model-writtenrun 1408 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Konstantin Knizhnik reported that amcheck falsely flagged a parent/child high-key mismatch after interrupted autovacuum left a half-dead B-tree page, and suggested checking the half-dead flag in #1.
- Mihail Nikalayeu agreed that half-dead pages should bypass the child high-key test and suggested an injection-point reproducer in #2; when Konstantin could reproduce only on PostgreSQL 16 in #3, Mihail argued that newer versions likely remained vulnerable in #4.
- Heikki Linnakangas demonstrated a simpler reproducer in #5 and explained that the problem was a persistent half-dead page after interrupted vacuum, rather than concurrent execution of vacuum and bt_index_parent_check().
- Peter Geoghegan proposed tracking half-dead pages similarly to incomplete splits in #6, but Heikki showed in #7 that a direct P_ISHALFDEAD check sufficed because a half-dead page itself lacks a downlink; his optional cleanup ideas for deleted-page checks were not included in the submitted series.
- Heikki submitted four patches in #8, pairing injection-point tests with the half-dead-page fix and a newly discovered false positive for interrupted root splits; Peter reviewed the whole series favorably in #9.
- The incomplete-split test landed unchanged as 1e4e5783e7, and its amcheck fix landed as 6c05ef5729 with an added Makefile dependency on contrib/amcheck, both from #8.
- The half-dead-page test and amcheck fix from #8 landed unchanged as c085aab278 and cbe04e5d72; Heikki confirmed the series was committed in #10, and both amcheck fixes were backpatched through PostgreSQL 14.
- Tom Lane reopened the testing issue in #11 after two buildfarm animals failed to produce the expected half-dead-page error and completion notice.
- Alexander Lakhin reproduced those failures and traced them to autovacuum snapshots preventing deleted tuples from becoming removable in #12, attaching a diagnostic delta that repeatedly ran the test.
- Heikki fixed the test separately in 1c64d2fcbe by waiting for the removable cutoff before vacuuming, rather than adopting the diagnostic repetition patch, and confirmed the fix was pushed in #13; the thread ended with the original fixes and the test stabilization committed.