model-writtenrun 1314 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Heikki Linnakangas proposed consolidating background-worker and regular-backend exit cleanup because every running background worker already had a BackendList entry, making separate searches and cleanup functions redundant #1.
- After a conflict-only rebase #2, Heikki committed the unchanged comment correction as ef4c35b416 and allocation-helper refactor as 63bef4df97, announcing both in #3.
- Heikki expanded the remaining work into an eleven-patch series covering background-process launching, connection-limit tests, dead-end shutdown, shared-memory exit callbacks and PMChild slots for auxiliary processes, with separate pools intended to prevent client connections from starving workers #3.
- Thomas Munro endorsed the initial refactoring, preserving meaningful Windows exit statuses and identifying unexpected children explicitly in logs; making unexpected-child handling fatal was left for separate consideration #4.
- The doubly linked list and restart-comment patches landed unchanged as b43100fa71 and a79ed10e6c, while unified cleanup landed as 28a520c0b7 after Heikki added duplicate-log suppression for Windows child exits #3 #5.
- The background-launch consolidation from #3 landed as 3354f85284 after comment cleanup, and Heikki resent seven remaining patches while raising the risk of confusing backend-type values with bitmasks #6.
- Alexander Lakhin reported binary garbage in crash logs and traced it to an uninitialized process-name buffer #7; Heikki corrected the argument passed to HandleChildCrash() in 56d23855c8 and confirmed the fix #8.
- Andres Freund reviewed the remaining series, requesting clearer exit-hook naming, safer bitmasks, less duplicated pool configuration and portable test timeouts #9; subsequent review identified incorrect pool sharing and potentially omitted autovacuum shutdown signaling #13 #15.
- Heikki deferred unifying PMChildSlots with ProcNumbers and replacing child-list searches with a hash table to keep the series manageable #11, while proposals to eliminate dead-end processes raised the problem of blocking client or SSL writes in the postmaster #13 #14 #16.
- The test branch was revised to exchange SSLRequest packets before opening each additional raw connection, avoiding socket accept-queue races #17; Thomas explained the platform differences, and Heikki corrected the affected platform to macOS #18 #19.
- After Dagfinn Ilmari Mannsåker explained newer Perl socket support #20, Heikki added a runtime capability check and skip path #22; Nazir Bilal Yavuz later reported that updating Windows CI's Perl allowed the test to pass #23.
- The final test attachments landed unchanged as 6a1d0d470e and 85ec945b78 from #22, while the exit-callback patch from #6 landed as 2bbc261ddb with Andres's requested registration-oriented naming.
- Heikki then posted four remaining patches with a distinct BackendTypeMask type, explicit per-type child pools, separate background-worker allocation, singleton auxiliary pools and consolidated shutdown-state handling, incorporating Andres's review #24.
- The backend-type conversion and dead-end termination changes from #24 landed as 18d67a8d7d and bb861414fe with comment edits.
- The child-slot implementation from #24 landed as a78af04270 after bootstrap and single-user initialization fixes, helper inlining and cleanup, followed by explicit slot passing as 5b00786857 with an added parameter comment; Heikki announced completion of the original series #25.
- Tomas Vondra later reported connection-limit test failures under Valgrind #26 and supplied a separate pg_ctl/pg_regress instrumentation patch to reproduce them #28.
- Heikki reproduced a race in which a rejected connection had reported its error but still occupied its PGPROC slot #29; pg_stat_activity could not observe that early failure, and possible sleeps or backend error-handling changes remained unresolved #30 #31.
- When the failure appeared on the buildfarm, Andres proposed exposing psql diagnostics and waiting for the postmaster's backend-exit log message #32 #33; Michael Paquier suggested injection points, while Noah Misch explained why pumpable() might still race with process exit #34 #35.
- Andres posted four test fixes implementing log-based synchronization, improved diagnostics and a 600-second authentication-timeout cap, leaving broader pump_until() changes and injection-point generalization aside #36.
- Heikki approved all four fixes and supplied an independent test-label and username cleanup #38; Andres then reposted identical code with completed commit descriptions #39.
- Tomas confirmed that the revised series fixed his reproducer #42, and its unchanged race fix and timeout cap landed as 71d1ed6fe1 and b48832cddb from #39.
- Heikki's independent cleanup from #38 subsequently landed as 2a943afcff, preserving both changed lines while rebasing them over the newly introduced synchronization helper.
- Michael Banck revived the thread with a Hurd failure caused by Perl's send() path requiring unsupported Unix-socket peer lookup, proposing an expanded raw_connect_works() probe and skip condition in duplicate submissions #44 #45.
- Heikki requested an explanatory comment and questioned sending arbitrary data as the probe #46, while Michael explained the Perl call path #48; the final reports say Debian's workaround passes by skipping affected tests, but ask whether the broader skips are intended and how to improve the probe, leaving that upstream follow-up unresolved #49 #50.