model-writtenrun 5558 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andres Freund opened the thread with intermittent FreeBSD CI failures in recovery-conflict tests, where replay waited indefinitely despite a conflicting buffer pin, and found evidence that the problem predated the latest changes #1 #2.
- After investigating latch wakeups and conflict signaling, Andres reproduced a self-deadlock on Linux: a timeout signal could interrupt cancellation code and acquire ProcArrayLock again through the same cancellation path #3 #4.
- Andres proposed making the signal handler set a flag, and Tom Lane supported that design; when Andres suggested a smaller timeout-reordering workaround for stable branches, Tom argued for applying the proper fix there too #5 #6 #7 #8.
- Andres developed a deadlock-conflict test to protect the fix and proposed backpatching recovery-conflict coverage, while raising separate concerns about disabling unrelated timeouts and assumptions about latch wakeups #9.
- The first submission included the server fix, a new deadlock test, and stable-branch test adaptations that omitted unsupported tablespace and statistics checks; Andres deferred broader timeout and wakeup cleanup to separate discussions #10.
- Tom supported retaining the test numbering and recommended backpatching shared helpers instead of inlining them, leading to the separate helper commit 5ab8e80148 #11.
- Andres landed the test extension from #10 as 21e184403b, the signal-handler fix as 8f1537d10e, and the adapted stable-branch tests as f74496dd61, confirming that the test extension had landed before the fix #12.
- Tom then reported that the deadlock test sometimes produced a buffer-pin timeout instead of the expected deadlock cancellation; Andres initially suspected missing lock synchronization, then recognized that the test already checked it and identified competing timeout behavior #13 #14 #15 #17.
- Andres submitted a timing fix that increased the standby delay and restarted the standby around the deadlock test, while noting an IPC::Run broken-pipe concern; that attachment landed unchanged as 9e6b7b45ca #19.
- Tom reported duplicate pump_until definitions left by the helper backpatch, and Andres agreed to remove them; 7fa95bb0ac switched the existing crash tests to the shared helper and its explicit timeout argument #20 #21.
- The timing fix was reverted in stable branches because adjust_conf was unavailable and restarting lost conflict statistics; Tom proposed postponing broader test-infrastructure changes until after the imminent releases, while Andres sought to preserve coverage by disabling only the failing section #21 #22 #23 #24.
- After finding that the observed failures involved the deadlock section, Andres submitted a narrow SKIP block in #25, which landed unchanged as f40d362a66; Tom's explicit approval followed the commit #26.
- A further failure on lapwing showed that disabling the deadlock section was insufficient; Andres suspected another genuine recovery-conflict bug, while Tom argued that shipping an unstable test would burden packagers without fixing that bug #29 #30 #31.
- Andres accepted disabling the entire test temporarily, and 55e5a5e0fa added the whole-file skip directly from that discussion #32 #33.
- In July, Alvaro Herrera asked about re-enabling the tests, but Tom and Andres said the stable-branch problems remained unresolved, with further recovery-conflict bugs also causing occasional failures on HEAD #35 #36 #37.
- The thread ends with Aleksei Fakeev reporting an IPC::Run broken pipe on arm64 under apparent load after successful deadlock detection in PostgreSQL 15; no resolution or re-enablement of the older branches is recorded #38.