Add a debug GUC to force batch sorted merge plan (#8384)

At the moment a significant part of the tests we have for batch sorted
merge relies on the completely broken cost and size estimations for
plain DecompressChunk. This prevents actually fixing these estimates. At
the same time, it is unfeasible to rewrite these tests entirely. Add a
GUC to force the batch sorted merge plan regardless of the cost, so that
we can test the correctness on small tables.

Author: akuzm

src/guc.c

@@ -144,8 +144,6 @@ TSDLLEXPORT char *ts_guc_license = TS_LICENSE_DEFAULT;
 char *ts_last_tune_time = NULL;
 char *ts_last_tune_version = NULL;
 
-bool ts_guc_debug_require_batch_sorted_merge = false;
-
 bool ts_guc_debug_allow_cagg_with_deprecated_funcs = false;
 
 /*
@@ -169,20 +167,23 @@ char *ts_current_timestamp_mock = NULL;
 
 int ts_guc_debug_toast_tuple_target = 128;
 
-#ifdef TS_DEBUG
-
-bool ts_guc_debug_have_int128;
-
 static const struct config_enum_entry debug_require_options[] = { { "allow", DRO_Allow, false },
 																  { "forbid", DRO_Forbid, false },
 																  { "require", DRO_Require, false },
+																  { "force", DRO_Force, false },
 																  { NULL, 0, false } };
 
+#ifdef TS_DEBUG
+
+bool ts_guc_debug_have_int128;
+
 DebugRequireOption ts_guc_debug_require_vector_qual = DRO_Allow;
 
 DebugRequireOption ts_guc_debug_require_vector_agg = DRO_Allow;
 #endif
 
+DebugRequireOption ts_guc_debug_require_batch_sorted_merge = false;
+
 bool ts_guc_debug_compression_path_info = false;
 bool ts_guc_enable_rowlevel_compression_locking = false;
 
@@ -1189,6 +1190,18 @@ _guc_init(void)
 							/* assign_hook= */ NULL,
 							/* show_hook= */ NULL);
 
+	DefineCustomEnumVariable(/* name= */ MAKE_EXTOPTION("debug_require_batch_sorted_merge"),
+							 /* short_desc= */ "require batch sorted merge in DecompressChunk node",
+							 /* long_desc= */ "this is for debugging purposes",
+							 /* valueAddr= */ (int *) &ts_guc_debug_require_batch_sorted_merge,
+							 /* bootValue= */ DRO_Allow,
+							 /* options = */ debug_require_options,
+							 /* context= */ PGC_USERSET,
+							 /* flags= */ 0,
+							 /* check_hook= */ NULL,
+							 /* assign_hook= */ NULL,
+							 /* show_hook= */ NULL);
+
 #ifdef TS_DEBUG
 	DefineCustomBoolVariable(/* name= */ MAKE_EXTOPTION("shutdown_bgw_scheduler"),
 							 /* short_desc= */ "immediately shutdown the bgw scheduler",
@@ -1271,17 +1284,6 @@ _guc_init(void)
 							 /* assign_hook= */ NULL,
 							 /* show_hook= */ NULL);
 
-	DefineCustomBoolVariable(/* name= */ MAKE_EXTOPTION("debug_require_batch_sorted_merge"),
-							 /* short_desc= */ "require batch sorted merge in DecompressChunk node",
-							 /* long_desc= */ "this is for debugging purposes",
-							 /* valueAddr= */ &ts_guc_debug_require_batch_sorted_merge,
-							 /* bootValue= */ false,
-							 /* context= */ PGC_USERSET,
-							 /* flags= */ 0,
-							 /* check_hook= */ NULL,
-							 /* assign_hook= */ NULL,
-							 /* show_hook= */ NULL);
-
 	DefineCustomBoolVariable(/* name= */ MAKE_EXTOPTION("debug_allow_cagg_with_deprecated_funcs"),
 							 /* short_desc= */ "allow new caggs using time_bucket_ng",
 							 /* long_desc= */ "this is for debugging/testing purposes",

src/guc.h

@@ -119,14 +119,15 @@ extern char *ts_current_timestamp_mock;
 
 extern TSDLLEXPORT int ts_guc_debug_toast_tuple_target;
 
