Multi-row `CREATE` followed by a second `CREATE` reports `vertex assigned to variable n0 was deleted`

[YGY-001]

Docker image apache/age:release_PG18_1.7.0

Reproduction

LOAD 'age';
SET search_path = ag_catalog, public;
SELECT ag_catalog.create_graph('age_issue_case5');

SELECT *
FROM ag_catalog.cypher('age_issue_case5', $$
UNWIND [-1452595301,16630552,2131315115] AS alias0
CREATE (n0 {id: 38, k0: 'lg', k1: -1164920309, klist: []})
WITH *
CREATE p0 = (n0) <-[r0:rt0 {k0: 'd', k1: 1141735256,
                              id: 90, klist: ['zt', '5', '4']}]-
            (n1 {id: 39, k0: 'i', klist: ['Ifg', 'h8']})
WITH alias0, n1.k0 AS alias1, lTrim(toString(n0.k0)) AS alias2,
     sqrt(abs(toFloat(r0.id))) AS alias3,
     n0, n1 AS alias4, r0 AS alias5, p0
RETURN collect(toString(alias5.k0)) AS alias6
$$) AS (alias6 agtype);

Expected

n0 remains bound across the WITH barrier for every row produced by
UNWIND, and the query returns the collected relationship property.

Actual

vertex assigned to variable n0 was deleted

[gregfelice]

I could not reproduce this on current master (4d9b2506, AGE 1.8.0, PostgreSQL 18.4). Running your reproducer verbatim returns the expected result with no error:

     alias6
-----------------
 ["d", "d", "d"]
(1 row)

No vertex assigned to variable n0 was deleted error.

Your report is against the apache/age:release_PG18_1.7.0 Docker image, so the most likely explanation is that this was fixed somewhere between 1.7.0 and 1.8.0.

To be careful about attributing it: I suspected #2495 ("Fix earlier clauses' writes being invisible to later clauses"), since it touches this area. I checked by rebuilding at 5f324f82, the commit immediately before that fix — your reproducer already passed there too, so #2495 is not what fixed this and I did not bisect further.

Could you retest against 1.8.0 / master and confirm? If it still fails for you there, the difference is environmental and worth knowing about. I would rather not close this pointing at a commit that demonstrably is not the fix.

[YGY-001]

I could not reproduce this on current master (4d9b2506, AGE 1.8.0, PostgreSQL 18.4). Running your reproducer verbatim returns the expected result with no error:

     alias6
-----------------
 ["d", "d", "d"]
(1 row)

No vertex assigned to variable n0 was deleted error.

Your report is against the apache/age:release_PG18_1.7.0 Docker image, so the most likely explanation is that this was fixed somewhere between 1.7.0 and 1.8.0.

To be careful about attributing it: I suspected #2495 ("Fix earlier clauses' writes being invisible to later clauses"), since it touches this area. I checked by rebuilding at 5f324f82, the commit immediately before that fix — your reproducer already passed there too, so #2495 is not what fixed this and I did not bisect further.

Could you retest against 1.8.0 / master and confirm? If it still fails for you there, the difference is environmental and worth knowing about. I would rather not close this pointing at a commit that demonstrably is not the fix.

You're right; it works fine in version 1.8.