Run our hypertable expansion only from the get_relation_info_hook (#9714)

We used to run it from preprocessing the query as well, but this
complicates working with the DML target expansion code. Let the standard
Postgres code run where possible, to reduce the divergence.

Author: akuzm

src/import/allpaths.c

@@ -34,6 +34,7 @@
 
 #include "allpaths.h"
 #include "chunk.h"
+#include "compat/compat.h"
 #include "cross_module_fn.h"
 #include "planner/planner.h"
 
@@ -224,6 +225,37 @@ ts_set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *parent_rel, Index pare
 			parent_rel->consider_startup = true;
 	}
 
+#if PG16_LT
+	/*
+	 * On PG15, create_append_path() calls get_appendrel_parampathinfo() instead
+	 * of the get_baserel_parampathinfo() like the later versions. The appendrel
+	 * parameterization info is not build for hypertable because of how we're
+	 * disabling the PG inheritance expansion, so here we have to compensate for
+	 * it. Fetch the parameterization info from the chunks here, following the
+	 * logic similar to add_paths_to_append_rel().
+	 */
+	if (parent_rel->reloptkind == RELOPT_BASEREL)
+	{
+		foreach (l, live_childrels)
+		{
+			RelOptInfo *child_rel = (RelOptInfo *) lfirst(l);
+			ListCell *lcp;
+
+			foreach (lcp, child_rel->pathlist)
+			{
+				Path *child_path = (Path *) lfirst(lcp);
+
+				if (child_path->param_info == NULL)
+					continue;
+
+				get_baserel_parampathinfo(root,
+										  parent_rel,
+										  child_path->param_info->ppi_req_outer);
+			}
+		}
+	}
+#endif
+
 	/* Add paths to the append relation. */
 	add_paths_to_append_rel(root, parent_rel, live_childrels);
 }

src/planner/planner.c

@@ -7,6 +7,7 @@
 #include <access/tsmapi.h>
 #include <access/xact.h>
 #include <catalog/namespace.h>
+#include <catalog/pg_inherits.h>
 #include <commands/extension.h>
 #include <executor/nodeAgg.h>
 #include <miscadmin.h>
@@ -447,20 +448,10 @@ preprocess_query(Node *node, PreprocessQueryContext *context)
 					if (ht)
 					{
 						/*
-						 * Mark hypertable RTEs we'd like to expand ourselves.
-						 * We always do this for SELECTs from hypertables.
-						 *
-						 * For DML, we also always expand the non-target relations.
-						 *
-						 * The hypertables that are not expanded by our custom code
-						 * here fall back to the standard Postgres inheritance
-						 * hierarchy expansion.
+						 * Hypertable expansion marking is done in the
+						 * get_relation_info_hook, which also handles
+						 * hypertables appearing after function	or view inlining.
 						 */
-						if (ts_guc_enable_optimizations && ts_guc_enable_constraint_exclusion &&
-							rte->inh && (Index) query->resultRelation != rti)
-						{
-							rte_mark_for_expansion(rte);
-						}
 
 						if (TS_HYPERTABLE_HAS_COMPRESSION_TABLE(ht))
 						{
@@ -1500,20 +1491,6 @@ timescaledb_set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, Index rti, Rang
 			break;
 