-#ifdef TS_DEBUG
 typedef enum DebugRequireOption
 {
 	DRO_Allow = 0,
 	DRO_Forbid,
-	DRO_Require
+	DRO_Require,
+	DRO_Force,
 } DebugRequireOption;
 
+#ifdef TS_DEBUG
 extern TSDLLEXPORT DebugRequireOption ts_guc_debug_require_vector_qual;
 
 extern TSDLLEXPORT DebugRequireOption ts_guc_debug_require_vector_agg;
@@ -136,7 +137,7 @@ extern TSDLLEXPORT DebugRequireOption ts_guc_debug_require_vector_agg;
 extern TSDLLEXPORT bool ts_guc_debug_compression_path_info;
 extern TSDLLEXPORT bool ts_guc_enable_rowlevel_compression_locking;
 
-extern TSDLLEXPORT bool ts_guc_debug_require_batch_sorted_merge;
+extern TSDLLEXPORT DebugRequireOption ts_guc_debug_require_batch_sorted_merge;
 
 extern TSDLLEXPORT bool ts_guc_debug_allow_cagg_with_deprecated_funcs;
 

tsl/src/nodes/decompress_chunk/decompress_chunk.c

@@ -990,13 +990,27 @@ ts_decompress_chunk_generate_paths(PlannerInfo *root, RelOptInfo *chunk_rel, con
 			batch_merge_path->custom_path.path.pathkeys = root->query_pathkeys;
 			cost_batch_sorted_merge(root, compression_info, batch_merge_path, compressed_path);
 
+			if (ts_guc_debug_require_batch_sorted_merge == DRO_Force)
+			{
+				batch_merge_path->custom_path.path.startup_cost = cpu_tuple_cost;
+				batch_merge_path->custom_path.path.total_cost = 2 * cpu_tuple_cost;
+			}
+
 			/* If the chunk is partially compressed, prepare the path only and add it later
 			 * to a merge append path when we are able to generate the ordered result for the
 			 * compressed and uncompressed part of the chunk.
 			 */
 			if (!consider_partial)
 				add_path(chunk_rel, &batch_merge_path->custom_path.path);
 		}
+		else if (ts_guc_debug_require_batch_sorted_merge == DRO_Require ||
+				 ts_guc_debug_require_batch_sorted_merge == DRO_Force)
+		{
+			ereport(ERROR,
+					(errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
+					 errmsg("debug: batch sorted merge is required but not possible at planning "
+							"time")));
+		}
 
 		/* If we can push down the sort below the DecompressChunk node, we set the pathkeys of
 		 * the decompress node to the query pathkeys, while remembering the compressed_pathkeys

tsl/src/nodes/decompress_chunk/exec.c

@@ -383,13 +383,22 @@ decompress_chunk_begin(CustomScanState *node, EState *estate, int eflags)
 		chunk_state->exec_methods.ExecCustomScan = decompress_chunk_exec_fifo;
 	}
 
-	if (ts_guc_debug_require_batch_sorted_merge && !dcontext->batch_sorted_merge)
+	if ((ts_guc_debug_require_batch_sorted_merge == DRO_Require ||
+		 ts_guc_debug_require_batch_sorted_merge == DRO_Force) &&
+		!dcontext->batch_sorted_merge)
 	{
 		ereport(ERROR,
 				(errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 				 errmsg("debug: batch sorted merge is required but not used")));
 	}
 
+	if (ts_guc_debug_require_batch_sorted_merge == DRO_Forbid && dcontext->batch_sorted_merge)
+	{
+		ereport(ERROR,
+				(errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
+				 errmsg("debug: batch sorted merge is used when it is forbidden")));
+	}
+
 	/* Constify stable expressions in vectorized predicates. */
 	PlannerGlobal glob = {
 		.boundParams = node->ss.ps.state->es_param_list_info,