model-writtenrun 9529 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Dec 2018 · 3 entries
- Andres Freund opened the thread with a minimal standby-decoding prototype that reused existing recovery-conflict WAL records plus a catalog-table flag, avoiding periodic catalog-xmin announcements and mandatory hot_standby_feedback #1.
- Robert Haas favored retaining invalid slots as evidence of a conflict, while Petr Jelinek warned that blocking replay behind a lagging logical consumer could stall physical replication; conflict policy remained an important design question #2 #4.
- Andres separated primary-side computation of page-level index-vacuum XID horizons because recovery could not safely inspect tables before consistency or identify arbitrary table access methods #3.
Mar - Jul 2019 · 8 entries
- Amit Khandekar took over implementation work, rebased both prototypes and adapted Craig Ringer's tests in #7; the horizon prerequisite subsequently landed as 558a9165e0, attributed to the later submission #421769 in thread #38991, which integrated table-AM support and Robert Haas's optimized prefetching work.
- Amit resent the remaining decoding work with conflicting-slot deletion and fixed a self-deadlock caused by reacquiring ReplicationSlotControlLock inside ReplicationSlotDropPtr #18.
- Reports of failures with insufficient primary wal_level led to checks during slot creation and decoding, plus removal of existing logical slots when replay encountered a lowering of wal_level #14 #26 #40 #41.
- Andres required standby slots to start at the replay position rather than the older redo position; Amit implemented valid-record and WAL-header handling, while upstream snapshot requests remained deferred because cascading replication complicated them #38 #49 #53 #60.
- Amit fixed apparent pg_recvlogical hangs by using the standby flush accessor instead of the primary-only accessor, which had incorrectly made the walsender sleep after each record #57.
- Review exposed unsafe handling of active slots, so Amit introduced cancellation signals, condition-variable waits, special handling for walsenders and recovery-conflict statistics #56 #59.
- Andres identified incorrect catalog checks on indexes and recursive catalog-access hazards; Amit began passing heap relations directly and documented the remaining constraints #66 #68.
- The work split into decoding, WAL metadata, conflict handling, tests and documentation, while Robert Haas drove simpler test helpers and explicit synchronization instead of sleeps #73 #77 #79 #83 #84.
Sep 2019 - Mar 2020 · 2 entries
- Rahila Syed's intermittent test failures exposed missing waits for hot-standby feedback, prompting corrected xmin checks and placement of wal_receiver_status_interval on the standby #90 #93 #94.
- After a long gap, Alvaro Herrera rebased the series, Fabrízio de Royes Mello revived it again, and Bertrand Drouvot restored missing conflict handling and repaired the test setup #101 #103 #105 #106.
Dec 2020 - Apr 2021 · 3 entries
- Bertrand corrected a bad ReplicationSlotReserveWal rebase and expanded statistics for inactive conflicting slots, with Fabrízio testing the work and supplying assertions, formatting and expected-output corrections #114 #115 #117 #119 #120.
- The inherited tests were replaced with checks for actual decoded output, invalidation and promotion, and a genuine row-removal conflict test was added because VACUUM FULL alone did not establish that behavior #126 #131 #142 #143.
- Andres traced an idle-primary delay to logical walsenders being awakened on WAL flush before replay advanced; Bertrand supplied an incremental replay-condition-variable fix, while testing continued to reveal slow slot creation #127 #146 #149 #150.
Jun - Oct 2021 · 2 entries
- Andres required prerequisites to precede feature enablement, found timeline initialization and slot-dropping races, and argued for invalidation rather than deletion; Bertrand implemented persistent invalidation in #130 following #124 and #128.
- Alvaro and Robert objected to opening table relations deep inside index operations, with Robert identifying deadlock risks under buffer locks; after another long pause, Bertrand tried propagating catalogness through pg_index instead #136 #152 #161.
Nov 2022 - Mar 2023 · 11 entries
- The full series returned with that pg_index design, updated timeline handling, active-slot-only conflict statistics and replay wakeups, followed by Windows TAP-process cleanup #163 #166.
- Robert's review improved the WAL layout and exposed hash and GiST alignment failures; Bertrand separately corrected misleading existing layout comments and introduced flexible arrays in the feature patches #168 #192 #193 #194 #195.