 		case TS_REL_HYPERTABLE:
-			/*
-			 * Set the indexlist for a hypertable parent to NIL since we
-			 * should not try to do any index scans on hypertable parents,
-			 * similar to how it works for partitioned tables.
-			 *
-			 * This can happen when building a merge join path and computing
-			 * cost for it. See get_actual_variable_range().
-			 *
-			 * This has to be after the hypertable is expanded, since the
-			 * indexlist is used during hypertable expansion.
-			 */
-
-			rel->indexlist = NIL;
-
 			if (!rte->inh)
 			{
 				/*
@@ -1565,15 +1542,21 @@ timescaledb_get_relation_info_hook(PlannerInfo *root, Oid relation_objectid, boo
 		{
 			/*
 			 * Mark hypertable RTEs we'd like to expand ourselves.
-			 * Hypertables inside inlineable functions don't get marked during
-			 * the query preprocessing step handled in preprocess_query().
-			 * Therefore we do an extra try here.
+			 * We always do this for SELECTs from hypertables.
+			 *
+			 * For DML, we also always expand the non-target relations.
+			 *
+			 * The hypertables that are not expanded by our custom code
+			 * here fall back to the standard Postgres inheritance
+			 * hierarchy expansion.
 			 *
-			 * For the explanation of the logic, see the comments in
-			 * preprocess_query().
+			 * `inhparent` goes to false in two cases: a hypertable without
+			 * chunks or a SELECT FROM ONLY hypertable. We still want to run our
+			 * hypertable expansion code for hypertables w/o chunks.
 			 */
-			if (ts_guc_enable_optimizations && ts_guc_enable_constraint_exclusion && inhparent &&
-				rte->ctename == NULL && rel->relid != (Index) query->resultRelation)
+			if (ts_guc_enable_optimizations && ts_guc_enable_constraint_exclusion &&
+				(inhparent || !has_subclass(rte->relid)) && rte->ctename == NULL &&
+				rel->relid != (Index) query->resultRelation)
 			{
 				rte_mark_for_expansion(rte);
 			}

test/expected/append-16.out

@@ -1746,7 +1746,7 @@ DROP INDEX :INDEX_NAME;
  Limit (actual rows=3.00 loops=1)
    ->  Merge Join (actual rows=3.00 loops=1)
          Merge Cond: (m2."time" = m1."time")
-         Join Filter: (m2.device_id = m1.device_id)
+         Join Filter: (m1.device_id = m2.device_id)
          Rows Removed by Join Filter: 4
          ->  Custom Scan (ChunkAppend) on join_limit m2 (actual rows=3.00 loops=1)
                Order: m2."time", m2.device_id

test/expected/append-17.out

@@ -1741,7 +1741,7 @@ DROP INDEX :INDEX_NAME;
  Limit (actual rows=3.00 loops=1)
    ->  Merge Join (actual rows=3.00 loops=1)
          Merge Cond: (m2."time" = m1."time")
-         Join Filter: (m2.device_id = m1.device_id)
+         Join Filter: (m1.device_id = m2.device_id)
          Rows Removed by Join Filter: 4
          ->  Custom Scan (ChunkAppend) on join_limit m2 (actual rows=3.00 loops=1)
                Order: m2."time", m2.device_id

test/expected/append-18.out

@@ -1744,7 +1744,7 @@ DROP INDEX :INDEX_NAME;
  Limit (actual rows=3.00 loops=1)
    ->  Merge Join (actual rows=3.00 loops=1)
          Merge Cond: (m2."time" = m1."time")
-         Join Filter: (m2.device_id = m1.device_id)
+         Join Filter: (m1.device_id = m2.device_id)
          Rows Removed by Join Filter: 4
          ->  Custom Scan (ChunkAppend) on join_limit m2 (actual rows=3.00 loops=1)
                Order: m2."time", m2.device_id

tsl/src/nodes/skip_scan/planner.c

@@ -1188,41 +1188,40 @@ static Var *
 get_distinct_var(PlannerInfo *root, Expr *tlexpr, IndexPath *index_path, Path *child_path,
 				 SkipKeyInfo *skinfo)
 {
-	RelOptInfo *rel = child_path->parent;
+	RelOptInfo *chunk_rel = child_path->parent;
 	RelOptInfo *indexed_rel = index_path->path.parent;
 
 	Assert(tlexpr && IsA(tlexpr, Var));
 	Var *var = castNode(Var, tlexpr);
 
 	RangeTblEntry *ht_rte = planner_rt_fetch(var->varno, root);
 
