model-writtenrun 3385 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Martín Marqués submitted wording corrections and documentation of two pg_combinebackup limitations: backups must be available in plain format, and enabling checksums between backups can produce reconstructed pages with invalid checksums, a problem found by Tomas Vondra #1.
- Tomas Vondra argued that documentation alone was insufficient and proposed detecting mixed checksum states and warning users; David Steele favored a fresh full backup and recording checksum state in the manifest #2 #3.
- David Steele questioned the storage cost of extracting tar backups, while Tomas Vondra and Magnus Hagander defended incremental delivery and discussed copy-on-write storage; Tomas reported limited success mounting archives, but David considered that workaround cumbersome and native archive processing remained future work #4 #5 #6 #9 #12.
- Magnus Hagander advocated failing unless users explicitly requested checksum disabling, while Tomas Vondra raised recovery complications and explained that backup-time detection lacked the previous pg_control file; both discussed adding checksum state to the manifest to detect problems earlier #7 #10 #11.
- Robert Haas committed only the spurious-word removal from Martín Marqués's initial submission as fbed6ebe41, declining the other wording changes because he preferred the existing text #1 #13.
- Robert Haas compared documentation, warnings, an explicit disabling option, manifest-based prevention, and automatic checksum recomputation, favoring documentation plus a warning after feature freeze and leaving recomputation for a future release #14.
- Daniel Gustafsson supported the conservative approach and warned that a checksum-disabling option could spread through copied tutorials; Robert Haas agreed with that concern #15 #16.
- Robert Haas supplied a replacement checksum patch that checks backup control files, warns when earlier checksum states conflict with an enabled final backup, and documents full-backup and checksum-reset workarounds #17.
- Daniel Gustafsson reviewed the patch and suggested specifying when checksum changes matter; Robert Haas explained that reconstructed intermediate backups make references to an initial backup ambiguous, and Daniel accepted the original wording in the thread's final message #19 #20 #21.
- The checksum patch subsequently landed unchanged as 205db0114e from #17; the two candidate commits contain the typo correction and checksum safeguards, but do not include the original plain-format documentation addition or implement native tar support.