model-writtenrun 8590 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Mar - Apr 2020 · 21 entries
- Jehan-Guillaume de Rorthais reports that primary crash recovery can discard unarchived WAL and proposes using StandbyModeRequested in XLogArchiveCheckDone() in #1; he later identifies Benoît Lobréau as the original reporter in #40.
- Fujii Masao explains that StandbyModeRequested is process-local and therefore unsafe for callers outside the startup process in #2, leading Jehan-Guillaume to replace the recovery boolean with a shared three-state enum and add a regression test in #3.
- Kyotaro Horiguchi identifies standby behavior problems in #4, while Fujii explores preservation requirements for snapshot backups and the distinction between archive recovery and standby mode in #6 and #8.
- Fujii proposes a separate shared standby-request flag in #17, but Jehan-Guillaume retains the recovery-state enum to avoid multiplying related booleans and revises the cleanup condition in #18.
- Fujii's review of initialization, naming, and recovery transitions in #19 drives the next revision #20, which updates shared state when crash recovery transitions into archive recovery.
- After Michael Paquier questions the placement of the regression test in #21, Jehan-Guillaume creates a dedicated archive-status test covering primary and standby behavior in #27.
- Michael's detailed review in #30 calls for locked shared-state reads, simpler initialization, consistent naming, and additional standby crash tests; Jehan-Guillaume incorporates these changes and expands coverage in #31.
- Michael takes over patch preparation in #33 and #35, consolidating the series, simplifying backend code, refactoring test segment names, and fixing a missing shared-state update at the start of REDO.
- Kyotaro accepts temporary .ready creation during standby crash recovery because archive recovery subsequently permits cleanup in #38, settling the immediate fix without adding more distinctions between requested recovery modes.
- Michael uses a deliberately failing copy command for portable archive-failure testing in #39, declining Kyotaro's proposed archiver error-handling change because it would alter unrelated behavior.
- The proposed reuse of ControlFile state is deferred after Michael explains the risks of depending on DBState ordering in #43, and Jehan-Guillaume recognizes that broader shared control-file work belongs in another discussion in #45.
- Jehan-Guillaume notices the unnecessary Windows skip in #46; Michael verifies Windows execution in #48 and supplies mainline and backport versions in #49, identifying the older branches' two-checkpoint recycling behavior.
- Jehan-Guillaume replaces the weak older-branch assertion with a two-restartpoint cleanup test in #51 and removes an empty configuration call in #52; Michael and Kyotaro then identify replay-comparison and interpolation issues in #53 and #54, addressed by #55.
- Michael reports applying the finalized series in #56, producing 4e87c4836a with the shared recovery-state fix and archive-status tests, plus branch-specific test adaptations and minor editorial changes.
- After the landing, Andres Freund questions whether process identity could replace shared state in #57; Michael explains in #60 that backup-history cleanup can call XLogArchiveCheckDone() from another process.
- Tom Lane reports buildfarm failures in #58, and Michael removes three unstable test assertions in #59 after suspecting that restartpoints run before standbys restore the required WAL segments.
- Jehan-Guillaume posts a delta restoring those assertions with explicit replay waits in #61, preserving the additional coverage while addressing the race.
- Michael reproduces the race with deliberately delayed archive restoration and verifies the patch in #62; Kyotaro confirms the restartpoint reasoning and requests explanatory comments in #63.
- Michael commits ebf6de8692 from #61, adding comments and using the LSN returned by pg_switch_wal() as the replay target, as explained in #65; the restored tests are limited to HEAD to avoid destabilizing stable branches.
- The thread ends with Michael reporting a quiet buildfarm in #67, leaving the backend fix landed across supported branches and the repaired additional coverage restored on HEAD.
- Later, 47a3a1c3d4 carries over the missed failed-archive test-label interpolation correction associated with #54 and #55, which had previously reached only PostgreSQL 10 and older.