model-writtenrun 7261 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
May 2020 · 2 entries
- Robert Haas began by separating base-backup transport, throttling, progress, archive generation and manifests into composable bbsink and bbarchiver objects, with preparatory cleanups in #1 and #2.
- Andres Freund questioned callback overhead and stressed progress measured before compression in #3; Suraj Kharage and Dipesh Pandit tested buffer sizes and backup performance in #9 and #13.
Jun - Jul 2020 · 1 entry
- After initial cleanups landed separately, Robert expanded the surviving series with client bbstreamers and a gzip prototype in #15, proposing one typed COPY stream with archive names and independent progress messages; Andres reviewed compatibility and recovery-file placement in #16.
Jul 2021 · 2 entries
- Following nearly a year’s gap, #19 replaced the experimental design with consolidated sinks, a new COPY subprotocol, server and blackhole backup targets, and usable server-side gzip; Robert explained in #27 that bbarchiver was dropped because alternate archive formats had no concrete demand.
- Tushar’s gzip failure report #20 led Dilip Kumar to supply the missing end_archive forwarding call in #21; after a reminder that it had been missed, Robert incorporated it in #40.
Sep 2021 - Jan 2022 · 16 entries
- Jeevan Ladhe’s LZ4 prototype #34 exposed output-buffer sizing requirements, prompting Robert to change sink buffer negotiation in #36 and Jeevan to adapt LZ4 in #45.
- Sergei Kornilov’s review #42 drove CHECKPOINT naming and compatibility corrections, while Jeevan proposed size_t buffer lengths in #43; #47 incorporated these changes, but the final BASE_BACKUP and CREATE_REPLICATION_SLOT patches were subsequently posted as #488471 in thread #42511, supplying 0ba281cb4b and 0266e98c6b.
- Jeevan found that server-target backups omitted tablespace_map and supplied #55; Robert folded it into the remainder #56, whose unchanged directory-conversion refactoring landed as 967a17fe2f.
- LZ4 review corrected preference initialization, autoflush and output-space accounting across #54, #58 and #62; the cleanup lifecycle added in #63 then exposed a missing gzip callback, fixed by Jeevan in #64.
- The bbsink and bbstreamer components of #63 landed as bef47ff85d and 23a1c6578c, with typedef-list omissions and small rebase adjustments, leaving protocol, targets and compression pending in #69.
- AIX buildfarm failures discussed by Tom Lane in #71 showed that bypassing tar parsing also bypassed required terminators; Robert’s patches #72 landed unchanged as client-padding fix 57b5a9646d and proper server termination 5a1007a508.
- Dmitry Dolgov diagnosed the wrong sink argument during manifest throttling in #75, leading directly to the one-line fix 1b098da200 without an attached patch.
- Robert documented tar termination and the new protocol in #77, retaining old-protocol compatibility and permitting -Xfetch with server targets; its documentation and protocol components landed unchanged as 81fca310b3 and cc333f3233.
- Jeevan extended the separate LZ4 series with initial server-side ZSTD in #88, while Robert narrowed the main series to targets and gzip in #89 and added end-to-end extraction and verification tests in #94.
- The target component of #94 landed as 3500ccc39b with a documentation wording correction, supporting server filesystem and blackhole destinations while streamed WAL remained client-side.
- Dipesh Pandit introduced client gzip decompression in #91; review required detection from the received archive name, proper stream finalization and dedicated extraction tests, developed through #95 and #99.
- Robert adapted the gzip remainder to client/server-prefixed compression options in #97, removed an overly verbose default-level notice and landed it after a trivial rebase as 0ad8032910.
- Noriyoshi Shinoda identified missing client/server help choices in #101, and Dagfinn Ilmari Mannsåker expanded the correction to documentation syntax in #103, which landed unchanged as e1f860f134.
- Tushar reported invalid default-level transmission and silently ignored -R with a target in #104; Robert’s fix #109 was tested and landed unchanged as dabf63bc9a.
- Robert consolidated Dipesh’s gzip extraction work in #115, and Dipesh removed redundant library guards in #116; these changes and a test correction landed as d45099425e.
- Tushar’s repeated-compression-option report #113 prompted the exact last-option-wins fix #114, committed as 51891d5a95; Jeevan’s LZ4 resubmission #119 also supplied the independent unused-test-variable cleanup 82331ed4dd.
