan
 			 */
-			Path sort_path; /* dummy for result of cost_sort */
 
-			cost_sort(&sort_path,
+			cost_sort(&dummy_sort_path,
 					  root,
 					  sort_info->required_compressed_pathkeys,
 #if PG18_GE
@@ -1538,7 +1544,29 @@ build_on_single_compressed_path(PlannerInfo *root, const Chunk *chunk, RelOptInf
 					  work_mem,
 					  -1);
 
-			cost_columnar_scan(compression_info, path_copy, &sort_path);
+			cost_columnar_scan(compression_info, path_copy, &dummy_sort_path);
+
+			decompressed_paths = lappend(decompressed_paths, path_copy);
+
+			columnar_scan_with_compressed_sort = path_copy;
+			compressed_path_for_cost = &dummy_sort_path;
+		}
+
+		if (chunk_rel->consider_startup &&
+			columnar_scan_with_compressed_sort->enable_bulk_decompression)
+		{
+			/*
+			 * Try a version with row-by-row decompression too, if the planner
+			 * requests the paths with cheap startup. Typically it happens with
+			 * ORDER BY + LIMIT. Row-by-row decompression is only useful if
+			 * there is no sort above the columnar scan, because a sort would
+			 * require a full decompression anyway.
+			 */
+			ColumnarScanPath *path_copy =
+				copy_columnar_scan_path(columnar_scan_with_compressed_sort);
+			path_copy->enable_bulk_decompression = false;
+
+			cost_columnar_scan(compression_info, path_copy, compressed_path_for_cost);
 
 			decompressed_paths = lappend(decompressed_paths, path_copy);
 		}

tsl/test/expected/columnar_index_scan-17.out

@@ -1700,10 +1700,9 @@ SELECT device, max(time), min(time) FROM metrics GROUP BY device ORDER BY 1,2,3;
  Aggregate
    ->  Merge Append
          Sort Key: metrics.device
-         ->  Custom Scan (ColumnarScan) on _hyper_1_1_chunk
-               ->  Sort
-                     Sort Key: compress_hyper_2_2_chunk.device
-                     ->  Seq Scan on compress_hyper_2_2_chunk
+         ->  Custom Scan (SkipScan) on _hyper_1_1_chunk
+               ->  Custom Scan (ColumnarScan) on _hyper_1_1_chunk
+                     ->  Index Scan using compress_hyper_2_2_chunk_device_sensor__ts_meta_v2_last_tim_idx on compress_hyper_2_2_chunk
          ->  Sort
                Sort Key: _hyper_1_1_chunk.device
                ->  Seq Scan on _hyper_1_1_chunk
@@ -1822,13 +1821,14 @@ SELECT device, max(time), min(time) FROM metrics GROUP BY device ORDER BY 1,2,3;
 -- DISTINCT on a segmentby column (Agg with no Aggrefs)
 :PREFIX SELECT DISTINCT device FROM metrics ORDER BY device;
 --- QUERY PLAN ---
- Sort
-   Sort Key: metrics.device
-   ->  HashAggregate
-         Group Key: metrics.device
-         ->  Append
+ Unique
+   ->  Merge Append
+         Sort Key: metrics.device
+         ->  Custom Scan (SkipScan) on _hyper_1_1_chunk
                ->  Custom Scan (ColumnarScan) on _hyper_1_1_chunk
-                     ->  Seq Scan on compress_hyper_2_2_chunk
+                     ->  Index Scan using compress_hyper_2_2_chunk_device_sensor__ts_meta_v2_last_tim_idx on compress_hyper_2_2_chunk
+         ->  Sort
+               Sort Key: _hyper_1_1_chunk.device
                ->  Seq Scan on _hyper_1_1_chunk
 
