model-writtenrun 7242 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Apr - Jun 2023 ยท 31 entries
- Regina Obe reported changed PostGIS sort results after ICU became the default, and strk demonstrated that a database reporting LC_COLLATE=C could still use different default ordering; the investigation established that the ICU locale was inherited independently of those libc settings #1 #4 #18.
- Andrew Gierth found that initdb --no-locale still selected an ICU locale from the environment, Tom Lane reproduced it, and Jeff Davis committed the one-line libc-provider override as 5cd1a5af4d without an attached patch #12 #13 #24.
- The immediate fix exposed a further inconsistency between --no-locale and --locale=C, while Tom Lane and Robert Haas warned that ICU-by-default threatened compatibility and lacked adequate user documentation #28 #20 #23.
- Jeff proposed removing the surprising conversion of C to en-US-u-va-posix, and Peter Eisentraut supported making users choose either libc's C behavior or an explicit ICU locale #31 #34.
- Jeff expanded the work to give ICU C/POSIX internal bytewise semantics and make LOCALE apply to ICU; Daniel Verite found a broken collation-version comparison and failure on fr_FR@euro, leading to corrected comparisons and an additional legacy-variant conversion patch #38 #39 #40.
- While rebasing, Jeff recognized that redefining existing ICU C collations could invalidate upgraded indexes, and laid out alternative catalog representations or a separate provider to distinguish old behavior from new bytewise semantics #41.
- Jeff separately submitted extensive ICU documentation covering language tags, comparison levels, settings, and examples, while leaving fuller ICU-rules documentation open because its practical use cases remained unclear #42.
- The conversion-removal component of #41 landed as f7faa9976c with uppercase C test inputs, but Tom Lane reported that some ICU versions still performed the conversion themselves; Jeff therefore reverted it as 455f948b0d #43 #44.
- Peter Eisentraut proposed removing the misleading read-only lc_collate and lc_ctype settings and supplied a separate patch replacing regression-test access with pg_database queries #33 #47.
- Jeff's next series introduced a none provider and automatic switching for ICU C locales to avoid upgrade ambiguity, alongside LOCALE propagation, a provider-default GUC, relaxed validation, and a separate collation-encoding fix #49.
- Peter challenged that series' scope, regarded the old libc locale variants as obsolete, questioned weaker validation, and argued that a new provider was invasive so late in the release cycle #51.
- The validation-default component of #49 landed unchanged as 6de31ce446; Jeff later acknowledged having missed Peter's objection and explained that WARNING reduced migration risks from ICU-version differences #76.
- Alexander Lakhin independently found an ASAN stack-buffer overflow when ICU returned an unterminated language name, then questioned the breadth of the defensive checks; Jeff's standalone fix #52 landed unchanged as 1c634f6647 #50 #54 #55.
- Jonathan S. Katz's detailed documentation review led to clearer language-tag and sensitivity explanations, followed by his approval; #58 landed as 1e16af8ab5 with a quote correction and table-row reordering #56 #57 #59.
- Jeff resent the remaining none-provider series after the validation change landed, adding refinements to provider-default selection; Daniel argued that explicit ICU requests should retain ICU semantics and favored selecting libc at the interface instead #64 #65.
- Tom Lane found PDF table overflow and unsupported glyphs in the new documentation, supplied column-width hints, and landed those hints with further table, example, and release-note rendering fixes as b62381d9a2 #66.
- Daniel tested the revised locale handling and demonstrated region loss, C.UTF-8 differences, and libc aliases that ICU did not recognize, reinforcing that ICU could not transparently replace every libc locale #73.
- Jeff then dropped legacy-variant conversion and the none provider, revived direct ICU C/POSIX handling with a pg_upgrade rejection check, and replaced the default-provider GUC with inheritance from the database; the unrelated encoding fix had also been removed from this thread's series #76.
- Peter's follow-up documentation review produced f4a9422c0c directly from discussion, adding table introductions, adjusting parameter capitalization and spacing, and reordering settings; Jeff acknowledged that deeper explanation of ICU's conceptual model remained useful future work #69 #77.
- Daniel supplied a standalone uloc_getDefault() patch to preserve the environment locale's country component, and Jeff tested it across ICU versions before committing it unchanged as ec1264f01e #78 #79.
- After a rebase, Daniel found that the direct C/POSIX handling still omitted initdb, and Jeff added the missing canonicalization bypass while continuing to argue for provider-independent LOCALE semantics #80 #81 #85.
- Joe Conway revived the separate-provider idea, Tom Lane objected to the name none, and Robert Haas suggested builtin; the ensuing naming discussion favored describing PostgreSQL's own implementation explicitly #82 #84 #87 #88 #93.
- Peter's GUC-removal patch #47 landed as b0f6c43716 with unchanged code and a revised commit message addressing Jeff's observation that the settings still represented the server environment #71 #94.
- Peter supported LOCALE propagation, requested removal of redundant createdb-side option expansion, and urged separating the PostgreSQL 16 release fix from new provider features and pre-existing behavior changes #96 #97.
- Jeff revived the earlier separate-provider implementation as builtin, adding documentation, tests, dump support, automatic ICU-to-builtin switching, and special initdb handling for C.UTF-8 #107.
- Daniel demonstrated that treating C.UTF-8 as C broke non-ASCII case conversion and created inconsistent initdb versus CREATE DATABASE behavior; Peter then explicitly rejected the new provider and other substantial behavior changes for PostgreSQL 16 #108 #110 #111.
- Jeff isolated LOCALE propagation into #112 and moved the deferred feature work to the PostgreSQL 17 commitfest; Peter approved the isolated patch and supplied SGML and Perl-test fixups in #114 #113.
- The isolated LOCALE patch #112, incorporating #114, landed as a14e75eb0b; the thread's final message confirms that landing, while the builtin-provider and CREATE COLLATION default work remained deferred #115.
- After this thread ended, 2535c74b1a restored libc as initdb's default from the later attachment #536874 in thread #48030; this thread contributed the earlier rollback arguments from Tom Lane and Robert Haas #20 #23.
- On June 21, 2023, Jeff re-landed the conversion-removal component of #41 as f3a01af29b, omitting the tests implicated by #43 and accepting that ICU's treatment of C varies by version, with canonicalization failures governed by icu_validation_level.
- The builtin-provider proposal in #107 was a precursor to the later work in thread #48860, whose March 2024 submission #556429 supplied the committed implementation for 2d819a08a1; that landing belongs to the later thread and excludes this thread's automatic provider switching and CREATE COLLATION default changes.