- Andres requested reuse of existing obsolete-slot invalidation machinery and centralized validity checks, leading Bertrand to merge the duplicated invalidation paths #198 #200.
- Andres rejected the pg_index propagation approach because of its complexity and interactions, so Bertrand replaced it with passing heap Relation through index APIs and vacuum state #205 #210.
- Bertrand fixed decoding across promotion, missing GiST catalog flags and stale xmin retention, while Bharath Rupireddy's monitoring feedback led to pg_replication_slots.conflicting #215 #221 #201 #224.
- Melanie Plageman improved relation-argument consistency and explanations, and Bertrand expanded cascading-standby tests and introduced pg_log_standby_snapshot to avoid expensive test checkpoints #216 #225 #228 #229.
- A proposal to require physical slots and hot_standby_feedback was rejected because those protections could not be guaranteed during archive recovery, leaving conflict detection necessary #231 #232 #233.
- Jeff Davis identified condition-variable queue and socket-wait problems and supplied a separate ConditionVariableEventSleep API experiment, later revising its timeout handling #240 #241 #252.
- During pre-commit review, Andres found that WAL-only catalog flags could corrupt neighboring visibility-map bits; the discussion produced the standalone assertion fix a88a18b125, while Bertrand isolated the WAL flags in the remaining series #251 #254.
- The table-relation plumbing was split from behavior changes in #254 and landed as 61b313e47e with only formatting and a stray-include removal; Bertrand then resent the remainder #259 #261.
- Amit Kapila and Jeff Davis proposed a simpler wakeup design that distinguished physical from logical walsenders, replacing the condition-variable experiment; Bertrand implemented it and reported that the tests no longer suffered the delay, including after promotion #255 #268 #274 #278.
Apr 2023 · 15 entries
- After the WAL-information prerequisite also landed, Bertrand removed it from the series and separated recovery-conflict handling from statistics and view changes #276.
- Subsequent review corrected effective-xmin conflict checks, translation of diagnostics, archive wakeups and replay wakeup guards, with Robert Haas requesting clearer explanations of the user-visible behavior #285 #286 #289 #293 #294 #320.
- Andres substantially rewrote the near-final series around explicit invalidation causes and centralized checks, fixing shared-catalog conflicts, effective_catalog_xmin handling and invalidated-slot retention; Bertrand supplied API and test-cleanup deltas #334 #335.
- The final full submission integrated those deltas, moved procsignal handling into the recovery-conflict component and forced WAL flushes after VACUUM in tests; its attachments could not be extracted because two tests shared a series number #337.
- A proposed subscription-on-standby test remained unfinished because its synchronization helpers invoked WAL-control functions unavailable during recovery #338 #341.
- Bertrand, Alvaro Herrera, Melanie Plageman and Amit Kapila reviewed the final rewrite, leading to clearer enum naming, switch handling and consistent error messages #340 #343 #344 #345 #346.
- The invalidation-enum prerequisite from #337 landed as 15f8203a59, including a slot-format bump and light naming edits; preserving restart_lsn during invalidation was explicitly left for later.
- The centralized CreateDecodingContext check from #337 landed separately as 4397abd0a2, removing duplicated validity checks with only light final edits.
- Generalized horizon and wal_level invalidation from #337 landed as be87200efd, incorporating the related pg_replication_slots.conflicting changes.
- Recovery-conflict integration from #337 landed as 26669757b6, incorporating conflict statistics and the required catalog and statistics-format changes.
- Selective physical and logical walsender wakeups from #337 landed as e101dfac3a with light comment and formatting edits.
- Andres announced that the feature series had been pushed, explaining that statistics and documentation components had been folded into the related commits and that he had corrected grammar and formatting #352.
- Noah Misch then reported hoverfly hanging because slot creation raced ahead of database replay; Bertrand added waits for both standby levels, and #357 landed unchanged as 57411c82ce #354 #356 #358.
- The same buildfarm report prompted lower test verbosity, and Tom Lane demonstrated an exact two-line deletion with smaller logs and faster execution #354 #359; the eventual 3034dc56ef is attributed to Bertrand's later matching attachment #533928 in thread #47736.
- At the last message, the feature and replication-wait fix were landed and the next buildfarm result was awaited; Noah agreed to reduce logging and left polling backoff, elapsed-time-based timeouts and fewer connections as future improvements #360.