 -- count(*) without GROUP BY

tsl/test/expected/columnar_index_scan-18.out

@@ -1702,10 +1702,9 @@ SELECT device, max(time), min(time) FROM metrics GROUP BY device ORDER BY 1,2,3;
  Aggregate
    ->  Merge Append
          Sort Key: metrics.device
-         ->  Custom Scan (ColumnarScan) on _hyper_1_1_chunk
-               ->  Sort
-                     Sort Key: compress_hyper_2_2_chunk.device
-                     ->  Seq Scan on compress_hyper_2_2_chunk
+         ->  Custom Scan (SkipScan) on _hyper_1_1_chunk
+               ->  Custom Scan (ColumnarScan) on _hyper_1_1_chunk
+                     ->  Index Scan using compress_hyper_2_2_chunk_device_sensor__ts_meta_v2_last_tim_idx on compress_hyper_2_2_chunk
          ->  Sort
                Sort Key: _hyper_1_1_chunk.device
                ->  Seq Scan on _hyper_1_1_chunk
@@ -1826,13 +1825,14 @@ SELECT device, max(time), min(time) FROM metrics GROUP BY device ORDER BY 1,2,3;
 -- DISTINCT on a segmentby column (Agg with no Aggrefs)
 :PREFIX SELECT DISTINCT device FROM metrics ORDER BY device;
 --- QUERY PLAN ---
- Sort
-   Sort Key: metrics.device
-   ->  HashAggregate
-         Group Key: metrics.device
-         ->  Append
+ Unique
+   ->  Merge Append
+         Sort Key: metrics.device
+         ->  Custom Scan (SkipScan) on _hyper_1_1_chunk
                ->  Custom Scan (ColumnarScan) on _hyper_1_1_chunk
-                     ->  Seq Scan on compress_hyper_2_2_chunk
+                     ->  Index Scan using compress_hyper_2_2_chunk_device_sensor__ts_meta_v2_last_tim_idx on compress_hyper_2_2_chunk
+         ->  Sort
+               Sort Key: _hyper_1_1_chunk.device
                ->  Seq Scan on _hyper_1_1_chunk
 
