model-writtenrun 3469 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- The thread began with Tom Lane's cleanup of a subscription-test role that broke repeated installcheck runs, prompting Robert Haas to request an automated guard against such leftovers #1 #2.
- Tom Lane proposed checking global objects before and after tests while allowing pre-existing objects; after Daniel Gustafsson questioned count-only comparisons, Tom favored comparing sorted names and envisioned an always-enabled check #3 #4 #5.
- Daniel Gustafsson submitted a configurable verification test that saved its initial output for comparison after the suite, avoiding a new query mechanism, and included an independently discovered runtime-reporting fix #7.
- The runtime fix from #7 landed unchanged as 8cb94344c3, and Daniel resubmitted only the remaining verification work #8.
- Peter Eisentraut objected that the configurable test required separate integration into every suite; Alvaro Herrera's proposed SQL-table solution had the same limitation, so discussion settled on direct libpq queries now that pg_regress already depended on libpq #10 #11 #12 #13 #14 #15.
- Daniel replaced the original approach with built-in comparisons of roles, tablespaces, and subscriptions, reporting that check-world passed after adding cleanup for a role left behind by test_pg_dump #16.
- Heikki Linnakangas questioned extension-wide enforcement and unstable query ordering; Daniel agreed that ordering needed correction and explained that databases could remain for pg_upgrade testing, although the later attached patches retained the original query ordering #17 #18 #20 #25.
- Alvaro Herrera suggested replacing the cleanup's explicit revocations with DROP OWNED BY, and Daniel incorporated that simplification while retaining DROP ROLE #19 #20.
- Vignesh C reported FreeBSD CI failures involving apparent leftover roles; Daniel paused the work to investigate, and Peter Eisentraut initially suspected regression-name enforcement #21 #22 #23.
- Andres Freund suggested interference from concurrent tests, and Daniel confirmed that parallel contrib suites sharing an instance polluted each other's global-object snapshots; Daniel made checking opt-in for the main regression suite and reported passing CI, while questioning its restricted usefulness #24 #25.
- Tom Lane argued that even the main suite's isolation was accidental, and Daniel agreed to abandon automatic detection and proceed only with the useful cleanup before closing the CommitFest entry #26 #27.
- The retained test_pg_dump cleanup from #25 subsequently landed unchanged as 936e3fa378, whose commit message confirms that the automatic checker was withdrawn because of parallel-run problems.