model-writtenrun 7148 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Oct - Nov 2021 · 3 entries
- Alexander Lakhin reported an uninterruptible PostgreSQL 14 autovacuum hang during concurrent regression testing in #1 and supplied a retry assertion and reduced workload in #2.
- Matthias van de Meent reproduced the problem and proposed preventing GlobalVisState->maybe_needed from moving backward in #4, then identified snapshot-registration problems that made that fix insufficient in #5.
- Matthias supplied broader xmin instrumentation in #7, but Andres Freund explained in #10 that horizons can legitimately retreat, including when standby feedback arrives, so monotonicity could not simply become a hard assertion.
Dec 2023 - Jan 2024 · 3 entries
- After a two-year gap, Noah Misch reported in #12 that an intervening bugfix stopped the old reproducer while production symptoms remained, and Peter Geoghegan confirmed recurring hangs on PostgreSQL 14 and 15 in #13.
- Noah raised concerns about retaining tuples older than OldestXmin after removal of the retry loop in #14; Peter proposed giving pruning an explicit immutable OldestXmin floor in #18, and Noah supported that design in #19.
- Peter's production investigation found a tuple whose deletion XID lay between the conflicting horizons in #23 and located the hang immediately after index vacuuming in #26, establishing an important lead for the eventual reproducer.
Mar - Jul 2024 · 14 entries
- Robert Haas analyzed the independent cutoff calculations and their unreliable ordering assumptions in #35, while Alexander refreshed the original reproducer in #41 and Robert successfully ran it against the original PostgreSQL 14 release in #42.
- Andres argued that the suspected corruption path would instead encounter defensive freezing errors in #46 and proposed checking OldestXmin before GlobalVisState in #48, preserving the ability to prune additional tuples using newer visibility information.
- Alena Rybakina supplied an independent isolation test of dead-tuple statistics in #62, but Matthias explained in #64 that retained tuples could reflect ordinary concurrent snapshots, so this did not establish a reproducer for the hang.
- Melanie Plageman constructed a standby-reconnection and index-locking reproducer for PostgreSQL 14 and 15 in #66, then implemented the OldestXmin-first fix in #68 and adapted it to PostgreSQL 15 in #70.
- Bowen Shi's interrupt experiments and Melanie's discussion with Andres established in #69 and #70 that the held buffer content lock prevented a simple CHECK_FOR_INTERRUPTS solution.
- Bowen supplied PostgreSQL 16 diagnostics in #73 and #75; Peter and Andres connected index processing with horizon recomputation in #80 and #81, and Bowen's empty dead-items array later confirmed the same pattern in #93 and #94.
- Melanie packaged a TAP reproducer with the unchanged fix in #84, initially treating the test as demonstration material because its timing was fragile.
- Noah questioned maintaining a separate tuple-removal rule in stable branches in #88 and #90, prompting Melanie to replace the cutoff fix with a manual backport that reused cached visibility results and removed the retry loop in #99.
- Melanie redesigned the reproducer in #100 to hold a buffer pin while reconnecting the standby and force index vacuuming before later pruning, reproducing the hang across PostgreSQL 14–16 and exposing freezing errors after the proposed backport.
- Melanie confirmed in #101 that the backport also failed on PostgreSQL 16 and that the development branch could encounter the freezing error, so merely removing the retry loop was abandoned.
- Noah checked the backport and confirmed the reproduction results in #102, after which Melanie returned to the OldestXmin-first fix and supplied branch-specific fixes and stabilized tests in #103, aided by Heikki Linnakangas's BackgroundPsql backport.
- At the thread's end, the final series was still proposed, with test stability the remaining concern: Noah accepted either commit arrangement in #104, and Michael Paquier supported separating the test from the fix in #105.
- The later commit 83c39a1f7f enforced the OldestXmin floor in pruning and updated explanatory comments, but its matching source is the July 17 submission #566995 in thread #49829, superseding the earlier attribution to this thread's July 8 submission #103.
- The recovery-test components represented by #103 landed separately as e2e992820f for PostgreSQL 15, d42f60ccf0 for PostgreSQL 14, and 01e2b7f0fd for PostgreSQL 16, preserving the requested separation in case test stability required removal.