Multi-hop query plans flip to full edge scans + Merge Join when an attribute index on the start vertex exists

[NotHimmel]

Describe the bug

A 2-hop MATCH query changes its plan from per-hop index probes (Nested Loop,
sub-second) to a full scan over all 9,000,000 edges followed by a Merge Join
(17.1 s) as soon as an attribute index (GIN) exists on the start vertex's
filter property. Dropping that index restores the fast index-probe plan. The
query results are identical in both cases (count = 300), only the chosen plan
differs.

How are you accessing AGE (Command line, driver, etc.)?

• psql command line

What data setup do we need to do?

CREATE EXTENSION age;
LOAD 'age';
SET search_path = ag_catalog, public;

SELECT create_graph('repro_graph');
SELECT create_vlabel('repro_graph','VTABLE');
SELECT create_elabel('repro_graph','E1TABLE');   -- E2TABLE / E3TABLE same

-- 3,000,000 vertices: 10,000 distinct person_id values, 300 vertices each
INSERT INTO "VTABLE" (id, properties)
SELECT ag_catalog._graphid(3, g),
       ag_catalog.agtype_build_map('person_id', 'B' || lpad((g % 10000)::text, 8, '0'))
FROM generate_series(1, 3000000) g;

-- 3,000,000 edges per table: endpoints concentrated on ~1,000 hub nodes
-- (skewed relation-graph shape, e.g. person→merchant→person;
--  a uniform distribution does NOT reproduce the problem)
INSERT INTO "E1TABLE" (id, start_id, end_id, properties)
SELECT ag_catalog._graphid(4, g),
       ag_catalog._graphid(3, ((g-1) % 999) + 2),
       ag_catalog._graphid(3, ((g::bigint * 7919) % 1000) + 1),
       ag_catalog.agtype_build_map('etype', 'E1')
FROM generate_series(1, 3000000) g;
-- E2TABLE / E3TABLE analogous; full script attached (01_load_data.sql)

CREATE INDEX v_prop_gin ON repro_graph."VTABLE" USING gin (properties);
ANALYZE;

(The queried start value matches 300 vertices; its actual traversal touches
only ~600 edges, so replays are safe and results are stable.)

What is the necessary configuration info needed?

• shared_preload_libraries = 'age' (or LOAD 'age'; per session)
• Everything else default; no special GUCs

What is the command that caused the error?

EXPLAIN (ANALYZE)
SELECT * FROM cypher('repro_graph', $$
  MATCH (a:VTABLE {person_id: 'B00000001'})-->()-->(c)
  RETURN count(DISTINCT c) $$) AS (r agtype);

Output (excerpt):

Merge Join (actual rows=300)
  -> Nested Loop (actual rows=9,000,300)
       -> Merge Join
            -> Index Scan using "E1TABLE_start_id_idx" (actual rows=3,000,000)
            -> Index Scan using "E2TABLE_start_id_idx" (actual rows=3,000,000)
            -> Index Scan using "E3TABLE_start_id_idx" (actual rows=3,000,000)
  -> Sort (rows=300)
       -> Bitmap Index Scan on v_prop_gin
Execution Time: 17113.490 ms

Dropping v_prop_gin and re-running the same command restores per-hop
start_id / end_id index probes and completes in under a second.

For diagnosis: with the index in place, SET enable_mergejoin = off; brings
the same query down to 0.54 s — confirming the data and results are correct
and only the plan selection is affected.

Expected behavior

Whether or not the start-point filter uses an attribute index should not
change how subsequent hops are accessed: the plan should keep probing the
per-hop start_id / end_id indexes regardless.

Environment (please complete the following information):

• Version: AGE 1.7.0 (release), PostgreSQL 18, Linux x86_64