model-writtenrun 7295 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Thomas Munro opened the thread with a patch adding missing_ok and elevel parameters because pgwin32_is_junction() silently treated every error as a negative result, while acknowledging an awkward frontend logging workaround #1.
- Michael Paquier explained that SyncDataDirectory() needed nonfatal handling for pg_wal and suggested flags controlling missing-file and hard-error behavior, but that API direction was subsequently superseded #2.
- After several months, Thomas replaced the original approach with Windows lstat() support and removal of pgwin32_is_junction(), also adapting unlink() for junctions; Melih Mutlu's accompanying MinGW CI patch was explicitly testing infrastructure outside the submission #3.
- Thomas separated unlink() support and restricted rmdir() to confirmed junctions so ordinary directories would still be rejected, and Andrew Dunstan successfully tested the three substantive patches on fairywren and drongo #4 #6.
- Thomas then populated symlink st_size through readlink(), strengthened related error handling, and removed redundant S_ISLNK guards #8.
- The August series landed as Windows lstat() support with an added explanatory comment c5cb8f3b77, unchanged junction-aware unlink() support f357233c9d, and unchanged helper replacement 5fc88c5d53, all drawn from #8; Thomas announced the push and highlighted the need for automated porting-layer tests #9.
- Zharkov Roman reported that the new stat() behavior broke initdb beneath a volume-GUID junction because pgreadlink() stripped a required NT prefix, and supplied a volume-specific exception #10.
- Thomas generalized the fix to preserve prefixes except for drive-letter paths, and Roman confirmed that it worked across several junction formats while suggesting an alternative reparse-buffer representation #11 #12 #14.
- Thomas retained the existing raw-path approach, tightened recognition to drive-absolute paths, and added an eight-hop stat() implementation after recognizing that junctions outside PGDATA could form chains #15.
- Roman tested both revised patches, found that the chain loop repeatedly read the original junction, and supplied a current-path correction that passed short-chain and hop-limit tests #16.
- At this thread's last message, Thomas intended to commit the fixes but was waiting for automated coverage, with a prototype C TAP suite and a nearby unlink() fix still under development #17.
- The fixes subsequently continued in thread #46799, where the later matching submission #518475 supplied the readlink() fix committed as f71007fbb3 and the recursive stat() fix with ELOOP committed as 4517358ee7; that thread's verdicts attribute both landings to the October submission, with its filesystem tests omitted #518717, superseding the previous attribution to this thread's August patches.