Apply sort transform optimizations to compressed chunks

src/planner/planner.c

@@ -56,6 +56,7 @@
 #include "partitioning.h"
 #include "planner/partialize.h"
 #include "planner/planner.h"
+#include "sort_transform.h"
 #include "utils.h"
 
 #include "compat/compat.h"
@@ -935,8 +936,6 @@ ts_classify_relation(const PlannerInfo *root, const RelOptInfo *rel, Hypertable
 	return TS_REL_OTHER;
 }
 
-extern void ts_sort_transform_optimization(PlannerInfo *root, RelOptInfo *rel);
-
 static inline bool
 should_chunk_append(Hypertable *ht, PlannerInfo *root, RelOptInfo *rel, Path *path, bool ordered,
 					int order_attno)
@@ -1183,14 +1182,47 @@ apply_optimizations(PlannerInfo *root, TsRelType reltype, RelOptInfo *rel, Range
 			break;
 		case TS_REL_CHUNK_STANDALONE:
 		case TS_REL_CHUNK_CHILD:
-			ts_sort_transform_optimization(root, rel);
+		{
 			/*
 			 * Since the sort optimization adds new paths to the rel it has
 			 * to happen before any optimizations that replace pathlist.
 			 */
-			if (ts_cm_functions->set_rel_pathlist_query != NULL)
-				ts_cm_functions->set_rel_pathlist_query(root, rel, rel->relid, rte, ht);
+			List *transformed_query_pathkeys = ts_sort_transform_get_pathkeys(root, rel, rte, ht);
+			if (transformed_query_pathkeys != NIL)
+			{
+				List *orig_query_pathkeys = root->query_pathkeys;
+				root->query_pathkeys = transformed_query_pathkeys;
+
+				/* Create index paths with transformed pathkeys */
+				create_index_paths(root, rel);
+
+				/*
+				 * Call the TSL hooks with the transformed pathkeys as well, so
+				 * that the decompression paths also use this optimization.
+				 */
+				if (ts_cm_functions->set_rel_pathlist_query != NULL)
+					ts_cm_functions->set_rel_pathlist_query(root, rel, rel->relid, rte, ht);
+
+				root->query_pathkeys = orig_query_pathkeys;
+
+				/*
+				 * change returned paths to use original pathkeys. have to go through
+				 * all paths since create_index_paths might have modified existing
+				 * pathkey. Always safe to do transform since ordering of
+				 * transformed_query_pathkey implements ordering of
+				 * orig_query_pathkeys.
+				 */
+				ts_sort_transform_replace_pathkeys(rel->pathlist,
+												   transformed_query_pathkeys,
+												   orig_query_pathkeys);
+			}
+			else
+			{
+				if (ts_cm_functions->set_rel_pathlist_query != NULL)
+					ts_cm_functions->set_rel_pathlist_query(root, rel, rel->relid, rte, ht);
+			}
 			break;
+		}
 		default:
 			break;
 	}

src/sort_transform.c

@@ -16,7 +16,10 @@
 #include <utils/lsyscache.h>
 
 #include "compat/compat.h"
+#include "cross_module_fn.h"
 #include "func_cache.h"
+#include "hypertable.h"
+#include "import/allpaths.h"
 #include "sort_transform.h"
 
 /* This optimizations allows GROUP BY clauses that transform time in
@@ -29,8 +32,6 @@
  * to an ordering on time.
  */
 
-extern void ts_sort_transform_optimization(PlannerInfo *root, RelOptInfo *rel);
-
 static Expr *
 transform_timestamp_cast(FuncExpr *func)
 {
@@ -384,8 +385,9 @@
  *	For example: an ORDER BY date_trunc('minute', time) can be implemented by
  *	an ordering of time.
  */
-void
-ts_sort_transform_optimization(PlannerInfo *root, RelOptInfo *rel)
+List *
+ts_sort_transform_get_pathkeys(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte,
+							   Hypertable *ht)
 {
 	/*
 	 * We attack this problem in three steps:
@@ -399,17 +401,16 @@
 	 *
 	 */
 	ListCell *lc;
-	List *transformed_query_pathkey = NIL;
-	List *orig_query_pathkeys = root->query_pathkeys;
+	List *transformed_query_pathkeys = NIL;
 	PathKey *last_pk;
 	PathKey *new_pk;
 	EquivalenceClass *transformed;
 
