model-writtenrun 6262 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Apr - May 2021 · 7 entries
- Bryn Llewellyn opened the thread with interval literals whose results depended on the order of negative years and positive months, also asking why fractional years stopped at months while fractional months spilled farther #1.
- Bruce Momjian traced the discrepancy to truncation during compound assignment and proposed rounding fractional-year conversions with rint(), initially changing only backend interval parsing #2.
- The discussion expanded into spill semantics: Bryn demonstrated existing cascading behavior, while Bruce proposed stopping at the next internal unit and extended his patch to ECPG, changing fractional weeks and months to stop at days #4 #8.
- Zhihong Yu tested the changes and identified two missing ECPG ISO-8601 conversions, which Bruce added to the next patch #16 #21 #22.
- Justin Pryzby showed that cascading fractional weeks and months was intentional and documented, challenged changing it near the release deadline, and requested clearer documentation; Bruce revised the prose and moved the work to PostgreSQL 15 #23 #25 #30.
- Bryn and Tom Lane questioned the rationale for approximate conversions at all, while Bruce defended fractional internal units as useful approximations; Zhihong tested more examples and restored his endorsement of the then-current patch #36 #38 #39 #43.
- After the release-cycle delay, Bruce refreshed the patch in May and resumed substantive work in July, discovering that fractional days were still truncated instead of rounded and replacing AdjustFractDays() with direct rounding #45 #54.
Jun - Aug 2021 · 11 entries
- Zhihong requested a regression test for fractional-month rounding, which Bruce added, and then raised the inconsistency between interval input and interval multiplication #55 #56 #57.
- Dean Rasheed demonstrated that rounding half a week to four days was a regression, and Tom urged preserving established behavior except for clearly broken cases #59 #60 #62.
- Bruce partially reversed course by restoring fractional-week cascading while retaining month-to-day rounding, then added negative full-month carry handling and factored the month logic into a helper #64 #69.
- Bryn asked how multiplication and division related to the proposed input rules and ultimately advocated fixing only the original defect, but Bruce briefly expanded the patch to change both arithmetic operators #70 #72 #74.
- Tom rejected the arithmetic expansion as an unsupported compatibility change, and Bryn reiterated that his request was for a narrow fix rather than new multiplication semantics #75 #77.
- Bruce then removed all changes to spill destinations and arithmetic operators, leaving only rounding when larger units become months and documentation of the existing conversion rules #81.
- Robert Haas supported a conservative fix, while Tom explored rejecting unrepresentable fractional inputs; those broader input restrictions were discussed but did not enter the final patch #82 #83 #84 #85.
- Tom reviewed the documentation and requested explicit truncation wording for older branches, microsecond terminology, and clearer examples; Bruce incorporated those corrections into the final substantive revision #90 #91.
- Robert questioned the scope and reported trouble applying the attachment, prompting an identical resend; Tom supported rounding over truncation but remained doubtful that it established order independence for every input #92 #94 #95 #96 #98.
- Bruce committed the complete final attachment unchanged as 95ab1e0a9d, making larger-unit conversions round to months on master and documenting that behavior #96 #99.
- The documentation-only portion of the same attachment became 9a01a2de85, with “rounded” changed to “truncated” for older branches; the final message confirms both applications, leaving the broader spill and arithmetic redesign abandoned and the general order-independence concern unresolved in the discussion #96 #95 #99.