model-writtenrun 672 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Jeff Davis opened the thread with #1, proposing provider-specific implementation files and method tables so extensions could substitute locale implementations, including different ICU versions, while leaving provider identity primarily a catalog concern.
- In #4, Jeff extended the design to remove runtime provider identity, hide provider data behind an opaque pointer, relocate ICU includes, and hook version reporting; his initial performance tests found no clear overhead, but catalog-driven providers and a demonstration extension remained open work.
- The ICU and libc file extractions from #4 landed as f244a2bb4c and 66ac94cdc7, both with only comment edits.
- Andreas Karlsson reviewed the remaining work in #5, requesting a builtin-provider file, combined case-mapping and ctype tables, consistency assertions, and investigation of classification overhead; he also rebased the remainder, whose locale-construction extraction landed with one comment removed as 3aa2373c11.
- Jeff and Andreas explored the extension interface in #6, #7, and #8: without a provider catalog, extensions would have to repurpose existing provider identities, while database-default collations and initdb complicated a cleaner catalog design.
- Jeff incorporated the builtin-provider file and combined ctype table in #9, reporting indistinguishable small-string timings but possible classification overhead on large strings and asking whether individual property callbacks would be preferable.
- Andreas expanded the series in #10 with provider-version, ICU validation, and locale-lookup cleanups, and approved the initialization refactor; Jeff committed that refactor with a declaration cleanup and rebase adjustments as 1ba0782ce9.
- Jeff combined their work in #11 and separated case-conversion relocation from method-table dispatch; Andreas's ICU feature-check extraction landed unchanged as 7167e05fc7.
- In #12, Jeff questioned the growing collection of provider-specific entry points, suggested a provider-level method structure, and reported a createdb test failure from the proposed ICU validation cleanup.
- The standalone case-conversion extraction from #11 landed unchanged as 86a5d6006a, and Andreas subsequently supplied the one-line createdb test fix in #13.
- The provider-version extraction from #11 landed with conditional-dispatch edits as 4f5cef2607, followed by the collation method table with an added Windows-specific table as a2f17f004d.
- Jeff posted a four-patch remainder in #14, replacing char_properties with individual classification callbacks; the extension hooks were absent, and he said the proposed namespace lookup cleanup also needed to address DefineCollation before proceeding.
- Worst-case tests in #15 exposed regex and single-byte ILIKE slowdowns, prompting Jeff to ask whether the cleaner method-table design justified them; #16 and #17 kept the work rebased while that tradeoff remained unresolved.
- After a four-month gap, Jeff renewed the proposal in #18, arguing that multiple provider-library versions justified the method tables and that the measured performance cases were unlikely to be a practical obstacle.
- Peter Eisentraut accepted the ctype table and provider-field removal in #19, suggested future unification of C-locale special cases, and opposed opaque provider data without a clearer justification; Jeff agreed in #20 to hold that patch and its dependent ICU-header cleanup.
- The first two patches of #18 landed unchanged as 5a38104b36 and 8af0d0ab01, and Peter closed the commitfest entry in #21 with the opaque-data and header patches withdrawn.
- Noah Misch reopened the correctness discussion in #22, demonstrating that 5a38104b36 made a UTF8 C-collation ILIKE query return false instead of true and forwarding additional review findings.
- Jeff supplied four fixes in #23: restrict lazy C-locale ILIKE matching to single-byte encodings, remove an obsolete comment, restore character-range guards needed on Windows, and add missing unsigned-char casts; he rejected several other reported issues as unsubstantiated.
- At the last message, #24, Noah accepted Jeff's issue triage and attached the underlying review for the archive, while explicitly saying he had not read the patches; the fixes were still pending at that point.
- The following day, the ILIKE correction landed as d34ce773d2, differing from the supplied #23 attachment only in the regression literal's encoding, and the comment deletion landed unchanged as 3ab2abc949.
- The guard restoration and unsigned-char casts from #23 also landed unchanged as dbbeafe61a and e6e08dc554, completing the four-fix round with backports through PostgreSQL 19.