Feb 2022 · 3 entries
- Jeevan refined server LZ4 levels and tests, while Dipesh supplied client compression and decompression in #127; server revision #130 landed as dab298471f with an assertion correction, and client revision #129 became 751b8d23b7 with option parsing, WAL handling and documentation edits.
- Robert proposed extension-registered backup targets and a configurable shell destination in #125; Abhijit Menon-Sen reviewed their integration design in #126, and documented revision #142 later supplied e4ba69f3f4 and c6306db24b, with an accidental Makefile.orig included in the first commit.
- Tushar’s large-backup ZSTD testing and Justin Pryzby’s review led to corrections in #141 and #149; Jeevan added client compression and decompression, separated build support for another thread and adopted clearer help syntax proposed by Álvaro Herrera in #147.
Mar 2022 · 13 entries
- Robert consolidated Jeevan’s ZSTD work and corrected exact-fit buffer flushing in #154; both attached components landed together unchanged as 7cf085f077, and Jeevan’s equivalent LZ4 correction #155 landed as 1d4be6be65.
- Robert standardized compression feature macros in #156, which Jeevan reviewed in #157 before it landed unchanged as 75eae09087; LZ4/ZSTD compression of separately streamed WAL remained outside this work because rewriting compressed tar headers required a different design, discussed in #152 and #153.
- Justin reported failure when combining server compression with standard output in #160; Robert chose to reject compressed manifest injection rather than silently decompress and recompress, landing #161 unchanged as b2de45f920, alongside Justin’s inline suffix-test cleanup #162 as d6f1cdeb9a.
- Dipesh proposed parallel ZSTD using @WORKERS in #167, but discussion with Justin pushed toward a general option grammar and server-side interpretation of compression details; Robert introduced that prerequisite in #175 to avoid freezing an inflexible interface.
- Justin separately explored negative ZSTD levels and reported an LZ4 output-buffer assertion in #174; the negative-level proposal continued through #194 and #215, but no candidate demonstrates its landing.
- Justin, Tom Lane and Ilmari reviewed compression-detail parsing, empty inputs, validation and documentation, with successive fixes ending in #190; it landed as ffd53659c4 with only SGML and typedef-list edits.
- Robert rewrote parallel ZSTD around workers=N in #192, reporting substantial speedups and a discrepancy between client and server compression; Andres Freund examined thread lifetime and cancellation in #193 and #198, with workers confined to libzstd data structures.
- The compression discrepancy revealed default-level and error-handling bugs: Robert supplied #203, Dipesh approved it in #204, and Justin republished it in #206 before the unchanged component landed as 61762426e6.
- Justin’s combined series #206 added worker-option fixes, long-distance matching and negative levels; Robert incorporated the worker feedback in #211 and #212, then Justin republished the final parallel patch in #215.
- The run_log return-value prerequisite was split from that final series into ad4f2c47de, and the remaining parallel client/server ZSTD component of #215 landed as 51c0d186d9.
- Tom’s Coverity report #205 led Robert to demonstrate and fix a valueless compression-option NULL dereference in #213; Justin confirmed it with long mode in #215, and 8e053dc6df retained Robert’s condition, omitted the demonstration keyword and added an explanatory comment requested in review.
- Ilmari proposed extending useful return values to more TAP wrappers in #220; Robert requested separate review, and Andrew Dunstan rejected the generalization in #228 because those underlying routines had no useful return value, while agreeing to fix the typo.
- Justin’s long-distance matching proposal in #215 continued in thread #47519, where #532330 supplied the final twelve-file implementation with shared compression APIs, pg_dump integration, validation and tests; that later patch supplied 2820adf775 in April 2023, superseding the earlier attribution to this thread.
Mar - Apr 2023 · 2 entries
- After a year’s gap, Thomas Munro’s report represented by #229 prompted Robert’s wait-event documentation and shell pipe-error patches in #231.
- After Justin’s reminder #232, the wait-event documentation component of #231 landed unchanged as 4b1ad19a4e, and Robert confirmed the final fixes committed in #233; the principal refactoring, targets and compression work was complete, while negative-level support and streamed-WAL compression were not shown landed here.