fix: do not let a custom scan hide INHERITS children

CHANGELOG.md

@@ -117,6 +117,22 @@ true until the next version shipped.
   `rte->tablesample` is set, so the query reaches the callbacks and raises. An
   unsampled query on the same table still takes the custom scan.
 
+- A custom scan no longer hides the children of an `INHERITS` hierarchy (#871).
+
+  A legacy inheritance parent is an ordinary `RELKIND_RELATION` carrying
+  `rte->inh`. `set_rel_pathlist_hook` still runs on that appendrel after core
+  has built the `Append` over its children, and a custom scan added there reads
+  only the parent's own storage. Every child's rows became invisible, with no
+  error and no warning. Measured with a heap mirror as the oracle, parent 1 row
+  and child 5000: `SELECT count(*) FROM parent` returned **1 where 5001 is
+  right**, from a plan containing no `Append`. The ungrouped vector aggregate
+  path had the same hole; the grouped path already refused this shape.
+
+  Both hooks now decline when `rte->inh` is set, leaving the hierarchy to the
+  ordinary `Append` plan. The children keep the custom scan, because a child's
+  own range-table entry has `inh` false: the resulting plan is an `Append` of
+  two `PgColumnarScan` nodes with pushdown intact, not a fallback to `Seq Scan`.
+
 - A shebang and the execute bit go together, and every directory that documents
   a command is swept (#856). Two things were left over from #852.
 

src/columnar_customscan.c

@@ -2535,6 +2535,18 @@ PgColumnarSetRelPathlist(PlannerInfo *root, RelOptInfo *rel, Index rti,
 	 */
 	if (rte->tablesample != NULL)
 		return;
+	/*
+	 * A legacy inheritance parent is a plain RELKIND_RELATION with rte->inh
+	 * set. set_rel_pathlist_hook still runs on that appendrel after core has
+	 * built the Append of its children, and a Custom Scan added here reads
+	 * only the parent's storage. Measured: parent 1 row, child 5000 rows,
+	 * SELECT count(*) FROM parent returned 1 (the heap mirror returned 5001)
+	 * from a plan with no Append. Leave the tree to Append. Children and
+	 * the parent-as-member still receive this scan: those RTEs have inh
+	 * false. Grouped vector aggregation already refuses the same shape.
+	 */
+	if (rte->inh)
+		return;
 	/*
 	 * A partition is RELOPT_OTHER_MEMBER_REL, not RELOPT_BASEREL, and excluding
 	 * it cost the custom scan entirely: no column projection, no zone-map

src/columnar_vector.c

@@ -961,6 +961,15 @@ PgColumnarCreateUpperPaths(PlannerInfo *root, UpperRelationKind stage,
 		return;
 	if (rte->tablesample != NULL)
 		return;
+	/*
+	 * A legacy inheritance parent is a plain RELKIND_RELATION with rte->inh set;
+	 * its children hold rows this single-relation scan would never see. Leave the
+	 * whole tree to the ordinary Append + Agg plan. The grouped path already
+	 * refuses this; the ungrouped path did not, so SELECT count(*) FROM parent
+	 * returned only the parent's own rows.
+	 */
+	if (rte->inh)
+		return;
 	if (!OidIsValid(rte->relid) || !PgColumnarIsColumnarRelation(rte->relid))
 		return;
 	relid = rte->relid;

