model-writtenrun 5765 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Peter Eisentraut proposed database-level collation version tracking, including a stored version, an actual-version function, a refresh command, and connection-time mismatch warnings, now useful because libc platforms supported version tracking #1.
- Julien Rouhaud supplied a straightforward rebase and reviewed missing documentation and tests, connection-blocking errors, concurrent database deletion, and unchecked template versions; he also favored retaining autovacuum warnings because mismatches risk index corruption #2.
- Peter documented the upgrade-oriented COLLATION_VERSION option and downgraded the connection diagnostic to a warning, while declining interactive completion for that option and explaining why absent recorded versions should not trigger upgrade-wide warnings #3.
- Displaying versions in psql remained a separate enhancement, with Michael Banck suggesting the extended database listing, while the proposed query identifying affected objects remained unimplemented after Julien explained its dependency-analysis difficulties #3 #7 #2.
- Julien pressed for a user-facing error when concurrent deletion invalidated the refresh lookup and for validation before copying a template database; Peter proposed rejecting mismatched templates, and Julien agreed while requesting an exemption for explicit versions used by pg_upgrade #6 #9 #10 #12.
- Alvaro Herrera supplied the current_database()/psql-variable technique for regression tests, and Peter's next patch incorporated those tests, a catalog-scan refresh implementation, and template-version checks with the upgrade exemption #11 #13.
- Julien confirmed that the revised patch resolved the reported problems and requested clearer diagnostics; Peter narrowed the repair instructions to objects using the default collation, while discussion of the refresh function's ObjectAddress return type exposed broader inconsistencies #14 #15 #16.
- Peter committed 37851a8b83 from #13 with only diagnostic wording, a comment correction, and the catalog-version bump, leaving broader return-type cleanup for separate discussion #17.
- After landing, Julien asked about the separate database-default ICU work and Peter promised an updated patch; the thread ended with Alvaro explaining the historical event-trigger rationale for DDL return types and suggesting void returns for shared-object commands, leaving that cleanup unresolved #18 #19 #20.