 	/*
 	 * nothing to do for empty pathkeys
 	 */
-	if (orig_query_pathkeys == NIL)
-		return;
+	if (root->query_pathkeys == NIL)
+		return NIL;
 
 	/*
 	 * These sort transformations are only safe for single member ORDER BY
@@ -420,7 +421,7 @@
 	transformed = sort_transform_ec(root, last_pk->pk_eclass);
 
 	if (transformed == NULL)
-		return;
+		return NIL;
 
 	new_pk = make_canonical_pathkey(root,
 									transformed,
@@ -434,30 +435,72 @@
 	foreach (lc, root->query_pathkeys)
 	{
 		if (lfirst(lc) != last_pk)
-			transformed_query_pathkey = lappend(transformed_query_pathkey, lfirst(lc));
+			transformed_query_pathkeys = lappend(transformed_query_pathkeys, lfirst(lc));
 		else
-			transformed_query_pathkey = lappend(transformed_query_pathkey, new_pk);
+			transformed_query_pathkeys = lappend(transformed_query_pathkeys, new_pk);
 	}
 
-	/* search for indexes on transformed pathkeys */
-	root->query_pathkeys = transformed_query_pathkey;
-	create_index_paths(root, rel);
-	root->query_pathkeys = orig_query_pathkeys;
+	return transformed_query_pathkeys;
+}
 
-	/*
-	 * change returned paths to use original pathkeys. have to go through
-	 * all paths since create_index_paths might have modified existing
-	 * pathkey. Always safe to do transform since ordering of
-	 * transformed_query_pathkey implements ordering of
-	 * orig_query_pathkeys.
-	 */
-	foreach (lc, rel->pathlist)
+/*
+ * After we have created new paths with transformed pathkeys, replace them back
+ * with the original pathkeys.
+ */
+void
+ts_sort_transform_replace_pathkeys(void *node, List *transformed_pathkeys, List *original_pathkeys)
+{
+	if (node == NULL)
 	{
-		Path *path = lfirst(lc);
+		return;
+	}
 
-		if (compare_pathkeys(path->pathkeys, transformed_query_pathkey) == PATHKEYS_EQUAL)
+	if (IsA(node, List))
+	{
+		List *list = castNode(List, node);
+		ListCell *lc;
+		foreach (lc, list)
 		{
-			path->pathkeys = orig_query_pathkeys;
+			ts_sort_transform_replace_pathkeys(lfirst(lc), transformed_pathkeys, original_pathkeys);
 		}
+		return;
+	}
+
+	Path *path = (Path *) node;
+	if (compare_pathkeys(path->pathkeys, transformed_pathkeys) == PATHKEYS_EQUAL)
+	{
+		path->pathkeys = original_pathkeys;
+	}
+
+	if (IsA(path, CustomPath))
+	{
+		/*
+		 * We should only see ChunkAppend here.
+		 */
+		CustomPath *custom = castNode(CustomPath, path);
+		ts_sort_transform_replace_pathkeys(custom->custom_paths,
+										   transformed_pathkeys,
+										   original_pathkeys);
+	}
+	else if (IsA(path, MergeAppendPath))
+	{
+		MergeAppendPath *append = castNode(MergeAppendPath, path);
+		ts_sort_transform_replace_pathkeys(append->subpaths,
+										   transformed_pathkeys,
+										   original_pathkeys);
+	}
+	else if (IsA(path, AppendPath))
+	{
+		AppendPath *append = castNode(AppendPath, path);
+		ts_sort_transform_replace_pathkeys(append->subpaths,
+										   transformed_pathkeys,
+										   original_pathkeys);
+	}
+	else if (IsA(path, ProjectionPath))
+	{
+		ProjectionPath *projection = castNode(ProjectionPath, path);
+		ts_sort_transform_replace_pathkeys(projection->subpath,
+										   transformed_pathkeys,
+										   original_pathkeys);
 	}
 }

