model-writtenrun 3841 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- David Steele proposed replacing backup_label with recovery information in pg_control, returning a safely constructed binary control file from pg_backup_stop() to prevent torn copies and make deleting recovery metadata fail visibly; the initial patch also updated pg_basebackup and pg_rewind, while deliberately retaining the existing recovery structure #1.
- David G. Johnston questioned compatibility and suggested additionally writing server-side files that backup tools could copy, but he and David Steele concluded that concurrent backups and the need to preserve the running cluster made that approach unattractive; explicitly rejecting an obsolete backup_label remained an option #2 #3 #4 #5.
- Michael Paquier's review led to clearer history-file naming, pg_controldata output, documentation corrections, pg_resetwal handling and better field placement across the next revisions, and his request to isolate sendFileWithContent() produced a separate prerequisite patch; a reported recovery-test failure could not be reproduced on repeated runs #8 #10 #11 #12 #13.
- Michael Paquier committed the sendFileWithContent() prerequisite from #13 as 7606175991, preserving its executable changes while polishing comments and the commit message, and confirmed that only this first patch had landed #14.
- Further review identified lost diagnostic information and undersized control files, so David Steele's subsequent revisions restored the 8kB file size, exposed recovery fields through SQL, and added backup labels and start/end timestamps to the manifest; Michael Paquier forwarded these patches after discovering that intermediate replies had accidentally omitted the mailing list #14 #15.
- Michael Paquier suggested retaining recovery fields as historical evidence, but David Steele preferred resetting them at recovery completion for now and considering retained history or a recovery-state variable separately #15 #16.
- Robert Haas requested broader consensus before another backup API change, asked for manifest tests, and raised conflicts with incremental backups whose safety metadata then lived in backup_label; David Steele acknowledged the need for wider review and discussed renamed functions or an explicit old-label rejection #17 #18.
- The proposed use of backup manifests for incremental-backup metadata remained disputed: Robert Haas emphasized that manifests were optional and that deleting one must not make an incremental backup appear bootable, while David Steele questioned whether incremental backups should permit omitting them #19 #20 #21 #22.
- Andres Freund argued that the redesign still allowed external backup tools to retain an incorrect pg_control and made inspection harder, proposing instead a control-file flag requiring backup_label for streamed backups; Michael Paquier found that narrower safeguard plausible, while David Steele wanted equivalent protection available to external tools #23 #27 #29 #30 #31.
- David Steele clarified that the main failure he wanted to prevent was users deleting backup_label after a restore error and mistaking a corrupt startup for success, and discussed retaining a separate human-readable backup-information file to address debugging concerns #33 #36.
- The manifest discussion also exposed a performance disagreement: Andres Freund supplied measurements showing substantial overhead, including with compression, and David Steele acknowledged the implementation bottleneck while defending checksums for long-lived backups and mentioning faster checksum algorithms #35 #37 #38 #39.
- Stephen Frost supported extending safety improvements to external tools and stressed the demonstrated torn-control-file risk, while Robert Haas defended concerns drawn from users of home-grown backup scripts and questioned whether repeated API changes produced enough benefit; the exchange did not establish consensus for removing backup_label #40 #41 #42.
- After the patches stopped applying and incremental backups introduced further dependencies on backup_label, Michael Paquier suggested returning the proposal with feedback, and David Steele withdrew it rather than undertake a substantial rework against the unresolved objections; the manifest and recovery patches had not landed in this thread #43 #44 #45.
- David Steele later proposed retaining backup_label while optionally returning a tear-free pg_control with a backupLabelRequired flag, allowing both pg_basebackup and external tools to gain protection without removing the existing mechanism; Stefan Fercot supported extending the backup functions for external orchestration #46 #47.
- At the thread's end, David Steele reported implementing that revised approach and moving it to a new discussion, leaving only the sendFileWithContent() prerequisite committed here and the original backup_label-removal proposal withdrawn #48 #45 7606175991.