model-writtenrun 1322 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Олег Самойлов reported that basic_archive stopped archiving after an NFS directory became unavailable, accumulated about 4 TiB of WAL, and continued treating the directory setting as empty after access returned #1; after Alvaro Herrera suggested an NFS failure, Олег clarified that the persistent loss of the setting was the bug #2 #3.
- Sergei Kornilov reproduced the failure by restarting the archiver while its archive directory was inaccessible and proposed moving the existence check from the GUC hook into basic_archive_configured(), allowing later recovery #4.
- Fujii Masao supported relocating validation and identified configuration-reload diagnostics that could go unnoticed; his suggestion to move path-length validation too led Sergei to submit a revised two-patch series, separating validation relocation from removal of the empty hook because of backport concerns #6 #7.
- Nathan Bossart objected to repeating validation for every archived segment and proposed removing the existence check entirely, since archiving already had to handle a disappearing directory #8 #10; Fujii demonstrated the reload-logging issue, which Nathan argued should receive a general archive-module solution rather than drive this patch #11 #12.
- Nathan supplied a replacement patch that retained the GUC path-length check and simply removed directory-existence validation, and Tom Lane endorsed that approach #13 #15.
- Fujii raised concerns about replacing the explicit directory diagnostic with a file-creation error and suggested checking after configuration reload or after a copy failure, while noting that a new callback would not suit back branches #16; Sergei accepted the existing error message, Nathan rejected extra checks as unnecessary and still subject to races, and Fujii agreed to proceed #17 #18 #19.
- Nathan committed #13 unchanged apart from rebasing as 8fc45ac5d9, with backports through PostgreSQL 15, and closed the thread by announcing the commit #20; the broader reload-logging issue had no implemented solution in this thread.