model-writtenrun 7245 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Daniel Verite opened the thread with examples of C.UTF-8 sorting differently across glibc releases and a one-line patch to record versions for libc C.* collations in #1.
- Thomas Munro questioned permanent versioning for an intended stable ordering in #2 and #4, while Daniel traced the incompatible behavior to older Debian/Ubuntu locale implementations and emphasized missing upgrade warnings in #5.
- Jeff Davis distinguished PostgreSQL’s built-in C semantics from externally supplied locale behavior in #3, then asked whether C.UTF-8 should bypass libc and ICU entirely in #6; Tom Lane supported built-in handling for C, POSIX and their dotted variants in #7.
- Jeff submitted a separate two-patch series in #8 implementing those built-in semantics and blocking upgrades of older ICU C/POSIX collations, while acknowledging unresolved compatibility risks for existing libc databases.
- Jeff proposed canonicalizing only newly created locale names to preserve existing libc behavior in #9, but Daniel objected that PostgreSQL would then sort differently from the operating system in #10.
- The ICU discussion shifted toward ordinary provider handling: Joe Conway supported that choice in #12, Daniel demonstrated incompatible sorting and character classification in #13 and #14, and Peter Eisentraut favored an error in #15; Jeff noted the potential impact on initdb with LANG=C.UTF-8 in #16.
- Returning to libc versioning, Thomas proposed limiting version capture to older glibc releases in #17, but Jeff raised evolving character classification in #18, and Thomas recognized that Daniel’s original patch could also warn about some ctype changes in #19.
- Jeff endorsed versioning everything passed to an external provider and described LC_COLLATE=C with LC_CTYPE=C.UTF-8 as an explicit user choice in #20; Daniel broadened the suspected compatibility problem to Red Hat-derived distributions in #21.
- Jeff committed f3a01af29b, removing forced ICU C/POSIX mapping and using ordinary canonicalization and validation; although this thread contributed discussion, the implementation came from #534326 in thread #47790, restoring an earlier landing that had been reverted for an ICU-version-dependent test failure, omitting that test and retaining regression validation adjustments.
- At the thread’s close, Thomas accepted the rationale for Daniel’s original libc versioning patch and planned a commit message targeting beta2 in #22; that fix remained pending in the supplied record, and the broader built-in semantics and pg_upgrade series had not landed.