model-writtenrun 4998 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Nov 2021 - Feb 2022 · 9 entries
- Nathan Bossart opened the thread with archive-module infrastructure carried over from an earlier parallel-archiving discussion, proposing archive_library, initialization callbacks, and a basic_archive test module in #1.
- Nathan clarified that the immediate goals were avoiding shell-launch overhead and enabling safer archiving; batching, parallelism, and restore-command modules were outside the initial implementation in #4.
- Michael Paquier questioned duplicating shared-library preloading in #3, while Robert Haas defended an explicit module interface and reloadable configuration in #5 and #12; Nathan consequently replaced mandatory preloading with SIGHUP-controlled archiver loading in #14.
- Fujii Masao proposed making basic_archive available to users in #2, and David Steele explained connection-startup overhead in #16; Nathan initially resisted promoting the module because existing archives and shared temporary filenames were not handled robustly in #17.
- Nathan split the feature into four independently reviewable patches in #21, revisited possible batching designs without adding them in #23, and refreshed the series with a Windows include fix and rebases through #25.
- Robert committed the unchanged shell-code extraction from #25 as aeb4cc9ea0, then requested a contrib module, identical-file acceptance, shutdown callbacks, and safer temporary filenames in #26 and #28.
- Nathan's remaining series in #30 moved basic_archive into contrib and added those robustness features; #32 ensured shutdown callbacks ran on archiver exit, and #33 added module-local error recovery so failures could use normal archiver retries.
- Robert's review in #34 led Nathan to use an empty archive_library as the sole shell-archiving selector and refine documentation in #35; Robert committed the complete remaining series with minor comment edits as 5ef1eefd76.
- The thorntail buildfarm failure reported in #38 showed that basic_archive tests inherited wal_level=minimal; Nathan's explicit replica setting in #41 landed unchanged as 00c360a89c.
Jul 2022 · 1 entry
- Benoit Lobréau, initially posting as talk to ben, reported that a working basic_archive configuration was absent from pg_settings in #44; Nathan explained process-local module loading in #45, and Alvaro Herrera proposed documenting explicit LOAD as a remedy in #48.
Aug - Nov 2022 · 11 entries
- Benoit's documentation patch in #51 initially covered both basic_archive and pg_settings, but Nathan and Tom Lane recommended documenting the generic behavior near the view in #52 and #53.
- Benoit adopted Nathan's backend-local loading explanation in #56, Nathan added links in #57, and Tom moved and polished the paragraph before committing 857808a411.
- Peter Eisentraut observed that the original feature had removed useful archive_command documentation in #63 and supplied a restoration patch in #65; Nathan reviewed its wording and duplicated explanations in #70.
- Peter incorporated that feedback and adjusted for intervening documentation work, then committed ba50834551, reporting the landing in #75.
- Peter separately objected to silently ignoring archive_command when archive_library was set in #66; Nathan and Tom explained why cross-variable GUC hooks were unsuitable in #67 and #69, leading Nathan to sketch warning-based suspension in #73.
- Peter distinguished invalid dual configuration from intentionally disabling archiving in #79, so Nathan replaced warnings with a library-loading error in #80, which Peter endorsed in #81.
- Bharath Rupireddy proposed GUC check hooks in #82, but Nathan demonstrated reload-order failures in #83 and Tom reinforced the objection in #84; Bharath withdrew that approach and resent Nathan's runtime-check patch unchanged in #86.
- Bharath also demonstrated that changing only archive_command bypassed the startup check in #86; Nathan added reload validation in #88, and Bharath simplified its control flow in #89.
- Kyotaro Horiguchi identified a misleading restart message before the configuration error in #90, while Michael argued that modules should retain flexibility to use archive_command in #91; Nathan's revised patch checked conflicts before restart handling in #94 and defended separate command and module interfaces in #95.
- Peter committed the final conflict checks from #94 with diagnostic and documentation wording edits as d627ce3b70, announcing a development-branch-only landing in #98.
- All six changes had landed by the thread's end, but backpatching the conflict check remained unsettled: Alvaro favored consistent behavior in #100, Nathan questioned changing behavior in a minor release in #101, and Michael reiterated both his design objection and the need for branch consistency in #102.