model-writtenrun 4226 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andrew Dunstan reported that fairywren's Windows pg_basebackup tests exceeded the 255-character path limit on REL_16_STABLE, whose branch name lengthened the test paths, and initially proposed shortening the test filename after other workarounds failed #1.
- Andrew instead pushed a short-symlink workaround that had worked in local testing #2; Daniel Gustafsson reported buildfarm failures #3, and Andrew reported a corrective fix that Daniel acknowledged #4 #5.
- The symlink workaround allowed file creation but still left the long-filename rejection test failing, apparently because the server could not see the file, so Andrew proposed skipping that test when its Windows path exceeded 255 characters #6; Daniel expressed reservations but had no alternative #7.
- Andrew submitted the conditional-skip patch #8, Daniel approved it #9, and Andrew committed it as 1fa25ce70f, keeping the shared filename variable outside the skip block for later tests and recording backpatching to all live branches #10.
- Skipping the first failing test exposed a further Windows junction-path failure involving pg_replslot; Andrew proposed using the shorter system temporary directory and then reported pushing a tested fix #11 #12.
- Months later, Alexander Lakhin reported that the pg_replslot change failed when the source tree and temporary directory were on different drives #13; Andrew suggested File::Copy::move() #14, but Alexander tested it and found that it also could not move directories across filesystems #16.
- Alvaro Herrera proposed dynamically sizing pgreadlink's DeviceIoControl buffer #15, while Andrew suggested an untested same-device temporary-directory workaround and continued seeking a better solution #17.
- After Andrew questioned the proposed Windows API behavior #18, Alvaro withdrew that approach and suggested replacing MAX_PATH with MAXPGPATH in dirmod.c #19; the thread ends with Andrew promising to test this, leaving the cross-drive regression unresolved in the supplied discussion #20.