model-writtenrun 9612 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Peter Eisentraut proposed an insensitive collation flag that would let ICU equality semantics survive PostgreSQL's bytewise shortcuts, with executor collation propagation and sort-key hashing, while identifying unfinished name, foreign-key and bpchar support #1.
- Daniel Verite proposed equality-oriented naming, Tom Lane objected to calling it natural, and Peter Geoghegan proposed Unicode's deterministic terminology; Eisentraut adopted deterministic and expanded name, foreign-key, pg_dump and psql support #2 #3 #4 #5 #7.
- Daniel's tests showed that grouping could return different byte representations of collationally equal strings, which Eisentraut considered valid because no preferred representative was defined; discussion with Andreas Karlsson also left normalization on input as a separate feature #8 #11 #13 #10 #12.
- Daniel found that building a unique index failed to reject collational duplicates, prompting a sort-support fix, then confirmed that fix and exposed DISTINCT aggregates ignoring the collation #14 #17 #18.
- Eisentraut extended collation propagation through aggregates, subplans and logical replication, completed additional string operations, and rejected unsupported pattern matching; Daniel's next review prompted similar rejection for substring searches and broader documentation #21 #22 #23.
- After a rebase, Peter Geoghegan questioned the ICU-only scope, suggested splitting the patch, and identified an unnecessarily disabled byte-equality optimization; Eisentraut retained the combined patch and ICU-only scope, citing test coverage and libc limitations, and deferred database-wide ICU support #24 #25 #26.
- Eisentraut restored safe equality shortcuts, added citext guidance and improved test portability, but reported that ICU versions older than 54 lacked the locale customization needed by several tests #27.
- Daniel demonstrated that LIKE prefix optimization could bypass the new rejection checks; Eisentraut fixed match_pattern_prefix() and also prohibited incompatible pattern operator classes, adding tests and resolving merge conflicts #28 #32.
- The older-ICU discussion moved from alternate locale syntax and a possible separate test file to Daniel's demonstration that ucol_setAttribute() provided the required behavior; Eisentraut then proposed a separate compatibility parser #29 #30 #31 #33.
- Eisentraut posted the compatibility parser and customization tests as an independent series, combined them with a comment correction in b8f9a2a69a, and rebased the main patch after that prerequisite landed #34 #35.
- The main patch from #35 landed as 5e1963fb76 with a catalog-version bump, expected-output correction and formatting; Eisentraut announced that the main feature was committed #36.
- Two years later, Jim Finnerty reopened the unsupported-LIKE question and explored lowercasing and explicit deterministic collations, while Daniel suggested an ICU prefix-range technique and clarified that catalog name comparisons do not use the database's default collation #37 #38 #39 #40 #41.
- Jim subsequently reported apparent multibyte wildcard failures, Daniel diagnosed a client-encoding mismatch, and Jim confirmed the diagnosis and demonstrated an explicit deterministic-collation ILIKE workaround; both original changes remained landed, with native nondeterministic LIKE support still unresolved and no further patch submitted #42 #43 #44.