model-writtenrun 3316 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Dec 2022 - Apr 2023 · 23 entries
- Nathan Bossart proposed loadable recovery modules in #1, pairing shell-command refactoring with three library settings and a basic_archive implementation intended to improve robustness and performance.
- Andres Freund argued for one configuration setting and runtime library switching in #2 and #4, with Michael Paquier supporting reloadability in #6; Nathan implemented a single SIGHUP-reloadable restore_library and shutdown callback in #8.
- Frontend compatibility remained disputed: Nathan left pg_rewind support unresolved in #1, while Andres stressed its importance for high-availability recovery performance in #7.
- Nathan fixed Windows path escaping in the test in #10, rebased onto shared placeholder expansion in #12, and extracted duplicated restartpoint filename construction in #14 following Michael's review.
- Michael committed the shell_restore.c extraction from #14 as 9a740f81eb, first correcting failure logging that incorrectly ran after successful commands, as explained in #15.
- Michael then proposed eliminating the now-thin BuildRestoreCommand() wrapper in #17; Nathan implemented this in #18, and Michael committed that component as 14bdb3f13d while retaining defensive NULL-placeholder checks, leaving the module feature in #20.
- Michael's review in #21 requested separate archive and recovery documentation sections and exposed checkpointer failure loops caused by conflicting command/library settings; Nathan split the series into independently reviewable pieces in #22.
- Michael committed the basic_archive-to-basic_wal_module rename from #22 as 0ad3c60caf, then forwarded the remaining patches with further documentation edits in #23.
- Nathan proposed warning and disabling cleanup callbacks instead of trapping the checkpointer in error loops in #24, and considered dropping cleanup and recovery-end support in #25; Andres argued these operations were essential to useful modules in #27.
- Andres challenged the callback API in #28 and #31, advocating constant callback tables, per-instance state, clearer naming, and dedicated headers so future implementations could support multiple modules.
- Nathan consequently opened a separate archive-module redesign series in #33; Michael questioned its double-pointer initialization interface in #34, and Andres required allocation to move into a startup callback in #36, producing #39.
- Further review introduced typed ArchiveModuleState and BasicArchiveData structures in #42, followed by state-first callback arguments, explicit long-lived allocation, include fixes, and documentation markup corrections through #48.
- Andres's review in #50 exposed redundant function-type assertions; Nathan added their removal in #53 and moved it ahead of the redesign in #55, allowing Andres to commit that component unchanged as 30b789eafe.
- After the assertion cleanup landed, Nathan resent the remaining redesign in #58 and followed Andres's review in #59 by squashing it, clarifying callback-table lifetime, and adding explicit shutdown cleanup in #60.
- Michael questioned shutdown memory-context switching and header placement in #62; Nathan moved the implementation into archive/ and temporarily removed cleanup in #63, but restored it with an assertion in #65 after Andres warned that omitting cleanup would entrench a single-module design in #64.
- Andres's final review in #69 led to small defensive cleanup changes in #70, which Michael committed as 35739b87dc with only formatting, wording, and typedef-list edits.
- The archive redesign also exposed concerns about module-local exception handling and misuse of AtEOSubXact_Files() in #71; Michael requested separate discussions in #73, and Nathan opened one in #75, leaving those issues outside the redesign commit.
- Later patch contents show that the early basic_wal_module rename and shell_restore.c extraction were no longer present in the restore series' base: #76 again extends basic_archive and extracts shell execution from xlogarchive.c, while mentioning a recent SIGTERM mistake whose full reversal history is not recorded in the supplied discussion.
- Nathan resumed the restore-library proposal in #76 using the new archive API pattern, separate WAL-segment and timeline callbacks, typed state, and restore/ files; checkpointer misconfiguration now warns and skips cleanup instead of blocking checkpoints or repeatedly restarting the server.
- The revised proposal in #76 retained synchronous one-file-at-a-time restoration, explicitly leaving asynchronous, parallel, and batched restoration for future work.
- Nathan refreshed the externally tracked SIGTERM prerequisites and made restore headers installable in #77, preserving the distinction between startup-process cleanup and immediate exit in a child forked by system().
- An intermittent recovery-test failure led Nathan to wait for WAL replay with poll_query_until() in #78; the accompanying discussion exposed confusing pg_ctl readiness behavior with hot_standby disabled, but Michael favored considering broader simplification in #79.
- Michael deferred the remaining feature to PostgreSQL 17 in #80, and Nathan continued maintaining it with #81.
Oct 2023 · 1 entry
- After a six-month gap, #82 reduced the series from seven patches to five because its two SIGTERM prerequisites were already present in the new base, leaving the GUC helper, shell refactoring, documentation changes, and restore_library feature pending.
Jan 2024 · 1 entry
- Rebases continued through #83 and #84, where Nathan explicitly deferred the sizeable, insufficiently reviewed proposal to PostgreSQL 18; subsequent submissions through #89 primarily tracked upstream changes.
Jul 2024 · 1 entry
- The thread ends with Nathan announcing his intention to withdraw the commitfest entry for lack of interest in #90; the restore-library feature remained uncommitted despite the earlier preparatory and archive API landings.