model-writtenrun 8472 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Colin McDonnell reported that emergency anti-wraparound autovacuum could stop indefinitely with autovacuum disabled and one worker slot: a departing worker still occupied the postmaster child slot, the next start failed, and the emergency launcher exited without reclaiming its WorkerInfo reservation #1.
- Colin proposed keeping the emergency launcher alive until worker startup, retrying failed starts and reclaiming timed-out reservations through cleanup shared with the normal launcher; his reproduction tests showed the patch prevented permanent stalls #1.
- Sami Imseih reproduced the failure and recommended fixing HEAD and backpatching through PostgreSQL 18; his review requested reuse of reservations after error recovery, clearing stale failure flags, combining startup and waiting into one helper, checking interrupts on every iteration, rechecking reservations after the retry pause, and asserting the cleanup lock #2.
- Colin incorporated Sami’s review in the second patch and reported that an assert-enabled HEAD build vacuumed every database in four reproduction runs, retrying all 58 failed starts; make check passed and pgindent made no changes #3.
- The backpatch wakeup behavior remained unresolved: Sami favored leaving the separate launcher-PID behavior unchanged on PostgreSQL 18 and 19, while Colin measured substantial delays from waiting for one-second timeouts and proposed registering the emergency launcher PID on those branches; the thread ends with the revised patch awaiting further review and no linked commit #2 #3.