model-writtenrun 7753 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Aug - Oct 2019 · 3 entries
- Magnus Hagander restarted his and Daniel Gustafsson's online-checksum proposal with a background-worker implementation layered over Andres Freund's experimental global barriers, attaching both components for testing in #1.
- Bruce Momjian asked whether standby switchover could replace in-place conversion in #5, but Magnus maintained that in-place switching was needed in #6.
- Tomas Vondra considered the barrier approach sufficient in #4, while Andres cautioned that the prerequisite itself had not received meaningful review in #9.
Dec 2019 - Jan 2020 · 3 entries
- Daniel took over posting revisions, first rebasing the combined series in #12 and then adapting it to Robert Haas's ProcSignalBarrier implementation in #16; the prerequisite's separate landing was reported in #17.
- Robert's detailed review in #18 identified unlocked mutable state, unsafe disable ordering, fragile worker retries and missing guarantees against checksum-state changes during I/O; Daniel addressed retries and diagnostics in #19 while continuing the concurrency work.
- Robert proposed persistent catalog progress in #20 and explained its transition-only meaning in #26, while Magnus questioned the cost for an infrequent operation in #21 and Andres preferred first establishing correctness without catalog state in #27.
Apr 2020 · 1 entry
- After the patch was returned with feedback in #29, Daniel revived it in #30 with backend-local state, revised synchronization, renamed workers, pg_class progress tracking and restart tests.
Jun 2020 - Mar 2021 · 15 entries
- Justin Pryzby's review in #35 and Robert's further synchronization review in #36 led to clearer diagnostics, Boolean SQL results and worker cleanup in #39, with deeper concurrency issues still under investigation.
- Daniel found that immediately disabling checksum writes could break readers still validating checksums, so #41 added an inprogress-off state that stops verification before writes cease and also repaired standby-test timing.
- Justin proposed per-relation checksum control in #43, but Daniel kept the feature cluster-wide in #44, explaining that catalog flags tracked conversion progress rather than authoritative checksum settings.
- Heikki Linnakangas identified unsafe state changes around I/O and races between SQL calls and launcher decisions in #47, then clarified that base backup needed per-read checking rather than holding interrupts throughout file transmission in #50.
- Daniel responded in #54 with interrupt protection and target-state requests to worker processing; Heikki's subsequent reviews in #55 and #57 prompted superuser checks, corrected WAL ordering, critical sections and checkpoint delays through #58.
- The January restart of work brought Michael Banck's documentation and code reviews in #60 and #61, followed by clearer state-transition explanations and naming in #63.
- Michael and Magnus supported extracting generally useful documentation in #65 and #66, and Daniel split it into a prerequisite patch in #67.
- Daniel added offline-after-interruption tests in #68 and recorded future optimization ideas in comments in #69, leaving lightweight WAL records, avoiding unnecessary dirtying and automatic restart for later work.
- Magnus committed the documentation component of #69 as cf621d9d84, changing a few terminology references as explained in #70; Daniel then resent only the remaining feature in #72.
- Heikki's review in #73 exposed unsafe signal-handler locking and inconsistent SQL return values, leading Daniel to fix those issues and a late-cancellation race in #75, followed by the catalog-tuple-copy correction in #77.
- Heikki supplied a substantial launcher rewrite in #78 that made SQL callers publish desired state and let the launcher decide operations and handle changed requests; Daniel supplied restart and interrupt fixes alongside it in #79, then squashed the series in #80.
- Heikki's review in #81 found that interrupt protection did not span the actual storage operations, while he and Michael Paquier questioned the feature's complexity in #81 and #82.
- Magnus attributed much of the added complexity to persistent progress tracking in #84, and Daniel removed that machinery while correcting I/O interrupt boundaries in #88, so interrupted conversions would start again from the beginning.
- At the thread's close, the documentation had landed but the online feature had not: Daniel marked the proposal withdrawn as effectively rejected in #91, directing standby-switchover discussion to the separate conversation begun by Michael Banck in #83.
- Years after the recorded discussion ended, f19c0eccae landed a substantially revised descendant of #88, retaining barrier-mediated online switching without persistent relation progress while adding reorganized state handling, progress reporting and extensive tests; the intervening development is absent from this thread.