model-writtenrun 9094 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Hans Buschmann reported that PostgreSQL 12.0 on Windows could not restart after creating physical replication slots because syncing a slot state file caused a bad-file-descriptor PANIC #1.
- Andres Freund traced the regression to a cleanup that had accidentally undone an earlier fix requiring read/write descriptors for fsync on some operating systems, and called for tests with fsync enabled #2.
- Andres restored O_RDWR in RestoreSlotFromDisk() and added a protective comment in a586cc4b6c, then announced the fix and explained the risks of temporarily disabling fsync as a workaround #3.
- Hans described difficulties building the Windows dependencies and requested a prepared dependency bundle #4; Michael Paquier pointed to the documentation and raised licensing and maintenance concerns about distributing such a bundle #6.
- Michael's audit found the same descriptor-mode error in logical rewrite checkpointing, and he submitted both the actual fix and diagnostic instrumentation that enabled fsync in TAP tests and detected read-only descriptors #5.
- Andres reviewed the checkpoint issue, suggested fstat() to distinguish directories in possible assertions, and emphasized the need for broader fsync test coverage #8.
- Michael committed the checkpoint fix unchanged as b8e19b932a and confirmed completion #9; the accompanying diagnostic instrumentation was left out because it had been supplied as a bug-finding aid #5.
- Both bug fixes had landed by the end of the thread, while test infrastructure remained follow-up work: Andres proposed removing pg_regress's forced fsync disabling #10, and Michael pointed to a separate discussion he had opened on that problem #11.