model-writtenrun 5032 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Aug - Nov 2019 · 4 entries
- Peter Eisentraut proposed database-wide ICU collation support in August 2019, using a global collator and new initdb and CREATE DATABASE options, while excluding nondeterministic defaults because they would break pattern matching and system queries #1.
- Andrey Borodin exposed configurations that allowed database creation but failed on connection #2, and Daniel Verite demonstrated that failing to initialize libc LC_CTYPE correctly broke full-text search #5.
- Marius Timmer rebased the patch #6, and Thomas Munro extended default-collation version detection while questioning the global terminology and highlighting missing dependency tracking #7 #8.
- Peter paused the work because making text search collation-aware would require broad changes, while Daniel warned that extensions using libc locale functions would remain affected #9 #10.
Dec 2021 - Mar 2022 · 12 entries
- Peter revived the feature in December 2021 with separate ICU and libc locale fields, abandoning the attempt to convert all legacy locale consumers to ICU and changing locale strings from name to text #12.
- Julien Rouhaud accepted the dual-locale design but found missing upgrade checks, confusing option inheritance, template compatibility gaps, and incorrect null handling #18.
- Daniel proposed broader metadata and references to pg_collation, but the discussion retained explicit locale fields and deferred nondeterministic database defaults because existing pattern-matching restrictions would break system functionality #23 #30 #32.
- Peter and Julien found no convincing performance penalty from calling pg_newlocale_from_collation() for the default locale, leading to the standalone simplification #33, which Julien tested and which landed unchanged as b99ccd2cb2 #19 #22 #37.
- Peter extracted the locale-field conversion into #39, Julien verified tests and documentation, and it landed with a catalog-version bump as 54637508f8 #40 #43.
- The remaining feature returned after those prerequisites, with database collation-version tracking moved to a separate thread and the public terminology changed to locale provider and ICU locale ID #44 #45.
- Julien's detailed review requested clearer options, actual sorting tests, and simpler DefineCollation() control flow; the standalone cleanup #47 landed unchanged as e94bb1473e after his endorsement #46 #48.
- Peter's final submission added an ICU test suite, required an explicit ICU locale in initdb, and improved error messages and documentation, although an unintended CREATE DATABASE fallback survived #49.
- Robert Haas questioned mismatched libc and ICU settings, prompting Peter to explain that legacy consumers still needed libc settings and that semantic consistency remained the user's responsibility #50 #53 #57.
- The feature landed from #49 as f2553d4306, incorporating Julien's unexpected-null checks and additional documentation on locale selection and deterministic defaults #51 #58.
- Noriyoshi Shinoda immediately supplied the missing daticulocale catalog documentation, which Peter confirmed applying, while Andres Freund reported sorting-test failures on an older ICU installation #59 #60 #62.
- Andres then identified an uninitialized-variable compiler warning and fixed src_locprovider directly in 97bddda61b, following his report and proposed initialization in #64.
Jun 2022 · 1 entry
- In June, Justin Pryzby found that pg_upgrade treated PostgreSQL 15 as lacking global ICU support; after confirmation from Julien and Michael Paquier, Peter changed the version boundary to less than 1500 in fa06a34d14 #65 #66 #69 #70.
Aug - Nov 2022 · 12 entries
- In August, Marina Polyakova demonstrated missed locale-provider mismatches during upgrades, a CREATE DATABASE assertion failure, and inconsistent ICU-locale fallback behavior, supplying inline fixes for the first two problems #72.
- Julien refined the assertion fix to inherit the ICU locale only after choosing the provider, and Marina supplied that inline revision along with separate diagnostic instrumentation showing old-ICU locale fallback and version discrepancies #73 #74.
- Peter committed the inline pg_upgrade cluster-version correction as f0b013b0fc and the revised CREATE DATABASE inheritance fix with a regression test as 36f729e2bc, confirming both fixes in #75.
- Peter concluded that deriving ICU_LOCALE from LOCALE was an unintended leftover and removed it in 8b808f189f, implementing the discussion in #75.
- After Andrew Bille found that pg_dump lost colliculocale, Marina supplied #78; Michael and Julien reviewed it, and Peter committed it with minor test cleanup as 396d348b04 #79 #80 #81.
- Marina next proposed restoring CREATE DATABASE's rejection of combined general and category-specific locale options #83, but Justin advocated relaxing createdb instead, and Marina supplied the accepted alternative inline #84 #85.
- That inline alternative landed with corresponding documentation changes as da5d4ea5aa, leaving the restrictive attachment unused and making createdb consistent with CREATE DATABASE and initdb #86.
- Marina demonstrated that ICU databases could be created with unsupported encodings and submitted validation checks, then reordered checks and output following Kyotaro Horiguchi's review #88 #89 #90.
- Peter based c7db01e325 on the smaller first patch, adding a helper, revised diagnostics, tests, and UTF8-default documentation while declining the broader reordering for PostgreSQL 15 #92 #94 #95.
- Marina also found that specifying an ICU locale failed when the provider should be inherited from the template; Peter supplied #104, Marina tested it and added a regression case, and it landed as e59a67fb8f #93 #105 #106.
- Marina continued the deferred check-order work in #107 and #109, but Peter declined additional PostgreSQL 15 changes without a faulty-behavior justification, leaving the broader reorganization uncommitted in this thread #108.
- The final investigation found that ICU-initialized clusters broke SQL_ASCII ECPG installchecks and defeated pg_regress --no-locale; Marina proposed selecting libc explicitly, and Michael confirmed the --no-locale problem in the last reply, leaving that fix awaiting action #110 #111.