model-writtenrun 8686 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jul 2015 · 6 entries
- Martijn van Oosterhout reported a broken index after crash recovery following CREATE TABLE and TRUNCATE in one transaction under wal_level=minimal #1, and Fujii Masao reproduced the failure #4.
- Tom Lane rejected rebuilding indexes during recovery because it would require unsafe catalog access #10, while Andres Freund explained that replaying truncation could also erase committed rows loaded by WAL-skipping COPY #11.
- Tom demonstrated another failure in which an INSERT initialization record erased subsequent COPY data #21; Andres favored new relfilenodes and fresh COPY pages #26, while Tom sought to preserve efficient repeated truncation #27.
- Heikki Linnakangas explained why triggers, full-page images, and hint-bit writes defeated narrow fixes #28, and proposed skipping all operations on new relations before choosing fsync or whole-file WAL at commit #33.
- Simon Riggs implemented two narrow alternatives #38, but Heikki's tests broke both, and his replacement tracked per-page WAL boundaries in a pending-sync hash and deferred synchronization until commit #39.
- Simon and Robert Haas questioned the cost of per-operation checks #40 #41; after a long pause, Heikki completed and rebased the prototype in February 2016 #45.
Feb - Apr 2016 · 1 entry
- Michael Paquier's testing showed that empty indexes still failed #47, and Kyotaro Horiguchi proposed an independent ambuildempty-based repair #49, which was set aside after objections that it was an incomplete, difficult-to-backpatch workaround #50 #51.
Jul 2016 · 1 entry
- Michael revived the broader fix in July 2016, first rebasing the wrong prototype #58, then correcting the base and adding TAP tests for INSERT/COPY, triggers, and TRUNCATE/COPY #59.
Sep - Dec 2016 · 1 entry
- Kyotaro moved pending-sync boundaries directly into RelationData to avoid repeated hash lookups #62, but Alvaro Herrera's review #71 and Tom's explanation of relcache invalidation #74 required restoring authoritative state outside relcache.
Jan - Apr 2017 · 1 entry
- Kyotaro restored the backend hash while caching pointers and negative lookups in RelationData #76 #77, then corrected an incomplete generated patch after Thomas Munro's compile test #85 #88 and repaired recovery-time debug access to InvalidBuffer #90.
Jul 2018 · 3 entries
- In July 2018, Michael restarted the narrow backpatch approach using Simon's COPY heuristic and selective disabling of in-place TRUNCATE #94, but Heikki showed that triggers, constraints, and datatype input functions could still cause unsafe interleaved WAL writes #97 #100.
- Kyotaro tried dynamically disabling COPY optimization after a logged insertion #106, while Robert favored preserving performance through a substantive fix #110 and Heikki recommended returning to the broader implementation #111.
- Andrew Dunstan reviewed that implementation's recovery behavior #114; Kyotaro removed the questioned recovery leniency, imported Michael's tests, added empty-btree initialization WAL, and split infrastructure from callers for review #115 #116.
Oct 2018 - Apr 2020 · 23 entries
- After further rebases, Noah Misch's tests exposed temporary-index WAL, lost pending syncs after subtransaction rollback, and UPDATEs spanning logged and unlogged pages #122; Kyotaro supplied fixups and expanded the series to cover CLUSTER, SET TABLESPACE, and btree operations #123 #125 #126.
- Noah identified further subtransaction and btree-split hazards #127, leading to revised table-AM interfaces and withdrawal of btree per-buffer integration #128, while Robert's review prompted additional UPDATE metadata assertions and corrections #129 #132.
- Noah then rejected the fragility of synthesizing INSERT and DELETE records for UPDATE and advocated an all-or-nothing WAL decision for each new relfilenode #135; Kyotaro implemented that redesign and discarded the special empty-btree fix #136.
- The new design gained precommit synchronization and reliable rd_firstRelfilenodeSubid tracking for CLUSTER and materialized views #137 #138, then moved ownership from access methods and relcache into storage.c at Noah's request #139 #143 #145.
- Noah's next test delta exposed unreliable state after nested TRUNCATE rollback and highlighted two-phase durability and all-fork synchronization requirements #151; Kyotaro repaired those paths, with Noah clarifying that synchronization must happen at PREPARE rather than COMMIT PREPARED #155 #156.
- Kyotaro separated pending syncs from pending deletes and implemented the size-based choice between whole-file WAL and fsync #161; Noah then exposed hint-bit WAL violations and proposed the wal_skip_threshold name #164, which the next revisions adopted #165 #167.
- Robert reviewed locking and subtransaction-field semantics #168, and Kyotaro revised those details #170 before Noah consolidated the series, documented its coding rules, and repaired further storage and relcache inconsistencies #171.
- Noah's consolidated version also revealed that GiST needed distinct real LSNs and that syncing many relations should avoid repeated shared-buffer scans #171; Kyotaro supplied separate GiST and bulk-sync proposals #174 #176.
- Noah explained that reusing the current insert LSN was insufficient and that truncated files could retain stale disk contents despite commit-time WAL #179, leading to a new GiST assignment record, truncation tracking, and combined bulk synchronization #185 #186.
- Review extended synchronization to inactive file segments #191 #192, while benchmarking informed the threshold default, which Noah ultimately set to 2 MB #197.
- Noah's stronger tests found lost WAL-skipping state after DROP followed by rollback, index-build reuse, and shrinking non-main forks #197; Kyotaro's January revisions preserved dropped relcache entries, transferred reused-index state, and required fsync after any truncation #202 #203 #204.
- Further review corrected dropped-state subtransaction handling and omitted field maintenance #206 #211, and Thomas's report of a patched regression failure #212 led Noah to remove an obsolete heap_create change and separate standalone prerequisites #215.
- Kyotaro reviewed the resulting documentation and tests #218, Noah posted the revised full series #223, and Kyotaro's supported-branch ports exposed additional recovery-test and multi-page WAL infrastructure requirements #225.
- Noah chose to backport the applicable recovery suite rather than only this patch's test infrastructure #227, producing 12034da6cc.
- The standalone TOAST-lock retention and log_newpage_range argument fixes from #223 landed unchanged as e629a01f69 and d3e572855b.
- For older branches, Noah introduced XLOG_FPI_MULTI so an old server would fail rather than silently replay only the first page #227, and the supporting backport landed as ae86e46c3b.
- The main member of #223 first landed as cb2fd7eac2, incorporating later recovery-test repairs, index-state restoration changes, and GUC integration described in #225 and #227.
- Buildfarm failures exposed parallel workers lacking pending-sync and relfilenode-newness state, plus a test threshold invalid on 32-bit platforms #228 #230; Noah supplied diagnostic instrumentation and reverted the main change with de9396326e, leaving the prerequisites in place.
- Kyotaro restored the main patch with serialized pending-sync state for workers and a smaller test threshold #231; Noah moved that transfer into parallel.c and prevented workers from performing redundant commit-time synchronization #232.
- Michael and Andres urged gaining development-branch experience before deploying the invasive fix to stable branches #236 #238, Noah abandoned the backpatch plan #239, and Robert explained the additional risk from changing WAL compatibility #240.
- The corrected series from #231 landed on the development branch as c6b92041d3, including Noah's worker-handling revisions and an XLOG_PAGE_MAGIC bump.
- The buildfarm then exposed a missing perl2host conversion in the new recovery test #242, which Noah fixed in 70de4e950c.
- Noah closed the commitfest entry at the thread's end, while recording an unreproduced catchup timeout with earlier occurrences on the same machine #244; the main fix was landed for the development branch, with stable-branch backporting deferred.