-	/* check whether a skip var is declared NOT NULL
-	 *  it's enough to check hypertable for NOT NULL
-	 *  as NOT NULL constraint will be propagated to and checked on all chunks
+	/*
+	 * Check whether a skip var is declared NOT NULL. It's enough to check
+	 * hypertable for NOT NULL, because the NOT NULL constraint will be
+	 * propagated to and checked on all chunks. Postgres doesn't set
+	 * RelOptInfo.notnullattnums for hypertable because it's an inheritance
+	 * parent, so check it against the catalog.
 	 */
-#if PG17_LT
 	skinfo->notnull = ts_get_attnotnull(ht_rte->relid, var->varattno);
-#else
-	RelOptInfo *baserel = ((Index) var->varno == rel->relid ? rel : rel->parent);
-	skinfo->notnull = bms_is_member(var->varattno, baserel->notnullattnums);
-#endif
 
 	/* If we are dealing with a hypertable Var extracted from distinctClause will point to
 	 * the parent hypertable while the IndexPath will be on a Chunk.
 	 * For a normal PG table they point to the same relation and we are done here. */
-	if ((Index) var->varno == rel->relid)
+	if ((Index) var->varno == chunk_rel->relid)
 	{
 		/* Get attribute number for distinct column on a normal PG table */
 		skinfo->indexed_column_attno = var->varattno;
 		return var;
 	}
 
-	RangeTblEntry *chunk_rte = planner_rt_fetch(rel->relid, root);
+	RangeTblEntry *chunk_rte = planner_rt_fetch(chunk_rel->relid, root);
 	RangeTblEntry *indexed_rte =
-		(indexed_rel == rel ? chunk_rte : planner_rt_fetch(indexed_rel->relid, root));
+		(indexed_rel == chunk_rel ? chunk_rte : planner_rt_fetch(indexed_rel->relid, root));
 
 	/* Check for hypertable */
-	if (!ts_is_hypertable(ht_rte->relid) || !bms_is_member(var->varno, rel->top_parent_relids))
+	if (!ts_is_hypertable(ht_rte->relid) ||
+		!bms_is_member(var->varno, chunk_rel->top_parent_relids))
 	{
 		return NULL;
 	}
@@ -1248,7 +1247,7 @@ get_distinct_var(PlannerInfo *root, Expr *tlexpr, IndexPath *index_path, Path *c
 		skinfo->indexed_column_attno = var->varattno;
 	}
 
-	var->varno = rel->relid;
+	var->varno = chunk_rel->relid;
 
 	return var;
 }

tsl/test/expected/transparent_decompression_ordered_index-16.out

@@ -424,7 +424,7 @@ GROUP BY device_id;
  GroupAggregate (actual rows=1.00 loops=1)
    Group Key: mt.device_id
    ->  Nested Loop (actual rows=48.00 loops=1)
-         Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (nd.node = mt.device_id))
+         Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (mt.device_id = nd.node))
          Rows Removed by Join Filter: 1493
          ->  Merge Append (actual rows=1541.00 loops=1)
                Sort Key: mt.device_id
@@ -452,7 +452,7 @@ WHERE mt.time > nd.start_time
 ORDER BY time;
 --- QUERY PLAN ---
  Nested Loop (actual rows=48.00 loops=1)
-   Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (nd.node = mt.device_id))
+   Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (mt.device_id = nd.node))
    Rows Removed by Join Filter: 1493
    ->  Custom Scan (ChunkAppend) on metrics_ordered_idx mt (actual rows=1541.00 loops=1)
          Order: mt."time"
@@ -1139,7 +1139,7 @@ ORDER BY "time", mt.device_id limit 5;
 --- QUERY PLAN ---
  Limit (actual rows=5.00 loops=1)
    ->  Nested Loop (actual rows=5.00 loops=1)
-         Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (nd.node = mt.device_id))
+         Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (mt.device_id = nd.node))
          Rows Removed by Join Filter: 1488
          ->  Custom Scan (ChunkAppend) on metrics_ordered_idx mt (actual rows=1493.00 loops=1)
                Order: mt."time", mt.device_id
@@ -1184,7 +1184,7 @@ ORDER BY "time", mt.device_id limit 5;
 --- QUERY PLAN ---
  Limit (actual rows=5.00 loops=1)
    ->  Nested Loop (actual rows=5.00 loops=1)
-         Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (nd.node = mt.device_id))
+         Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (mt.device_id = nd.node))
          Rows Removed by Join Filter: 1488
          ->  Custom Scan (ChunkAppend) on metrics_ordered_idx mt (actual rows=1493.00 loops=1)
                Order: mt."time", mt.device_id

