model-writtenrun 6190 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Dec 2019 · 3 entries
- Nathan Bossart reported that a crash could leave an archived WAL segment containing the beginning of an incomplete record that recovery subsequently overwrote, and supplied a reproducer plus an initial delayed-notification patch in #1.
- After initially questioning the report, Kyotaro Horiguchi accepted the archiving bug and showed that WAL-buffer eviction could still write incomplete records with the revised patch, ruling out merely moving notification to the end of XLogWrite() in #5.
- Nathan found parsing WAL metadata too complicated and rejected a proposed complete-record flag as vulnerable to indefinite notification delays, then moved notification into a WAL-writer proof of concept that compared insert and flush positions in #10, #12, and #13.
May - Jul 2020 · 2 entries
- Ryo Matsumura objected to polling delays and shutdown synchronization in #14, then supplied an alternative that recorded crossing-record endpoints in shared memory and moved standby notification into replay in #16.
- Kyotaro simplified Ryo's design to track the latest crossing record and restrict what walsender could transmit in #17, but Ryo demonstrated that multiple buffered crossing records could invalidate its assumptions in #18 and #22.
Sep - Oct 2020 · 1 entry
- Heikki Linnakangas questioned exposing flush positions inside records in #24, and Kyotaro revised his implementation to extend writes and defer unsafe flush positions while leaving cascaded-replication concerns unresolved in #25.
Nov 2020 - Mar 2021 · 3 entries
- Nathan challenged record-size and flushing assumptions and introduced a shared hash table of crossing-record endpoints in #27, while Kyotaro continued a competing bounded-range approach through #28, #29, and #32.
- Andrey Borodin reviewed lock usage in Kyotaro's version in #34, while Nathan and Kyotaro debated hash-table memory consumption and time spent holding WALWriteLock before Nathan rebased his own branch in #35, #37, and #39.
- Nathan expanded the hash-table branch into a two-patch series with persisted notification progress in #40, explicitly leaving streaming-replication repair for separate work and deciding that archive_timeout required no change, then rebased the series in #41.
Jul - Sep 2021 · 16 entries
- When Alvaro Herrera resumed review, he questioned the persistence patch's suitability for backpatching; Nathan's crash tests showed that existing checkpoint cleanup recreated missing status files, so that patch was abandoned in #44 and #45.
- Alvaro took over revisions of the remaining patch, batching archiver wakeups and experimenting with lock ownership before restoring separate spinlock and LWLock responsibilities in #48 and #53.
- Kyotaro's reviews led Nathan to publish notification progress earlier, initialize it during startup, and move file creation outside the map lock across #59, #62, and #64.
- Alvaro's naming and API review produced explicit segment-boundary terminology and simpler lookup interfaces in #66, followed by explicit flush-position arguments in #71.
- Alvaro argued against creating status files during individual record insertion, and the discussion settled on XLogWrite() as the primary notifier with a WAL-writer backstop for registration races, implemented in #81.
- Alvaro and Nathan refined the wakeup comparison, ensured registration preceded refreshing the local flush position, and reorganized the implementation across #82, #83, and #84.
- Robert Haas and Alvaro challenged the shared hash table's bounds and cost in #85, #86, and #92, prompting Nathan to replace it with earliest/latest boundary slots in #93.
- Alvaro isolated the new state behind a dedicated spinlock in #95 and folded single-caller helpers into the notification routine in #99.
- Alvaro briefly suspected that segments without continuation records would never be notified, but Nathan explained exact-boundary registration and Alvaro withdrew the concern in #102 and #104.
- Alvaro committed #99 as 515e3d84a0 with naming and comment edits, credited Nathan as author, and announced backpatching to all live branches in #106.
- Nathan and Alvaro agreed to move the streaming-replication fix into another thread in #107 and #108, while Fujii Masao's post-commit question reopened discussion of the WAL-writer race backstop in #109 and #110.
- Andres Freund raised concerns about atomic operations, divisions, lock duration, single-user mode, and the separation from streaming-replication repair in #111 and #114.
- Nathan proposed registering boundaries before releasing insertion locks and removing the backstop in #118, but Andres objected to extending that critical section and identified a case where a partially filled page could leave archiving delayed indefinitely in #119.
- After Nathan found an unguarded notification call when archiving was disabled in #120, Andres clarified that background WAL generation need not rescue an unflushed partial page in #121 and #123.
- Nathan supplied a second follow-up that made background flushing advance to the earliest pending boundary and guarded the disabled-archiving case in #124, while acknowledging that single-user mode still posed a problem.
- Andres requested reverting the change and Alvaro agreed that a different approach was needed in #125 and #126; after Michael Paquier also reported a Windows warning in #127, Alvaro ended the thread inclined to revert once the way forward was clearer in #128, with no completed revert or follow-up landing shown.