model-writtenrun 7659 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Michael Paquier opened the investigation in #1 after multiple Windows users reported WAL recycling permission failures in PostgreSQL 13, identifying the switch from renaming to hard-link creation followed by unlinking as a suspected regression.
- Magnus Hagander proposed testing a revert and supplying replacement binaries to affected users in #2 and #4; Michael reported that his initial revert test ran without failure, but that reproducing the original failure remained unreliable in #3.
- Ranier Vilela submitted #5, proposing a rename() attempt followed by a MoveFileEx() fallback with MOVEFILE_COPY_ALLOWED in src/port/dirmod.c; this alternative was not incorporated into the eventual fix, and the thread records no explicit rejection rationale.
- Michael reported distributing two builds to affected users for comparison in #6, relying on their environments because his own archive-command failure remained difficult to reproduce.
- Andres Freund questioned the broader hard-link-and-unlink design in #7, explaining how unchecked unlink failures could leave one physical file at multiple future WAL positions, and proposed reproduction scenarios and link-count diagnostics in #8; Michael confirmed similar testing and instrumentation without a reliable reproduction in #9.
- Michael announced the landing of 909b449e00 in #10 after reporter testing supported the fix; its latest matching patch came from #679508 in thread #70637, restoring rename() on Windows in durable_rename_excl() with final documentation edits.
- The thread ended with the immediate regression addressed, while Michael left broader review of the file-recycling design for future work and explicitly avoided further stable-branch changes in #10.