src/sort_transform.h

@@ -7,4 +7,12 @@
 
 #include <postgres.h>
 
+#include "hypertable.h"
+
 extern Expr *ts_sort_transform_expr(Expr *expr);
+
+extern List *ts_sort_transform_get_pathkeys(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte,
+											Hypertable *ht);
+
+extern void ts_sort_transform_replace_pathkeys(void *node, List *transformed_pathkeys,
+											   List *original_pathkeys);

tsl/src/chunkwise_agg.c

@@ -437,7 +437,7 @@ generate_agg_pushdown_path(PlannerInfo *root, Path *cheapest_total_path, RelOptI
 										  copy_append_like_path(root,
 																partially_compressed_append,
 																partially_compressed_sorted,
-																subpath->pathtarget));
+																partial_grouping_target));
 			}
 
 			if (can_hash)
@@ -446,7 +446,7 @@ generate_agg_pushdown_path(PlannerInfo *root, Path *cheapest_total_path, RelOptI
 										  copy_append_like_path(root,
 																partially_compressed_append,
 																partially_compressed_hashed,
-																subpath->pathtarget));
+																partial_grouping_target));
 			}
 		}
 		else

tsl/src/nodes/decompress_chunk/planner.c

@@ -961,6 +961,50 @@
 	plan->plan.parallel_safe = lefttree->parallel_safe;
 }
 
+/*
+ * Find a variable of the given relation somewhere in the expression tree.
+ * Currently we use this to find the Var argument of time_bucket, when we prepare
+ * the batch sorted merge parameters after using the monotonous sorting transform
+ * optimization.
+ */
+static Var *
+find_var_subexpression(void *expr, Index varno)
+{
+	if (IsA(expr, Var))
+	{
+		Var *var = castNode(Var, expr);
+		if ((Index) var->varno == (Index) varno)
+		{
+			return var;
+		}
+
+		return NULL;
+	}
+
+	if (IsA(expr, List))
+	{
+		List *list = castNode(List, expr);
+		ListCell *lc;
+		foreach (lc, list)
+		{
+			Var *var = find_var_subexpression(lfirst(lc), varno);
+			if (var != NULL)
+			{
+				return var;
+			}
+		}
+
+		return NULL;
+	}
+
+	if (IsA(expr, FuncExpr))
+	{
+		return find_var_subexpression(castNode(FuncExpr, expr)->args, varno);
+	}
+
+	return NULL;
+}
+
 Plan *
 decompress_chunk_plan_create(PlannerInfo *root, RelOptInfo *rel, CustomPath *path,
 							 List *output_targetlist, List *clauses, List *custom_plans)
@@ -1136,18 +1180,22 @@
 					continue;
 				}
 
-				Ensure(IsA(em->em_expr, Var),
-					   "non-Var pathkey not expected for compressed batch sorted merge");
-
 				/*
-				 * We found a Var equivalence member that belongs to the
-				 * decompressed relation. We have to convert its varattno which
-				 * is the varattno of the uncompressed chunk tuple, to the
-				 * decompressed scan tuple varattno.
+				 * The equivalence member expression might be a monotonous
+				 * expression of the decompressed relation Var, so recurse to
+				 * find it.
 				 */
-				Var *var = castNode(Var, em->em_expr);
+				Var *var = find_var_subexpression(em->em_expr, em_relid);
+				Ensure(var != NULL,
+					   "non-Var pathkey not expected for compressed batch sorted merge");
+
 				Assert((Index) var->varno == (Index) em_relid);
 
+				/*
+				 * Convert its varattno which is the varattno of the
+				 * uncompressed chunk tuple, to the decompressed scan tuple
+				 * varattno.
+				 */
 				const int decompressed_scan_attno =
 					context.uncompressed_attno_info[var->varattno].custom_scan_attno;
 				Assert(decompressed_scan_attno > 0);