model-writtenrun 7479 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tang reported a publisher-side lazy_scan_heap() assertion failure during concurrent transactions with logical replication in #1, and Michael Paquier identified the assertion as recently introduced in #2.
- Peter Geoghegan proposed an inline fix resetting all_visible_according_to_vm when releasing the visibility-map buffer in #3, but Tang tested it and reported that the failure persisted in #4.
- Masahiko Sawada explained in #5 that the visibility-map value can become stale before the heap buffer is locked, demonstrated how an inserted delay exposes the race, and attached a patch removing the assertion because later code already rechecks the visibility map.
- Dilip Kumar confirmed the locking-race analysis in #6, while Andres Freund questioned whether other uses of the stale variable could prevent marking a page all-visible and proposed refreshing it after locking in #7.
- Peter Geoghegan committed the assertion-removal patch from #5 unchanged as c9787385db and announced the push in #8.
- Peter Geoghegan argued in #9 that Andres Freund's concern involved the inverse case, but Andres clarified in #10 that a stale true value could still suppress setting the visibility-map bit; the assertion fix was landed, while this broader concern remained unresolved at the thread's end.