model-writtenrun 7063 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Sep - Nov 2024 · 11 entries
- Jonathan S. Katz reported that large vector-search index expressions exceeded pg_index's tuple-size limit and supplied a reproducer in #1.
- Nathan Bossart proposed adding a TOAST table to pg_index in #2, and Tom Lane agreed in #3 that circularity was unlikely because system indexes do not use the affected expression and predicate columns.
- Nathan committed #2 with catalog-version and comment updates as b52c4fc3c0, then posted optional circularity checks in #7 while explaining that he had left them out.
- Alexander Lakhin found that dropping a large-expression index concurrently failed without a snapshot in #8, prompting Nathan's first follow-up fix in #9.
- Alexander identified the same problem in concurrent creation and reindexing in #11, and Nathan expanded the fix in #13; Alexander's stress testing found no further snapshot failures in #14.
- Nathan narrowed snapshot lifetimes to avoid interfering with deadlock-sensitive waits and added comments and tests in #15, while Michael Paquier approved the placement and suggested broader catalog assertions in #16.
- The pg_index snapshot fix landed as b52adbad46 from #15, with a small assertion adjustment and comment edits, as confirmed in #19.
- Michael's assertion proposal led Nathan to discover additional problems involving pg_constraint, pg_subscription, and pg_replication_origin in #17; Michael initially favored retaining the replication-origin TOAST table and endorsed the broader checks in #18.
- Nathan moved the assertions into heapam.c and introduced a RelationGetToastRelid cleanup in #20, then separated backpatchable fixes from development-only changes in #21 and added Michael's requested assertion-build guard in #25.
- Backpatch preparation delayed the work, and Nathan questioned whether to retain the macro cleanup in #27; Michael supported it in #28, and Nathan subsequently planned to proceed with it in #29.
- Nathan questioned requiring snapshots for updates that leave varlena columns untouched in #29, but Michael favored strict catalog-wide checks for future error detection in #30; uncertainty about replication-origin snapshot semantics remained in #31.
Mar - May 2025 · 9 entries
- After several months, Amit Kapila revived the alternative of removing pg_replication_origin's TOAST table in #33 and #35, and Euler Taveira's account of short real-world origin names in #37 helped Michael accept a 512-byte limit in #38.
- Nathan posted the removal patch in #39, then incorporated Euler's documentation and cosmetic feedback from #40 into #41 and corrected the regression-test name prefix in #42.
- Nisha Moond tested the removal patch and suggested moving validation to the SQL entry point in #43; Nathan explained that the limit mainly improves error reporting in #44 and adopted the relocation in #45.
- Tom Lane requested that the snapshot rationale be preserved in #46 and enforced by an assertion in #49, leading to clearer comments and a no-TOAST invariant in #48, #50, and #52.
- Although Nathan initially intended to wait for the next development cycle in #52, Tom argued that the removal qualified as a bug fix in #53, and Nathan reported release-team agreement to include it in PostgreSQL 18 in #56.
- Michael requested restoring validation to replorigin_create() so external callers also receive the check in #55 and #57; Nathan committed that modification to #52 as 16bf24e0e4 and announced it in #58.
- Nathan returned to the remaining snapshot fixes with branch-specific attachments in #60, and Michael reviewed the refreshed versions and the older-branch replication-origin exception in #61; those attachments are not represented among the available indexed patchsets.
- The concluding fix round produced b7ba2c0308, whose supplied diff adds snapshots for concurrent index swapping and autovacuum deletion; its provenance remains this thread's final staging despite links to the earlier bug reports in thread #72125 and thread #72342, and no indexed patchset establishes its complete source.
- Nathan's final message #62 confirms commitment of the remaining staged fixes; b7ba2c0308 explains that replication-origin problems remain on older releases because catalog removal cannot readily be backpatched and extremely long names are unlikely, while strict assertions were omitted on PostgreSQL 13 and 14.