model-writtenrun 3558 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Sep - Nov 2023 · 12 entries
- Jeff Davis proposed a UTEXT type that would normalize input and reject unassigned Unicode code points, combining canonical-equivalence comparisons with fast byte comparison and forward-compatible normalization #1.
- Peter Eisentraut preferred database-wide nondeterministic collations over a nonstandard type, while Robert Haas emphasized that accepting unassigned characters could make later collation updates break index ordering #2 #3.
- Jeff narrowed the immediate goal to internal and SQL functions for detecting unassigned characters, usable by CHECK constraints or extensions, while Nico Williams explored preserving input representation and normalizing during comparison #6 #7 #10.
- The discussion also explored per-column encodings and Unicode-specific types, but Robert and Jeff highlighted the infrastructure cost and insufficient demonstrated demand; Matthias van de Meent subsequently supplied migration and storage use cases without an implementation emerging #26 #28 #44.
- Peter explained that ICU collation versions reflect Unicode-version changes and warned that ICU and PostgreSQL could use different Unicode data, establishing a compatibility question that remained relevant throughout the work #33 #35.
- Jeff's first attachment implemented general-category lookup using generated ranges and binary search, deliberately avoiding an ICU runtime dependency while comparing results against ICU in tests #45.
- After Peter asked what the narrower feature would accomplish, Robert identified CHECK constraints and finding existing unassigned characters as applications, and Jeff added unicode_is_valid(), unicode_version(), and icu_unicode_version() to the patch #47 #49 #50.
- Peter pointed out that unassigned characters are still valid Unicode, so Jeff renamed the predicate unicode_assigned(); Jeff also outlined future database-wide restrictions, normalization, and Unicode-based character classification #52 #53 #55.
- Daniel Verite warned that database-wide restrictions would reject newly assigned characters on older PostgreSQL releases and could abort transactions too late, while Jeff defended optional restrictions as a way to guarantee stable normalization #56 #60.
- Jeff committed the primitive functions from #55 as a02b37fc08, with light naming, documentation, diagnostic, assertion, and test-build cleanup; the proposed UTEXT type and database-wide enforcement were outside this change.
- Thomas Munro reported unresolved unicode_category symbols in the legacy MSVC build, and Jeff and David Rowley independently identified the missing common-source entry, with David reporting that he had supplied the initial repair #62 #67 #68.
- Thomas then found that the source entry had been made conditional on an unrelated OpenSSL setting, and David corrected it to be unconditional in 151ffcf6d8, confirming the fix in the thread #73 #74.
Mar 2024 · 2 entries
- In March 2024, Jeff submitted a separate STRICT_UNICODE database option that rejected unassigned code points, supported disabling the restriction later, and integrated database creation, dumping, and client tools; whether enforcement should also account for ICU's potentially different assignment data remained open #75.
- Jeff accepted deferring STRICT_UNICODE to PostgreSQL 18, explaining that planned case-folding support could provide another reason to restrict inputs to assigned characters #76.
Dec 2024 · 1 entry
- Jeff ultimately withdrew STRICT_UNICODE for lack of interest in December 2024, leaving the Unicode primitive functions and MSVC correction landed but database-wide enforcement uncommitted #77.