model-writtenrun 5685 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Oct 2021 - Mar 2022 ยท 23 entries
- Alexander Lakhin reported a PostgreSQL 14 crash during concurrent catalog activity and VACUUM in #1, and Dmitry Dolgov reproduced it in #2.
- Testing ruled out an already-fixed parallel VACUUM bug, while stronger assertions exposed invalid heap references and pruning of non-heap-only tuples; Alexander and Dmitry reported these results in #6 and #7.
- Peter Geoghegan's initial inline guard prevented one assertion failure but left catalog errors, and he explained in #10 that Andres Freund had traced the underlying corruption to an LP_REDIRECT whose target pruning marked unused.
- Peter proposed traversing complete HOT chains before separately processing disconnected heap-only tuples in #11, then consolidated tracking into one array and reported successful stress testing in #12.
- Andres explained how repeated visibility checks could disagree in #13, and Peter argued in #14 for limiting the backpatch to PostgreSQL 14 because older versions had no demonstrated instance of this corruption.
- The investigation also exposed undocumented relationships between pruning horizons, OldestXmin, and freezing; Peter requested colocated initialization in #21, while Andres explained the required horizon relationship in #24 and #25.
- Andres's detailed review in #27 prompted clearer visitation bookkeeping and assertions in #33; Peter then made pruning conservative across non-DEAD chain members and tightened disconnected-tuple checks in #36.
- Andres questioned assumptions about concurrent subtransaction aborts in #37 and proposed caching each tuple's visibility result as a smaller backpatchable fix in #41, leaving the broader redesign for development.
- Dmitry's savepoint-and-rollback stress test found a new assertion failure in Peter's revised patch in #49; Peter diagnosed erroneous fallthrough from DEAD handling into a path requiring a valid update XID in #50.
- Andres justified the existing RECENTLY_DEAD chain rule in #51, and Peter accepted pursuing the minimal fix first and offered to review it in #52.
- Andres refined the reproducer to require horizon recomputation triggered by another tuple after catalog invalidation in #53, then attached an isolation test demonstrating incorrect index results in #54.
- Andres combined that test with the cached-visibility fix and whole-page redirect assertions in #57; Peter supported the fix and stronger checks in #58, while test reliability and corruption detection remained concerns.
- The December revision #60 selected the minimal fix, separated redirect assertions and block-number caching, and proposed a VACUUM horizon-initialization cleanup; Andres dropped the isolation test because its index-lock scheduling mechanism was too fragile without explicit wait points.
- Peter approved the pruning changes in #62 but challenged the placement of horizon initialization and its interaction with relation size in #63, while Dmitry reported successful testing of the fix and assertions in #64.
- Andres landed the cached-visibility fix as 18b87b201f and redirect assertions as bb42bfb5cc, each extracted from #60 with comment edits, and announced completion in #69.
- After Alexander noticed the stable-branch fix was missing in #70, Andres confirmed pushing it in #71; the separate block-number optimization from #60 subsequently landed unchanged as c702d656a2.
- For release-note recovery guidance, Tom Lane asked whether REINDEX was sufficient in #72; Andres described remaining uncertainty and limited HOT-corruption detection in #73, and Tom chose to recommend REINDEX in #74.
- Peter resumed the uncommitted VACUUM initialization work in #75, keeping setup in heap_vacuum_rel(), and separately revived his broader pruning design atop the committed visibility cache in #76.
- Andres reviewed both follow-ups in #77 and #78, suggesting optimizations and questioning pruning edge cases; Peter said in #79 that further pruning-hardening discussion would move to a dedicated thread.
- Andres requested separating the rel_pages rename in #80, and Peter's revised series #86 did so while also moving index pg_class updates out of lazy_scan_heap().
- The initialization and documentation component of #86 landed as 73f6ec3d3c with additional comment and truncation-condition edits, and its mechanical rename component landed unchanged as e370f100f0.
- Discussion continued over fresher initial horizons versus periodically refreshed pruning horizons in #87, #88, and #89, with Andres withdrawing his concern that refreshing vistest would itself break lazy_scan_prune().
- At the thread's end, the five candidate changes had landed, the isolation test remained omitted, and broader pruning hardening had moved to separate discussion; Peter introduced a new range-based visibility-map skipping proposal in #90 and suggested precomputed ranges for future VACUUM prefetching.