tsl/test/expected/transparent_decompression_ordered_index-17.out

@@ -424,7 +424,7 @@ GROUP BY device_id;
  GroupAggregate (actual rows=1.00 loops=1)
    Group Key: mt.device_id
    ->  Nested Loop (actual rows=48.00 loops=1)
-         Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (nd.node = mt.device_id))
+         Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (mt.device_id = nd.node))
          Rows Removed by Join Filter: 1493
          ->  Merge Append (actual rows=1541.00 loops=1)
                Sort Key: mt.device_id
@@ -452,7 +452,7 @@ WHERE mt.time > nd.start_time
 ORDER BY time;
 --- QUERY PLAN ---
  Nested Loop (actual rows=48.00 loops=1)
-   Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (nd.node = mt.device_id))
+   Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (mt.device_id = nd.node))
    Rows Removed by Join Filter: 1493
    ->  Custom Scan (ChunkAppend) on metrics_ordered_idx mt (actual rows=1541.00 loops=1)
          Order: mt."time"
@@ -1139,7 +1139,7 @@ ORDER BY "time", mt.device_id limit 5;
 --- QUERY PLAN ---
  Limit (actual rows=5.00 loops=1)
    ->  Nested Loop (actual rows=5.00 loops=1)
-         Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (nd.node = mt.device_id))
+         Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (mt.device_id = nd.node))
          Rows Removed by Join Filter: 1488
          ->  Custom Scan (ChunkAppend) on metrics_ordered_idx mt (actual rows=1493.00 loops=1)
                Order: mt."time", mt.device_id
@@ -1184,7 +1184,7 @@ ORDER BY "time", mt.device_id limit 5;
 --- QUERY PLAN ---
  Limit (actual rows=5.00 loops=1)
    ->  Nested Loop (actual rows=5.00 loops=1)
-         Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (nd.node = mt.device_id))
+         Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (mt.device_id = nd.node))
          Rows Removed by Join Filter: 1488
          ->  Custom Scan (ChunkAppend) on metrics_ordered_idx mt (actual rows=1493.00 loops=1)
                Order: mt."time", mt.device_id

tsl/test/expected/transparent_decompression_ordered_index-18.out

@@ -1148,7 +1148,7 @@ ORDER BY "time", mt.device_id limit 5;
 --- QUERY PLAN ---
  Limit (actual rows=5.00 loops=1)
    ->  Nested Loop (actual rows=5.00 loops=1)
-         Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (nd.node = mt.device_id))
+         Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (mt.device_id = nd.node))
          Rows Removed by Join Filter: 1488
          ->  Custom Scan (ChunkAppend) on metrics_ordered_idx mt (actual rows=1493.00 loops=1)
                Order: mt."time", mt.device_id
@@ -1193,7 +1193,7 @@ ORDER BY "time", mt.device_id limit 5;
 --- QUERY PLAN ---
  Limit (actual rows=5.00 loops=1)
    ->  Nested Loop (actual rows=5.00 loops=1)
-         Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (nd.node = mt.device_id))
+         Join Filter: ((mt."time" > nd.start_time) AND (mt."time" < nd.stop_time) AND (mt.device_id = nd.node))
          Rows Removed by Join Filter: 1488
          ->  Custom Scan (ChunkAppend) on metrics_ordered_idx mt (actual rows=1493.00 loops=1)
                Order: mt."time", mt.device_id