model-writtenrun 7834 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Sep 2019 · 2 entries
- James Coleman proposed documenting waits between concurrent index builds on different tables in #1; Bruce Momjian questioned whether the existing snapshot explanation already sufficed in #2, while James argued that this cross-table effect was operationally surprising in #6.
- Alvaro Herrera suggested exempting simple, non-partial index builds from snapshot accounting in #7, and Michael Paquier highlighted related deadlocks involving REINDEX CONCURRENTLY in #8.
Feb - Apr 2020 · 2 entries
- Andres Freund preferred fixing unnecessary waits and proposed a distinct process flag in #11 and #15, explaining that reusing PROC_IN_VACUUM could allow tuples to disappear underneath an index build.
- James and Andres separated snapshot waits from vacuum's inability to remove dead tuples in #19 and #20; their discussion in #21 identified expression indexes, catalog lookups and multiple xmin requirements as obstacles to a broader vacuum optimization.
Jul 2020 - Jan 2021 · 14 entries
- David G. Johnston reviewed the documentation in #23 and requested integration into the existing explanation; James responded in #24 with a reproducible two-session example, more precise waiting text, and a separate vacuum-impact patch.
- Alvaro posted an experimental PROC_IN_CIC implementation in #25 to eliminate waits between concurrent index builds, explicitly leaving the vacuum problem unresolved, then identified a missing end-of-transaction mask update in #26.
- While examining process flags, Alvaro discovered that PROC_IN_ANALYZE had never been used and proposed its removal in #27; Andres mildly objected because of possible future bloat-control uses in #28.
- Simon Riggs proposed periodically refreshing ANALYZE snapshots in #29, while Robert Haas and Tom Lane supported removing the unused flag in #30 and #31; the ensuing discussion explored TOAST safety, sampling consistency and planner estimates without producing an attached implementation.
- Alvaro committed the cleanup unchanged from #27 as cea3d55898, explaining in #42 that the discussion favored removal and that alternative approaches could address ANALYZE bloat.
- Tom pointed out that the cleanup discussion had displaced the documentation work and confused CI in #44, prompting James to restore the unchanged two-patch documentation series in #45.
- Michael reviewed the vacuum paragraph in #47, requested consideration of REINDEX CONCURRENTLY and questioned its location; David examined the explanation's clarity in #48 and supplied a context-only rebase in #49.
- Alvaro committed the waiting-documentation portion of #49 as 58ebe967f8, adding the development-branch qualification for partial and expression indexes and backpatching the general warning, as confirmed in #54.
- The remaining vacuum warning became a maintenance.sgml footnote in #52, but James argued in #53 that users planning an index build needed to encounter the warning in the command documentation; Alvaro agreed that REINDEX needed equivalent coverage in #54.
- Alvaro abandoned the footnote because it did not fit the maintenance discussion and posted replacement CREATE INDEX and REINDEX notes in #57, explaining the effect through the familiar behavior of long-running transactions.
- James requested explicit cross-table wording and questioned whether the warning should apply only to concurrent builds in #59; Alvaro clarified in #60 that ordinary builds also constrain vacuum, and James accepted that explanation in #61.
- Alvaro committed #57 with the cross-table wording additions as 93c39f987e, backpatched through PostgreSQL 12, and announced completion in #62.
- After committing, Alvaro raised and then withdrew a concern about database boundaries in #62 and #64, and closed the commitfest entry in #65.
- The thread ended with Tom suggesting simpler wording in #70, without another patch or commit; both documentation concerns and the flag cleanup had landed, while the PROC_IN_CIC experiment in #25 had no demonstrated landing among this thread's supplied candidates.