model-writtenrun 8591 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Sep - Oct 2019 · 4 entries
- Robert Haas proposed backup manifests after recognizing that CREATE DATABASE and tablespace moves undermine LSN-only incremental-backup detection; he wanted file inventories, optional checksums, and a way to verify backups #1.
- David Steele advocated timestamps, versioning, accessible manifests, and JSON, while suggesting that constructing manifests retrospectively could wait; Robert initially preferred a tab-delimited format to avoid substantial frontend infrastructure work #2 #9.
- Robert's first implementation paired COPY/tar refactoring with server-generated manifests containing file metadata and SHA256 checksums, including a checksum of the manifest itself #15.
- Vignesh's review found warnings and error-handling gaps, and Rushabh Lathia supplied a small corrective patch #17 #19.
Nov 2019 - Feb 2020 · 9 entries
- After a substantial gap, Rushabh rebased the series and made file checksums optional because Suraj Kharage's measurements showed substantial SHA256 overhead #21.
- Suraj added verification inside pg_basebackup using a filename hash table, and Jeevan Chalke reviewed both generator and validator details #25 #26 #27.
- Checksum discussions led to explicit algorithm selection and labeled checksum values, implemented in Rushabh's delta and then incorporated into the complete stack #29 #44 #45 #47.
- The COPY/tar refactor from #47 landed as 431ba7bebf, and Rushabh consolidated the remaining manifest work into #50.
- Jeevan's review prompted a manifest context structure, and Suraj combined that work with shared checksum functions, the validator, and Rajkumar Raghuwanshi's TAP tests #52 #53 #56.
- Robert's reviews drove replacement of the custom hash table with simplehash and a rewrite of parsing around a whole-file buffer, followed by bounded scans and improved diagnostics #57 #59 #60 #61 #78.
- The manifest's own checksum became unconditional SHA256, independent of the optional algorithm selected for individual data files #63 #65 #66.
- January discussion continued over JSON extensibility and frontend infrastructure: Stephen Frost and David favored a standard format, while Robert and Tom Lane resisted making a large infrastructure project a prerequisite #79 #86 #88 #94.
- Robert returned with a major rewrite that spooled manifests to temporary files, provided configurable checksum helpers, moved verification into standalone pg_validatebackup, and introduced JSON generation and parsing #96.
Mar 2020 · 6 entries
- Testing exposed invalid-path crashes, permission-related double frees, unchecked manifest checksums, and backups that passed verification despite unusable WAL; the immediate crash and checksum problems were fixed while WAL validation was initially considered outside scope #97 #99 #100 #105 #107.
- The next revisions removed the alternate text format, separated JSON parsing into its own module, expanded tests and documentation, and added encoded-path handling and manifest suppression #111 #115 #119.
- Amit Kapila's Windows testing exposed an incorrect option name, an initializer problem, and tablespace-junction removal failures, producing fixes incorporated into the full series #121 #122 #127 #145.
- Stephen challenged claims of backup validity without required-WAL verification and preferred stronger default hashes; documentation improved, but Robert retained configurable CRC32C as the inexpensive default with support from Andres Freund and Noah Misch #123 #140 #145 #174.
- David's malformed-manifest testing improved documentation and clarified that parsing errors can precede checksum verification, while more extensive diagnostic changes were deferred #152 #192 #196 #207.
- Andres supplied a pg_waldump parsing-only prototype and explained how timeline history could describe required WAL ranges; Robert integrated both into automatic backup validation #184 #188 #201.
Apr 2020 · 15 entries
- WAL tests and documentation followed, and manual timeline-switch testing found bugs that Robert fixed alongside Noah's proposal to publish the manifest only after WAL retrieval and fsync completed #204 #208.
- Andres requested a comparison between bytes actually read and the manifest size, producing the final attached series; reliable automated timeline-switch tests and broader tool consolidation remained deferred #211 #212 #217.
- The final series #212 landed in three pieces: pg_waldump --quiet as ac44367efb, checksum helpers as c12e43a2e0, and manifest generation plus validation as 0d8c9c1210, with only small final edits.
- Immediate buildfarm failures exposed test portability problems; Álvaro Herrera's path-conversion suggestion became 460314db08 without a matching attached patch #219 #222 #223.
- Justin Pryzby submitted post-commit documentation corrections, and Tom's report that make clean left tmp_check behind led directly to 9f8f881caa #224 #227.
- Further buildfarm investigation uncovered inaccessible test directories, a time_t versus pg_time_t mismatch, and an invalid resource-owner assumption in replication connections, with follow-up repairs restoring most hosts #234 #235 #236 #244 #246.
- Fujii Masao discovered that stale backup_manifest files could be copied into new backups; his two attachments became the exclusion fix 1ec50a81ec and documentation correction c4f82a779d after Robert's review #262 #263.
- Andrew Dunstan addressed MSYS2 argument conversion and junction removal; Tom and Robert warned that unconditional unlink would break other Windows environments, leading to the rmdir/unlink fallback in c3e4cbaab9 #264 #265 #266.
- Fujii's next review fixed --no-manifest with tar output to stdout, removed an unused option, and corrected protocol documentation; Michael Paquier reproduced the failure, and all three attachments landed unchanged as a2ac73e7be #272 #273.
- Robert supplied a separate manifest-format chapter, revised its spelling, and incorporated review clarifications about timelines, LSNs, and fixed SHA256 before committing 149f2ae88a #275 #277 #278 #284.
- The format discussion retained a byte-for-byte checksum excluding the final line, preserving a route to incremental verification, while reviewers debated timestamp representation and the explanation of checksum choices #279 #282 #284.
- Fujii added an index entry to make the new documentation discoverable, committed unchanged as 4db819ba40 #299 #300.
- External checksum storage was reconsidered without agreement to change the design, and David suggested a future manifest-only verification option beyond this release #301 #302 #304.
- Fujii's pg_verifybackup option cleanup landed unchanged as 0a89e93bfa, removing an unused option and correcting help output #307 #308.
- Andres's optimized-build warning report was resolved by Robert's callback noreturn annotation, which Andres tested before it landed unchanged as 05021a2c0c #253 #310 #311.
May 2020 · 1 entry
- By the final messages the feature and candidate fixes had landed, but timestamp representation remained unresolved: Fujii questioned consistency with backup_label, Tom favored its precedent, and David continued advocating GMT or epoch timestamps for machine processing #313 #315 #316 #318.