model-writtenrun 5680 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Peter Eisentraut opened the thread with the landing announcement for global ICU locale selection, explaining that databases still retain libc locale settings for legacy code alongside the new ICU locale field #1.
- Michael Paquier reported an invalid-locale test failure, Julien Rouhaud noticed missing collator cleanup, and Thomas Munro identified the old ICU version as a clue; Julien confirmed that ICU below 54 requires explicit attribute handling #2 #3 #4 #6.
- Julien's first patch added backend validation and collator cleanup while copying attribute parsing into initdb, and he reported that it fixed the failures with ICU 50 #7.
- Tom Lane independently confirmed the CREATE DATABASE memory leak with Valgrind and found a separate ICU test failure under the C locale, which he addressed by forcing UTF8 encoding; Peter agreed with that test adjustment #8 #9 #11.
- Peter proposed letting backend initialization validate the locale instead of duplicating validation in initdb, Tom supported that approach, and Andres Freund favored keeping ICU-specific handling inside pg_locale.c #10 #12 #13.
- Julien's replacement patch removed initdb's ICU validation and header dependency, updated the expected backend error, and changed test locales to syntax supported by old ICU; he reported check-world passing with both older and newer ICU versions #15.
- Christoph Berg raised a separate usability concern that the added locale columns made psql database listings too wide, proposing a consolidated locale column #16.
- Peter committed the second attachment as 3a671e1f7c, additionally centralizing unsupported-ICU handling in the helper and removing initdb's ICU link dependency; he announced completion, and Andres subsequently confirmed that the change fixed his CentOS 7 meson build #15 #17 #20.
- Peter acknowledged the listing-width concern, Christoph suggested leaving locale details only in verbose listings, and Tom preferred a summary locale column in ordinary listings while keeping column order consistent #18 #21 #22.
- Christoph later demonstrated that psql incorrectly repeated the libc collation in the ICU Locale column when connected to a pre-v15 server, identifying the faulty expression and suggesting an empty value #23.
- Euler Taveira proposed NULL to match non-ICU databases on v15, Tom agreed, and 7129a9791e implemented that one-line discussion-derived fix independently of the unresolved listing redesign #24 #25.
- The compatibility and display correctness fixes had landed, but the broader listing-width proposal remained unresolved when Christoph renewed his request for more concise default output and offered to prepare a patch #26.