model-writtenrun 6347 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Bharath Rupireddy opened the thread after observing that children sleeping in pg_usleep(), including an autovacuum launcher using PostAuthDelay, could remain after postmaster death #1.
- Thomas Munro recommended latch infrastructure for long waits and polling loops, explaining that CHECK_FOR_INTERRUPTS() did not detect postmaster death and that signal-only handling leaves races; broader interrupt-based death detection remained a separate design question #2.
- Bharath surveyed additional sleep sites #3 and submitted three patches covering vacuum truncation, backup completion, and authentication delays, reporting passing regression tests #4; Aleksander Alekseev subsequently reported successful testing on macOS #6.
- Michael Paquier supported the vacuum change but challenged the backup patch's use of the startup process's recoveryWakeupLatch and questioned changing developer authentication delays #7; Bharath initially accepted leaving those delays alone #8, while Tom Lane argued that surviving children interfere with subsequent debugging #9.
- Michael committed the vacuum component of #4 as 70685385d7, preserving its logic while renaming the wait event to VacuumTruncate and refining its documentation; his announcement also clarified that backup completion runs in backends or WAL senders, not the startup process #11.
- Michael supplied a corrected backup patch that always waits on MyLatch, simplifying wait-event reporting and improving responsiveness to signals and postmaster death #13.
- Kyotaro Horiguchi remained neutral about changing developer delays but requested an explanation if pg_usleep() stayed #14; Stephen Frost endorsed the corrected backup patch and supported eliminating surviving children #15, and Bharath accepted the latch correction #16.
- Bharath resent the original authentication-delay patch together with Michael's corrected backup patch for CI #17; the backup component landed unchanged as 2aca19f298, with application confirmed in #18.
- Michael found that the authentication patch skipped PostAuthDelay because MyLatch was already set, proposed waiting only for timeout or postmaster death #18, and confirmed the failure through testing #20; Bharath removed latch-set handling and resets from those waits and corrected the PreAuthDelay timeout variable in #21.
- Michael requested consistent flags across authentication-delay paths but declined further review during that commitfest because the developer-option changes needed closer examination #22; Bharath applied the same timeout-and-death-only behavior to PreAuthDelay in #23.
- The thread ended with vacuum truncation and backup completion committed, while authentication-delay changes remained uncommitted; Michael marked the commitfest entry committed and recommended a new thread for separate evaluation of the remaining work #24.