 -- count(*) without GROUP BY

tsl/test/expected/compress_sort_transform.out

@@ -44,9 +44,10 @@ group by 1, 2 order by 1, 2 limit 1;
          Group Key: ht_metrics_partially_compressed.device, (time_bucket('@ 1 min'::interval, ht_metrics_partially_compressed."time"))
          ->  Merge Append (actual rows=3.00 loops=1)
                Sort Key: ht_metrics_partially_compressed.device, (time_bucket('@ 1 min'::interval, ht_metrics_partially_compressed."time"))
-               ->  Custom Scan (VectorAgg) (actual rows=2.00 loops=1)
-                     ->  Custom Scan (ColumnarScan) on _hyper_1_1_chunk (actual rows=1440.00 loops=1)
-                           ->  Sort (actual rows=3.00 loops=1)
+               ->  Partial GroupAggregate (actual rows=2.00 loops=1)
+                     Group Key: _hyper_1_1_chunk.device, time_bucket('@ 1 min'::interval, _hyper_1_1_chunk."time")
+                     ->  Custom Scan (ColumnarScan) on _hyper_1_1_chunk (actual rows=3.00 loops=1)
+                           ->  Sort (actual rows=1.00 loops=1)
                                  Sort Key: compress_hyper_2_4_chunk.device, compress_hyper_2_4_chunk._ts_meta_v2_first_time, compress_hyper_2_4_chunk._ts_meta_v2_last_time
                                  Sort Method: quicksort 
                                  ->  Seq Scan on compress_hyper_2_4_chunk (actual rows=3.00 loops=1)
@@ -56,9 +57,10 @@ group by 1, 2 order by 1, 2 limit 1;
                            Sort Key: _hyper_1_1_chunk.device, (time_bucket('@ 1 min'::interval, _hyper_1_1_chunk."time"))
                            Sort Method: quicksort 
                            ->  Seq Scan on _hyper_1_1_chunk (actual rows=960.00 loops=1)
-               ->  Custom Scan (VectorAgg) (actual rows=1.00 loops=1)
-                     ->  Custom Scan (ColumnarScan) on _hyper_1_2_chunk (actual rows=1512.00 loops=1)
-                           ->  Sort (actual rows=3.00 loops=1)
+               ->  Partial GroupAggregate (actual rows=1.00 loops=1)
+                     Group Key: _hyper_1_2_chunk.device, time_bucket('@ 1 min'::interval, _hyper_1_2_chunk."time")
+                     ->  Custom Scan (ColumnarScan) on _hyper_1_2_chunk (actual rows=2.00 loops=1)
+                           ->  Sort (actual rows=1.00 loops=1)
                                  Sort Key: compress_hyper_2_5_chunk.device, compress_hyper_2_5_chunk._ts_meta_v2_first_time, compress_hyper_2_5_chunk._ts_meta_v2_last_time
                                  Sort Method: quicksort 
                                  ->  Seq Scan on compress_hyper_2_5_chunk (actual rows=3.00 loops=1)
@@ -68,9 +70,10 @@ group by 1, 2 order by 1, 2 limit 1;
                            Sort Key: _hyper_1_2_chunk.device, (time_bucket('@ 1 min'::interval, _hyper_1_2_chunk."time"))
                            Sort Method: quicksort 
                            ->  Seq Scan on _hyper_1_2_chunk (actual rows=1008.00 loops=1)
-               ->  Custom Scan (VectorAgg) (actual rows=1.00 loops=1)
-                     ->  Custom Scan (ColumnarScan) on _hyper_1_3_chunk (actual rows=507.00 loops=1)
-                           ->  Sort (actual rows=3.00 loops=1)
+               ->  Partial GroupAggregate (actual rows=1.00 loops=1)
+                     Group Key: _hyper_1_3_chunk.device, time_bucket('@ 1 min'::interval, _hyper_1_3_chunk."time")
+                     ->  Custom Scan (ColumnarScan) on _hyper_1_3_chunk (actual rows=2.00 loops=1)
+                           ->  Sort (actual rows=1.00 loops=1)
                                  Sort Key: compress_hyper_2_6_chunk.device, compress_hyper_2_6_chunk._ts_meta_v2_first_time, compress_hyper_2_6_chunk._ts_meta_v2_last_time
                                  Sort Method: quicksort 
                                  ->  Seq Scan on compress_hyper_2_6_chunk (actual rows=3.00 loops=1)

tsl/test/expected/merge_append_partially_compressed.out

@@ -905,10 +905,7 @@ SELECT * FROM test1 WHERE x4 % 11 != 0 ORDER BY x1, x2, x5, time DESC, x3 ASC, x
          ->  Custom Scan (ColumnarScan) on _hyper_3_7_chunk (actual rows=4.00 loops=1)
                Filter: ((x4 % 11) <> 0)
                Rows Removed by Filter: 10000
-               ->  Sort (actual rows=13.00 loops=1)
-                     Sort Key: compress_hyper_4_8_chunk.x1, compress_hyper_4_8_chunk.x2, compress_hyper_4_8_chunk.x5, compress_hyper_4_8_chunk._ts_meta_v2_last_time DESC, compress_hyper_4_8_chunk._ts_meta_v2_first_time DESC, compress_hyper_4_8_chunk._ts_meta_v2_first_x3, compress_hyper_4_8_chunk._ts_meta_v2_last_x3, compress_hyper_4_8_chunk._ts_meta_v2_first_x4, compress_hyper_4_8_chunk._ts_meta_v2_last_x4
-                     Sort Method: quicksort 
-                     ->  Seq Scan on compress_hyper_4_8_chunk (actual rows=13.00 loops=1)
+               ->  Index Scan using compress_hyper_4_8_chunk_x1_x2_x5__ts_meta_v2_last_time__ts_idx on compress_hyper_4_8_chunk (actual rows=13.00 loops=1)
          ->  Sort (actual rows=1.00 loops=1)
                Sort Key: _hyper_3_7_chunk.x1, _hyper_3_7_chunk.x2, _hyper_3_7_chunk.x5, _hyper_3_7_chunk."time" DESC, _hyper_3_7_chunk.x3, _hyper_3_7_chunk.x4
                Sort Method: quicksort