model-writtenrun 5478 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tushar reported that PostgreSQL 15 beta pg_upgrade failed after a successful --check because pg_upgrade_output.d already existed #1; Daniel Gustafsson initially considered documenting manual removal, while Justin Pryzby showed that failed invocations also prevented retries #2 #3.
- Daniel proposed revisiting timestamped directories, and Tom Lane suggested adding a timestamp level beneath the existing pg_upgrade_output.d root so whole-directory cleanup remained straightforward #4 #5.
- Michael Paquier first submitted a narrow fix that removed logs after successful --check unless --retain was specified and added a TAP check before upgrading #9; Justin demonstrated that this still left retries after failures broken #10.
- Michael adopted timestamped per-run directories, centralized cleanup, and added failure/success coverage in #12; Daniel then identified that successful cleanup would wrongly delete earlier retained logs and requested readable timestamps and path-length checks #13.
- Michael revised cleanup to remove only the current run and remove the root only when effectively empty, added path checks, and kept cleanup restricted to successful exits so failure logs survived #14; Daniel endorsed the revised cleanup #15.
- Justin suggested ISO-style timestamp formatting #16, and Michael retained milliseconds to avoid collisions during quick failed runs #17; the next submission also repaired recursive discovery of failure logs in the TAP test #18.
- Justin caught overlapping snprintf input/output and requested a dot before milliseconds plus prose corrections #19; Michael incorporated those changes, additionally zero-initialized directory buffers, and committed #18 as 4fff78f009, announcing completion in #20.
- After Justin questioned buildfarm readiness #21, Michael submitted a separate TestUpgrade.pm patch and pull request for recursive log discovery, explaining that fixing the broken beta upgrade workflow had taken priority #22.
- Andrew Dunstan explained that TestUpgrade.pm already does nothing when the pg_upgrade TAP test exists #23, so Michael agreed to withdraw the unnecessary buildfarm pull request #24; the thread ended with the PostgreSQL fix landed and Andrew planning to remove redundant buildfarm logic that had caused confusion #25.