model-writtenrun 9088 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tom Lane proposed checking child-slot availability before spawning background workers after a reported connection surge caused a postmaster crash; he reused canAcceptConnections() and initially deferred separate process pools because later PGPROC exhaustion could still prevent worker startup #1.
- Tom committed the crash fix with comment-only additions as 3887e9455f, based on #1, and backpatched through PostgreSQL 9.5; the commit explains that 9.4 needed a different fix and had too little background-worker usage to justify it.
- After the commit, Robert Haas explained that parallel queries tolerate failure to reserve worker slots but error out when registered workers fail to start, while Tom argued that resource exhaustion should permit graceful fallback; Robert defended reporting fork failures rather than silently running slower #2 #3 #4 #5 #6.
- Tom then posted a delta reserving separate child-process capacity for autovacuum workers and background workers, preventing connection floods from consuming those slots, while questioning whether the added complexity was preferable to more robust parallel-worker startup #7.
- When discussion resumed in November, Alvaro Herrera emphasized keeping the postmaster alive, Robert clarified that his objection concerned allowing individual queries to fail, and Tom confirmed that the postmaster crash was already fixed while graceful handling of fork failures remained disputed #8 #9 #10 #11.
- Andres Freund and Stephen Frost noted that strict memory-overcommit accounting can cause fork failures on otherwise usable systems; Andres also proposed using worker error queues and exit behavior to distinguish harmless startup failures from query failures #12 #13 #14.
- Robert warned that worker-status reuse, incomplete error reporting, and shutdown races complicate safe fallback, and defended exposing startup failures so users and developers would notice underlying problems #15.
- Andres concluded that existing stopped-worker and queue checks could support removing the initialization error, leaving the fallback policy unresolved; the crash fix had landed, but the separate-capacity patch had no recorded landing in the supplied thread #16 #7.