model-writtenrun 1074 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Fujii Masao opened the thread with a SIGTERM patch to prevent a slotsync worker stuck waiting for the primary from delaying promotion, also guarding against replacement workers starting during promotion #1.
- Tom Lane raised shutdown-semantics concerns #2, while Amit Kapila explained that the SQL synchronization backend should report an error and remain connected, making unconditional SIGTERM unsuitable #4 #9.
- Fujii supplied a network-drop reproducer #8, Nisha Moond confirmed the hang and explored TCP timeouts #10, and Fujii argued for a direct fix because timeout tuning trades failover delay against false failure detection and lacks uniform platform support #11.
- Amit proposed a dedicated procsignal reason that sets interrupt flags while retaining SIGUSR1 #12, and Nisha implemented and tested that design for both workers and SQL callers in #13.
- Shveta Malik and Amit reviewed redundant shutdown checks and interrupt-handler structure #14 #17 #18, leading Nisha to remove the old wrapper and consolidate shutdown handling through ProcessInterrupts() in #15 and #19.
- Fujii requested clearer naming, shared-state documentation, and startup-refusal messages #20, while Shveta examined why shared stopSignaled state and a local pending flag were both needed #22; Nisha incorporated these refinements in #21 and #24.
- Fujii's tests revealed duplicate promotion logs #23, and discussion settled on lowering the replacement worker's refusal message to DEBUG1 instead of adding shared-memory checks to the postmaster #25 #28.
- Zhijie Hou and Shveta identified the need to ignore shutdown requests after SQL synchronization had already finished #26 #27, and Nisha added that guard together with the logging refinement in #30.
- Fujii requested final comment cleanup #32, Nisha supplied it in #33, and Fujii reposted the same code with a revised commit message and prepared stable-branch variants in #34.
- Amit questioned whether the prerequisite SQL-backend signaling change needed backpatching #35, but Fujii favored consistent shutdown behavior across branches #36; after Amit found no concrete safety problem #37, Fujii landed db93032a7c from #34 without substantive code edits and confirmed backpatching through PostgreSQL 17 #38.
- The backports changed existing ProcSignalReason values, so Fujii reported an ABI problem #39 and posted PostgreSQL 17 and 18 repairs that moved the new signal reason after existing values #40.
- Nisha checked extension usage and argued that PostgreSQL 17 could retain NUM_PROCSIGNALS as its final enum member #41; Fujii accepted this at the thread's last message, leaving the repairs awaiting application at that point #42.
- The supplied commit record shows that the ABI repair subsequently landed as acf49bfede from #40, with the PostgreSQL 17 variant following the adjustment agreed in #42.
- The remaining increase in NUM_PROCSIGNALS was subsequently recorded as an accepted compatibility exception by 29d8bd9085, implementing the .abi-compliance-history follow-up anticipated in #40 using the rationale settled in #42.