model-writtenrun 6079 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Mark Dilger reported TAP failures caused by premature WAL removal under smaller max_wal_size settings and proposed explicit retention safeguards for recovery, pg_basebackup, and bloom tests #1 #3.
- Noah Misch and Tom Lane questioned hardening tests against every possible setting, but supported the narrower goal of preventing premature WAL removal; Tom considered missing safeguards likely oversights #2 #4 #6 #7 #8.
- Mark submitted replication-slot additions for the promotion and bloom tests, a wal_keep_size increase for pg_basebackup, and bloom TAP enablement, while acknowledging that bloom's historical instability might have another cause #10.
- After a successful local bloom test, Tom proposed enabling it independently to observe the buildfarm, and Mark agreed to defer retention changes to avoid confusing the diagnosis #11 #12; the unchanged Makefile hunk landed as 7d1aa6bf1c, announced in #13.
- Michael Paquier questioned what could have fixed the old failure, Tom could not reproduce it even with historical source on the original host, and Andrew Dunstan knew of no explanatory buildfarm change #14 #15 #16.
- A fresh intermittent buildfarm failure reopened the investigation, and Tom identified that the test waited for WAL to be written rather than replayed before comparing query results #17 #19.
- Mark raised unordered result comparisons and then analyzed missing standby rows, but Tom explained that bitmap-scan ordering should match under physical replication and rejected adding ORDER BY because it could obscure real defects #18 #20 #21.
- Tom considered Mark's proposed physical-file comparison outside bloom's scope, reproduced the failure by delaying WAL replay, and supplied a separate fix using wait_for_catchup plus improved assertion locations #22.
- The corrective attachment from #22 landed unchanged as 6bc6bd47cf, excluding the diagnostic delay; the commit also records backpatching the fix and test enablement to supported branches beginning with v12.
- Michael supported backpatching both changes to prevent others copying the faulty test, but his response followed the corrective commit and therefore does not establish pre-commit review coverage #23 6bc6bd47cf.
- Nathan Bossart asked whether the remaining WAL-retention patches should proceed, and Mark withdrew them to avoid complicating investigation of intermittent failures; the bloom enablement and replay-wait fix remained landed, with broader hardening deferred #24 #25.