model-writtenrun 4759 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Apr - Oct 2021 · 22 entries
- Robert Haas opened the thread because long data-directory synchronization, unlogged-relation reset and WAL replay could leave administrators unable to tell whether startup was progressing, proposing periodic logs and optional remote monitoring in #1.
- Tom Lane and Magnus Hagander favored logging because remote monitoring introduced authentication and complexity concerns, and Robert chose to implement logging first in #5, #6 and #7.
- Alternatives involving pg_control, special monitoring connections and recovery percentages remained discussion topics: Tom warned about additional control-file writes, while Andres Freund explained why recycled WAL makes progress estimates difficult in #11, #21 and #25.
- Nitin Jadhav supplied the first logging patch in #30, but Robert and Tom rejected checking the clock for every WAL record as too expensive; #34 instead used the existing timeout framework and a signal-set flag following #32 and #33.
- Justin Pryzby's reviews established distinct fsync/syncfs messages, a reloadable configuration setting and clearer operation codes, which Nitin incorporated across #34, #38 and #40 after #31, #37 and #39.
- Amul Sul reported test failures and confusing completion APIs, Bharath Rupireddy identified redundant helpers and message issues, and Robert requested a useful default interval and reduced timer drift in #41, #43 and #42; Nitin revised the implementation in #44.
- Justin identified misplaced initialization and the need to distinguish unlogged-relation cleanup from initialization in #46, leading to #47, then supplied an experimental function consolidation in #50.
- Robert argued that generic directory traversal should remain independent of startup logging in #49, and Nitin moved logging into the fsync callback in #52 without adopting Justin's proposed function consolidation.
- Robert and Bharath argued that single-user startup should also report progress in #53 and #54, prompting Nitin to remove startup-process-only guards and add timeout registration for that path in #56.
- Robert proposed placing message strings at their call sites behind a small reporting macro in #51, and Nitin replaced the centralized operation enums and switches with that design in #59.
- Justin found missing syncfs reporting and a zero-setting crash in #60, while Robert recommended absolute timeout scheduling, removing routine completion messages and moving the machinery into startup.c in #61; those changes shaped #64.
- Further review simplified phase initialization, made the signal flag safe, added pre-fsync reporting and configuration documentation, and reduced public header dependencies across #66, #68 and #73, following #65, #67 and #69.
- Justin reported endless standby progress messages in #70, and Robert distinguished standby operation from finite point-in-time recovery in #71; Nitin added StandbyMode guards while preserving single-user support in #79.
- Michael Paquier and Justin found inconsistent millisecond units and defaults in #75 and #78, which Nitin corrected in #80.
- Robert's testing exposed broken disabling semantics in #81 and #84; after an incomplete revision in #82, Nitin added an explicit zero-value guard and incorporated clearer documentation and Álvaro Herrera's translation and naming corrections in #89.
- Álvaro questioned drift compensation in #90, while Tom supported it but requested a reusable facility inside timeout.c in #96 and #99; Robert supplied that separate prerequisite and a temporary autovacuum testing patch in #100.
- Robert and Tom agreed not to make a broad clock-handling rewrite a prerequisite for startup logging in #101 and #102, and Robert then added an explicit first-firing timestamp to the repeating-timeout API in #106.
- Amul questioned the reporting macro's implicit formatting arguments in #104, but Robert found no clearly better implementation in #105; the macro remained in the final submission.
- Nitin combined Robert's prerequisite with the revised startup-reporting patch in #107, replacing private rescheduling logic with enable_timeout_every() and consolidating registration in StartupXLOG(); the temporary autovacuum testing patch was omitted.
- Robert committed the two unchanged components of #107 separately as 732e6677a6 and 9ce346eabf, then announced completion in #108.
- Bharath proposed storing additional recovery timing summaries in #109, but Robert treated that as separate work requiring its own proposal in #110.
- Robert observed an immediate progress message at the start of a new phase and supplied an inline fix in #111; after Nitin's review and agreement in #112 and #114, 5ccceb2946 disabled the old timer before clearing its flag and starting the next phase.
Nov 2022 · 5 entries
- More than a year later, Thomas Munro reported periodic SIGALRM delivery to otherwise idle standbys in #116, and Bharath began a new fix series in #117 to disable a timer whose progress messages were already suppressed.
- Simon Riggs initially favored disabling only while waiting, but Robert clarified the existing standby logging policy and endorsed encapsulated enable/disable helpers in #118 and #119; Bharath adopted the helpers in #120.
- Robert objected to disabling the timer for every WAL record in #122, so Bharath handled entry into the redo loop and later transitions into standby mode explicitly in #123.
- Robert proposed reactive disabling when a timer expiration encountered standby mode in #126, and Thomas approved that alternative in #127; Bharath instead centralized standby-mode transitions in EnableStandbyMode() in #128, citing per-record overhead and module coupling.
- Robert requested protection against accidental timer re-enabling in #129, while Kyotaro Horiguchi preferred the reactive approach for robustness in #131; Bharath tried a standby guard in #130 and then a narrower assertion in #133 after finding that the broad guard suppressed legitimate unlogged-relation progress.
Feb 2023 · 4 entries
- After a gap, Robert judged the assertion machinery too complicated in #134, and Bharath returned to the simpler EnableStandbyMode() version with comment changes in #135.
- Robert committed #135 unchanged as 8a2f783cc4 and backpatched it to PostgreSQL 15, announcing that result in #136.
- Tom objected to introducing the noncritical fix on release-wrap day in #137 and requested postponement in #139, so Robert reverted only the PostgreSQL 15 backport with 1eadfbdd7e; the development-branch fix remained.
- After Tom cleared reapplication in #142, Robert restored the unchanged PostgreSQL 15 fix with ecb01e6ebb, and Bharath confirmed restoration and closed the CommitFest entry in the thread's final message, #143.