model-writtenrun 2447 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Nov 2024 · 4 entries
- Peter Eisentraut proposed updating the generated Unicode data to version 16.0.0, explicitly reopening discussion because earlier debate had questioned how such updates should be handled #1.
- Joe Conway tested the update and demonstrated changed UPPER() and INITCAP() results, while Laurenz Albe opposed updates that could require rebuilding indexes using the builtin provider #2 #3.
- Peter argued that Unicode updates must continue for existing functionality, and Michael Paquier verified that regenerating unaccent.rules produced no changes #4 #5.
- Jeff Davis proposed extension hooks for preserving older character semantics and later suggested bundling multiple Unicode versions; Peter preferred runtime selection and emphasized that normalization, SCRAM, and psql formatting also depend on these tables #6 #8 #9 #13 #14.
Jan - Apr 2025 · 17 entries
- Jeff explained that the builtin provider guarantees stability within a PostgreSQL major version, while cross-version immutability problems also affect normalization, expression indexes, and partitioning independently of that provider #16.
- Following the pre-FOSDEM developer meeting, Jeff submitted a best-effort pg_upgrade check that searches stored expression trees for Unicode-dependent functions and operators, warns about affected relations, and deliberately omits pre-17 clusters and dependencies hidden inside user functions #17.
- Peter endorsed the warning check's structure, and Jeff combined both patches into a rebased series, adding warning-status output and changes to normalization test expectations #18 #20.
- Nathan Bossart reviewed the UpgradeTask usage, identified task allocations stranded by early returns, and questioned whether warnings were sufficient; Jeff retained warnings because detected dependencies can be false positives #22 #23.
- After Nathan requested remediation guidance, Jeff supplied worked examples showing how to inspect and rebuild indexes, resolve uniqueness and check-constraint problems, and move rows into the correct partitions after upgrading #24 #25.
- Peter found the normalization-test changes unjustified, prompting Jeff to discard his faulty Unicode rebase, rebuild from Peter's original patch against the optimized case-mapping implementation, and resend both patches with the pg_upgrade logic unchanged #26 #38.
- The policy debate continued: Jeremy Schneider, Laurenz, and Joe favored preserving older behavior, while Tom Lane argued that timely Unicode updates reduce the chance of users storing newly assigned characters before PostgreSQL understands them #27 #28 #29 #30.
- Jeremy suggested treating changing character semantics as STABLE, but Jeff and Tom argued that this would prevent useful expression indexes without solving the underlying need to track dependencies and invalidate derived data #33 #35 #37.
- Jeff proposed revisiting optional rejection of unassigned code points, while Robert Haas advocated overlapping support for old and new versions so users could schedule semantic changes independently of PostgreSQL upgrades #39 #43 #47.
- Jeff outlined future provider catalogs, separate collation and character-semantics dependencies, extension support, and repair commands; these remained proposals beyond the submitted Unicode update and warning check #49.
- Robert rejected automatic background index repair as potentially disruptive and described an alternative based on strong versioning and user-controlled migration; subsequent discussion explored defaults for new objects and metadata-based restrictions on accepted code points #55 #56 #58 #59 #60.
- Peter reiterated that broader versioning work should not block existing Unicode-update policy, Robert explicitly supported proceeding, and Jeff warned that prerequisite refactoring would likely slip beyond the current development cycle #61 #62 #63.
- Jeremy noted that rejecting unassigned code points could reject newly introduced emoji before databases were updated, leaving that proposed safeguard's usability unresolved #66.
- Peter committed the Unicode-update component of #38 as 82a46cca99, with only rebase differences, and announced that Jeff should land the pg_upgrade check next #67.
- Jeff committed the unchanged pg_upgrade component of #38 separately as b81ffa13e3, preserving warnings rather than making detected dependencies fatal.
- After Coverity flagged the previously identified allocation leak, Nathan posted an inline fix delaying task creation until after the early returns and committed it as de48056ec7 #68 #69.
- At the final confirmation, both submitted features and the memory-leak fix had landed, while multiple-version support, stronger dependency tracking, and improved migration tooling remained future work #69 #49 #64.