model-writtenrun 7200 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Greg Stark opened the thread after a long-lived monitoring connection caused a wraparound outage despite using ON COMMIT DELETE ROWS, proposing that truncation reset freezing horizons and table statistics because autovacuum cannot vacuum another session's temporary tables #1.
- Noah Misch endorsed the basic idea but questioned resetting planner statistics, suggested reusing vac_update_relstats(), and raised rollback safety; Greg Stark, Masahiko Sawada, and Noah concluded that rolling back these truncations also discards the affected pg_class tuple version #2 #3 #4 #5.
- After nearly a year, Greg added autovacuum warnings identifying old temporary tables and their owning backend PID, while flagging an unresolved case involving stale temporary tables and a newly connected session that never initializes its temporary schema #6.
- Nathan Bossart gave the patch an initial favorable review and suggested a statistics-based PID lookup; Greg retained direct namespace-owner lookup, renamed the changed API, and separated the potentially fragile regression test during the March 2022 refresh #7 #10.
- Andres Freund requested separate warning changes, questioned RecentXmin validity and the cost of global multixact calculations, and suggested using the backend's oldest visible multixact; Greg subsequently fixed a wrong TOAST relation argument, handled transactions that lock temporary tables without taking a snapshot, and extended TOAST coverage #12 #15 #16.
- Greg added GetOurOldestMultiXactId() for temporary-table creation and truncation, leaving non-temporary truncation's multixact horizon unchanged, then rebased the three-patch series in June 2022 #18 #19.
- After the patch missed a release, Tom Lane questioned cross-session catalog inspection, ABI compatibility, the multixact optimization, code placement, and test reliability, reopening design issues despite its earlier Ready for Committer status #20 #21.
- Andres objected that ResetVacStats() bypassed the table-access-method abstraction, while Greg explained that the warning code only inspected catalog data and explored the safety and catalog-bloat implications of in-place updates #23 #24.
- Greg relayed an alternative developed with Ronan Dunklau that would keep temporary-table freezing horizons locally, but Andres explained that transaction-status retention still required exported horizons and questioned the extra machinery; no attached version adopted this alternative #25 #26 #27 #28.
- Greg moved the catalog update into heapam_relation_nontransactional_truncate() and reused vac_update_relstats() in the December revision, retaining tests mainly for reviewer validation; subsequent discussion with Robert Haas supported distinguishing catalog inspection from cross-session modification #30 #33 #34.
- While investigating CI failures, Andres distinguished the temporary-table regression failure from unrelated failures and diagnosed artifact uploads failing because the test server continued changing its log; he committed a FreeBSD failure-cleanup step directly from discussion in 209f0f0e85 and announced it in the thread #37 #38, an attribution consistent with the supplied verdict for the related discussion in thread #46599.
- Justin Pryzby later reported continuing temporary-table test failures, and Greg reproduced them under Meson with concurrent test suites, reinforcing his proposal to omit the tests because concurrent transactions could hold back RecentXmin #39 #41 #43.
- Andres argued that the behavior should remain tested, proposed the relation-specific GetOldestNonRemovableTransactionId(rel), and objected to calling vac_update_relstats() from heapam_handler.c; Greg agreed to investigate and suggested landing the warning patch separately if the remaining work took too long #44 #45.
- Discussion of the relation-specific horizon exposed confusing documentation about which tuples become removable, leading Andres to propose clearer wording that Peter Geoghegan endorsed #46 #51 #52.
- Greg's April revision restored ResetVacStats() in heap.c and adopted relation-specific horizons for creation and truncation, reporting that he could no longer reproduce the test race; profiling then raised concerns about repeated catalog work and locks across every ON COMMIT DELETE ROWS table #53 #54.
- The temporary-table warnings, truncation fix, and tests remained uncommitted in the supplied thread, with performance and possible horizon reuse still open; after nine months without activity, Peter Smith marked the submission Returned with Feedback in January 2024 #54 #55.