model-writtenrun 6051 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Oct 2021 ยท 14 entries
- Alexander Lakhin reported that pg_amcheck failed when another session owned temporary tables and indexes, unlike tools that silently exclude them, in #1; Mark Dilger supplied temporary-relation filtering and a regression test in #3.
- Alexander expanded the problem to invalid indexes created concurrently and associated deadlocks in #5, while Peter Geoghegan identified a separate missing case for unlogged heap relations on standbys in #8.
- Mark initially argued that skipping requested indexes could misleadingly report success in #9 and #11, while Peter argued that invalid indexes should be treated as a catalog-visibility concern in #15; Robert Haas supported excluding relations that could not usefully be verified in #18.
- Alexander's concurrent regression-suite experiments also exposed deadlocks and disappearing-relation errors in #22, while Pavel Borisov reported that he and Maxim Orlov had reviewed the initial temporary-table patch and successfully run its tests in #24.
- Robert rejected silently downgrading explicitly requested verification options in #31 and urged keeping policy in the underlying amcheck functions in #34; Peter clarified that parent checks should still fail on standbys, and Mark accepted that distinction in #44.
- Mark's expanded patch in #47 combined temporary-relation filtering, execution-time checks of index validity and recovery state, a backend fix for unlogged heaps, documentation, and a new primary-and-standby test script.
- Peter's review in #48 challenged client-side recovery filtering, and his proposal in #58 resolved the concern about excessive notices by moving consistent unlogged-relation handling into amcheck with DEBUG1 messages; Mark implemented that proposal in #60.
- Peter explained that the documentation warning should emphasize the disruptive locks taken by parent checks in #61, and Mark asked him to revise the wording and commit in #62.
- Peter first committed the backend portion of #60 as 292698f158, with explanatory comment additions, then posted the unchanged client-side remainder in #63.
- Peter questioned redundant temporary-relation checks in #63, but Mark explained the concern about state changes before parallel workers execute in #64; the checks remained, and d2bf06db37 landed that remainder with only the documentation warning revised, as confirmed in #65.
- Windows buildfarm failures appeared after the new tests landed: Tom initially considered a port-binding failure transient in #67, but Andrew reported similar trouble on another animal in #69.
- Mark suspected the unfinished background_psql harness in #70, while Tom questioned that explanation and proposed deleting the tests because their maintenance and buildfarm costs outweighed their value in #72.
- Mark accepted removal in #77, and Peter deleted the entire test script in cd3f429d95, confirming the action in #80; the production fixes remained committed.
- Andrew's final investigation in #83 tied the failure to starting and pumping background_psql before starting the standby, identified Windows IPC::Run behavior and insufficiently robust port allocation as concerns, and recommended starting servers first; the removed tests were not restored and the infrastructure issue remained unresolved in this thread.