model-writtenrun 7545 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Sep 2018 · 5 entries
- Thomas Munro opened the thread with an experimental administrator-supplied command for libc collation versions, motivated by impending glibc collation changes, and separately supplied a hypothetical native version API in #1 and #2.
- Christoph Berg argued for automatic detection and suggested using the glibc release in #3; Thomas implemented that proxy in #7, while Peter Eisentraut identified false alarms and missing database-default coverage in #8.
- Thomas proposed per-index version tracking in #19, and Douglas Doole's CHECK-constraint example in #20 helped broaden the design to versioned object dependencies stored in pg_depend, proposed in #30.
- Stephen Frost advocated column- or type-associated storage in #31 and #33, while Thomas argued in #32 and #36 that separately created indexes and constraints need independently captured versions.
- The prototype in #38 removed pg_collation.collversion, added pg_depend.refobjversion, and checked indexes and CHECK constraints, but did not yet handle the database-default collation.
May 2019 · 1 entry
- After an eight-month gap, Thomas rebased the core index series in #48, leaving CHECK support out because it needed more work and had already demonstrated the generality of dependency storage.
Oct 2019 - Nov 2020 · 28 entries
- Peter revived the independent glibc feature in #49, requested direct __GLIBC__ detection, and committed its component from #50 as 9f90b1d08d.
- Peter immediately reported reverting that landing because pg_dump and pg_upgrade tests failed in #52; Thomas corrected the dump expectation in #53 and documented the limitations in #55.
- Thomas relanded the complete revised glibc patch as d5ac14f9cc, then reported another buildfarm failure with fuller locale installations in #57, while default-collation and per-object tracking remained unfinished.
- Julien Rouhaud tested the index prototype and found that REINDEX itself emitted warnings recommending REINDEX in #58; Thomas rebased it in #59 and invited further development.
- Julien took over extending the series in #67, adding default-collation dependencies despite system pinning and preserving index versions through binary upgrades.
- Michael Paquier questioned silently assuming compatibility for old indexes in #70, Laurenz Albe proposed an explicit compatibility override in #72, and Julien developed unknown-version handling and an upgrade option.
- Thomas's tests in #76 exposed missing expression, predicate, and composite-type dependencies, leading to recursive type analysis, regcollation, and expanded upgrade handling in #86.
- Julien moved checks into planner and VACUUM paths and suppressed repeated checks in #88, avoiding warnings during corrective REINDEX operations and adding a user command to acknowledge a current version.
- Testing in #89 exposed duplicate warnings and missing range dependencies, and Tom Lane highlighted nondeterministic hashing in #91; Julien addressed these cases in #94.
- Peter's review in #95 drove nullable text storage, automated tests, internal upgrade functions, and clearer refresh syntax, implemented across #97 and #98.
- New warning tests exposed confusion between unsupported version reporting and an unknown stored version, corrected in #102, with further dependency handling and default-collation lookup fixes in #105 and #107.
- Thomas consolidated documentation and tests in #109, and Michael's subsequent reviews led to a pg_depend TOAST table, standalone regcollation tests, and fuller documentation in #113 and #116.
- Peter committed the independent regcollation component from #124 as a2b1faa0f2, adding the missing pg_upgrade rejection check and documentation cleanup described in #125.
- Michael's concurrent-reindex tests in #122 and #128 showed that interrupted rebuilds could associate versions with the wrong physical index; Julien repaired dependency swapping and resent the series without regcollation in #130.
- The remaining series underwent repeated spring and summer rebases, after which Thomas removed overlooked collversion behavior from pg_dump and collationcmds.c in #138.
- Michael and Peter objected to hardcoded hash/bloom exemptions in #139 and #140, leading to their removal in #143 and further API and dependency-registration cleanup in #146.
- Julien tried an extensible access-method flag in #148 and added pattern_ops exceptions and tests in #157, but Thomas deferred the flag while retaining the narrower operator-class exceptions in #158.
- Thomas substantially simplified the implementation in #160, allowing duplicate dependency rows while suppressing duplicate warnings and consolidating version updates; Julien then caught a case where that suppression could hide a real warning in #161.
- Cross-version upgrade testing produced fixes for OID formatting, collation-name resolution, STRICT handling, and unknown versions in #163.
- Thomas changed binary-upgrade restoration to direct catalog updates and made pessimistic warnings for pre-feature indexes opt-in in #164, explaining that default warnings would otherwise create excessive noise.
- Julien identified a platform-dependent regression test failure in #165; Thomas fixed it in #166 and also deferred VACUUM warnings, special CollateExpr handling, and the expanded pg_dump TAP tests.
- The first two components of #166 landed unchanged as 7d1297df08 and cd6f479e79, replacing collation-level version storage with dependency-version infrastructure.
- The index-tracking component landed as 257836a755 after Julien's final review in #167 moved warning-list registration inside the mismatch branch; Thomas confirmed the change and landing in #168.
- Windows buildfarm failures appeared immediately in #170 and #171, and David Rowley traced them to traditional locale names rejected by GetNLSVersionEx(), supplying a conversion patch in #174.
- Thomas questioned the reliability of automatic name conversion in #177 and replaced it with tolerance of failed version lookups in #181; David tested the result in #183, and it landed unchanged as 9f12a3b95d.
- A maintained Windows conversion table was proposed in #185 but drew maintenance objections in #186 and #187; Thomas left reliable normalization and more nuanced version compatibility for future work in #189.
- Thomas revived the previously uncommitted FreeBSD provider work in #190, adding C/POSIX exclusions and documentation, and it landed unchanged as ca051d8b10.
- By the thread's conclusion, the index feature and platform support had landed, while CHECK tracking, access-method-specific exemptions, VACUUM warnings, and reliable conversion of traditional Windows locale names remained deferred, as explained in #63, #158, #166, and #192.
Mar - Apr 2021 · 1 entry
- The Windows limitation became a release open item in #191, and the final submission #192 documented it; 1bf946bd43 landed that text with terminology corrections.