test/inheritance.sh

@@ -0,0 +1,89 @@
+#!/usr/bin/env bash
+#
+# pgColumnar: a legacy INHERITS parent must plan as Append of its children, not
+# as a custom scan (or vectorized aggregate) of the parent's storage alone.
+#
+# PgColumnarSetRelPathlist and the ungrouped vector-aggregate path both run on
+# the inheritance appendrel (RELKIND_RELATION with rte->inh set). A scan added
+# there reads only the parent's own file. Measured: parent 1 row, child 5000
+# rows, SELECT count(*) FROM parent returned 1 where the heap mirror returned
+# 5001, from a plan with no Append. Grouped vector aggregation already refused
+# this shape; the other two hooks did not.
+#
+# Declarative partitions are a different RTE (RELKIND_PARTITIONED_TABLE) and
+# already go through Append of OTHER_MEMBER_REL children. That path stays.
+#
+# Usage:  test/inheritance.sh [PG_CONFIG]
+# Written fresh for pgColumnar.
+
+set -uo pipefail
+. "$(dirname "${BASH_SOURCE[0]}")/lib.sh"
+pgc_setup "${1:-/usr/local/pg17/bin/pg_config}"
+
+plan() {
+	env PATH="$PGC_BINDIR:$PATH" psql -h 127.0.0.1 -p "$PGC_PORT" -U postgres \
+		-d "$PGC_DB" -Atq -c "$1"
+}
+
+psql_run "CREATE TABLE inh_h (id int, v text);"
+psql_run "CREATE TABLE inh_hc () INHERITS (inh_h);"
+psql_run "INSERT INTO inh_h VALUES (1, 'parent');"
+psql_run "INSERT INTO inh_hc VALUES (2, 'child');"
+
+psql_run "CREATE TABLE inh_p (id int, v text) USING pgcolumnar;"
+psql_run "CREATE TABLE inh_c () INHERITS (inh_p) USING pgcolumnar;"
+psql_run "INSERT INTO inh_p VALUES (1, 'parent');"
+psql_run "INSERT INTO inh_c VALUES (2, 'child');"
+
+check "INHERITS parent includes the child (heap oracle)" \
+	"$(q "SELECT count(*) FROM inh_h")" "2"
+check "INHERITS parent includes the child (columnar)" \
+	"$(q "SELECT count(*) FROM inh_p")" "2"
+check "ONLY parent is still just the parent" \
+	"$(q "SELECT count(*) FROM ONLY inh_p")" "1"
+check "the child is visible through the parent" \
+	"$(q "SELECT v FROM inh_p WHERE id = 2")" "child"
+check "row identities match the heap" \
+	"$(pgc_set_hash 'SELECT id, v FROM inh_p')" \
+	"$(pgc_set_hash 'SELECT id, v FROM inh_h')"
+
+star_plan="$(plan "EXPLAIN (COSTS OFF) SELECT * FROM inh_p;")"
+check "SELECT * from an INHERITS parent plans an Append" \
+	"$(printf '%s' "$star_plan" | grep -c Append)" "1"
+
+count_plan="$(plan "EXPLAIN (COSTS OFF) SELECT count(*) FROM inh_p;")"
+check "count(*) from an INHERITS parent plans an Append" \
+	"$(printf '%s' "$count_plan" | grep -c Append)" "1"
+
+# A cheap parent and a large child is the shape that made the custom scan win:
+# the appendrel is costed from the parent's empty-looking storage.
+psql_run "INSERT INTO inh_c SELECT g, 'c'||g FROM generate_series(10,5000) g;"
+psql_run "INSERT INTO inh_hc SELECT g, 'c'||g FROM generate_series(10,5000) g;"
+psql_run "ANALYZE inh_p; ANALYZE inh_c; ANALYZE inh_h; ANALYZE inh_hc;"
+
+check "a large child is still visible through the parent" \
+	"$(q "SELECT count(*) FROM inh_p")" \
+	"$(q "SELECT count(*) FROM inh_h")"
+check "filtered rows match the heap" \
+	"$(pgc_set_hash 'SELECT id FROM inh_p WHERE id > 100')" \
+	"$(pgc_set_hash 'SELECT id FROM inh_h WHERE id > 100')"
+
+vec_count="$(q "SET pgcolumnar.enable_ungrouped_vector_agg = on;
+                SELECT count(*) FROM inh_p;")"
+check "ungrouped vector aggregation also sees the child rows" \
+	"$(printf '%s' "$vec_count" | tail -1)" \
+	"$(q "SELECT count(*) FROM inh_h")"
+
+# Declarative partitions were the #436 case: the custom scan MUST still fire on
+# the child, or this refusal of rte->inh would put them back on a seqscan.
+psql_run "CREATE TABLE inh_prt (id int, v text) PARTITION BY RANGE (id);"
+psql_run "CREATE TABLE inh_prt1 PARTITION OF inh_prt FOR VALUES FROM (0) TO (100000)
+          USING pgcolumnar;"
+psql_run "INSERT INTO inh_prt SELECT g, 'p'||g FROM generate_series(1,2000) g;"
+check "a partitioned parent still returns every row" \
+	"$(q "SELECT count(*) FROM inh_prt")" "2000"
+prt_plan="$(plan "EXPLAIN (COSTS OFF) SELECT * FROM inh_prt;")"
+check "and the partition is still a PgColumnarScan (#436)" \
+	"$(printf '%s' "$prt_plan" | grep -c 'Custom Scan (PgColumnarScan)')" "1"
+
+pgc_summary

test/run_all_versions.sh

@@ -112,6 +112,7 @@ SUITES=(
 	index_delete_liveness
 	index_fetch_penalty_width
 	index_only
+	inheritance
 	int8_agg_int128
 	isolation
 	local_open_race_free