model-writtenrun 7541 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tom Lane opened the thread with a reproducer showing collation versions leaking into unrelated dependency entries and unnecessary dependencies on an unversioned default collation, supplying both a complete fix and a compatibility-preserving alternative in #1.
- Julien Rouhaud tested both alternatives, explained why a domain's own collation dependency need not be versioned, and then reproduced the regression failure under --no-locale; he proposed giving the test column an explicit C collation in #2 and #4.
- Tom adopted Julien's C-collation test adjustment while declining additional default-collation tests with alternative expected outputs because their maintenance cost outweighed their coverage, posting the revised fix in #5.
- After Julien confirmed the revision in #6, Tom committed it unchanged as ef387bed87 and announced the push in #7.
- Buildfarm failures immediately exposed another locale-sensitive test problem: Andres Freund supplied his ICU and C.UTF-8 environment, and Tom traced the unexpected default-collation dependencies to leaf constants whose collations did not affect expression semantics in #8, #9, #10 and #12.
- Tom posted a separate rewrite that collected function and operator input collations and corrected range and array handling, while acknowledging missing record-comparison dependencies and questioning the use of pg_depend for version tracking in #13.
- Thomas Munro defended the benefits of tracking individual indexes and avoiding duplicate dependency infrastructure in #14, then supported landing the rewrite temporarily in #16; Tom committed #13 as f24b156997, adding dropped-attribute handling and deduplication cleanup, and confirmed completion in #17.
- Julien argued that missed dependencies were more dangerous than unnecessary rebuild warnings and favored retaining the feature in #20, while Tom demonstrated that the earlier analyzer could also miss dependencies when composite values came from casts or functions in #21.
- Andres challenged the suggestion that rebuilding everything was an adequate substitute in #22, and Tom supplied an ALTER TYPE ADD ATTRIBUTE example where an existing index acquired a collation dependency that its recorded metadata could not reflect in #23.
- Peter Geoghegan argued for a clearer specification of container comparisons rather than treating their behavior as inherently unknowable in #27, while Tom favored conservatively treating record operations as potentially collation-dependent and raised pinned-object and catalog-size objections to pg_depend in #26 and #29.
- Thomas considered reanalysis after type changes but recognized substantial locking problems in #28 and #30; Julien agreed that the ALTER TYPE case could not readily be fixed for PostgreSQL 14 in #32.
- Peter and Julien distinguished warnings about genuine version mismatches from incorrectly recorded dependencies in #33 and #34, and Thomas explored recording versions when composite attributes were added while noting further problems with dropped attributes in #35.
- Andres proposed retaining a narrower corruption-warning feature even without complete dependency coverage in #36, but after a quiet interval and Andrew Dunstan's release-status inquiry in #37, Thomas announced that he would revert and resubmit for PostgreSQL 15 in #38.
- Thomas removed the per-index tracking feature, including both fixes landed from this thread, in ec48314708; his final report in #40 explains the toast-OID and catalog-default adjustments needed to restore per-collation tracking.
- Thomas separately preserved useful platform documentation in b65431ca5e, explaining in #40 that restored per-collation tracking still supported more operating systems; the thread ended with the revert complete and the per-index redesign deferred.