model-writtenrun 4737 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Jeff Davis opened the thread with a broad collation rework covering provider-independent comparison and transformation APIs, locale internals, replaceable ICU/libc libraries, and hook tests, while leaving the design of core multiple-library support open #1.
- Early style feedback from Ted Yu prompted John Naylor to clarify that the existing return style and NUL terminology were acceptable; Jeff subsequently supplied platform fixes and expanded hook tests, reporting passing CI while keeping libc replacement experimental #2 #3 #4 #5.
- Peter Eisentraut objected that refactoring, multiple ICU libraries, and hooks were too much to assess together; Jeff explained the hooks' testing and version-lockdown uses, then agreed to focus this thread on refactoring and move provider-library work elsewhere #6 #7 #8 #11.
- Peter Geoghegan argued that comparison and transformation belong in one patch and that deterministic tie-breaking belongs with callers; Jeff implemented both changes, retained ICU's NUL-terminated comparison path, and posted a reduced two-patch series #12 #13.
- After Peter Geoghegan requested performance evidence, Jeff found no significant hash-aggregation penalty from the non-inline accessor and traced surprising sorting gains partly to differing compiler optimization levels; abbreviated keys were sometimes slower in his tests #14 #15 #16.
- Jeff added separate GUCs to make abbreviated-key experiments reproducible and reported an approximately five-percent ICU sorting improvement, while making clear that the GUCs needed independent support before commitment #17.
- Peter Eisentraut's detailed review questioned the internal-header rearrangement and repetitive transformation-size checks, but supported extracting pg_locale_deterministic(); Jeff dropped the larger locale rearrangement, split out the accessor and USE_ICU cleanup, and moved the testing-only GUCs last #18 #20 #21.
- Jeff defended caller-managed transformation buffers as necessary for flexible allocation and explained why size-consistency checks remained with callers; Peter Eisentraut endorsed the revised arrangement without claiming a complete reread or supporting the GUC patch #21 #23.
- Jeff committed the comparison/transformation component of #21 as d87d548cd0, revising ICU length types and transformed-length semantics while explicitly preserving the existing inclusion of the terminating NUL in nondeterministic hashes #24.
- The accessor component of #21 landed unchanged as 6974a8f768, followed by its USE_ICU cleanup as e0b3074e89, rebased over the revised hash code.
- The thread ended with the three core refactoring components committed; the GUCs remained testing-only, the internal-header rearrangement was deferred, and provider hooks and multiple-library support had been redirected to another thread #13 #21 #24.