model-writtenrun 7638 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Alexander Kukushkin reported that autovacuum workers could continue for 20–25 minutes after the postmaster was killed, providing reproduction steps and backtraces showing vacuum-delay sleeps in #1.
- Tom Lane explained that busy backends also detect postmaster death late and opposed constant polling in #3, while Stephen Frost argued that long-running autovacuum warranted checks during its existing sleeps in #4.
- Tom Lane and Alvaro Herrera identified vacuum_delay_point() as a suitable place to check before sleeping in #5 and #6, and Andres Freund proposed replacing pg_usleep() with a latch wait in #8.
- Broader detection remained exploratory: Alvaro suggested periodic buffer-access checks in #10, Tom opposed additional syscalls on normal execution paths in #11, and Stephen proposed peer signaling or I/O-time checks in #12.
- Stephen posted the narrowly scoped WaitLatch() patch and reported that light testing addressed the original reproduction in #13; Robert Haas discussed detecting client disconnects too in #14, while Thomas Munro described timer-based alternatives and supported fixing this particular sleep in #15.
- After Thomas's endorsement in #16, Stephen posted an unchanged rebase with a commit message in #17 and proposed backpatching, asking which wait event older releases should use.
- Robert Haas recommended a dedicated wait event if backpatched in #18, and Michael Paquier opposed backpatching while explaining safe enum placement in #19; Stephen defended backpatching in #20 but ultimately acknowledged the impasse and chose to proceed on HEAD in #21.
- Stephen committed #17 unchanged as 4753ef37e0 and confirmed completion in #22; the HEAD fix landed, while backpatching remained unresolved and broader detection outside vacuum sleeps was not implemented here.