model-writtenrun 9257 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Apr - Aug 2019 ยท 15 entries
- Tom Lane announced the tzdata 2019a update, which made UCT an alias of UTC, and Christoph Berg subsequently reported that initdb selected unexpected UCT spellings and broke Docker users' tests #1 #2.
- Andrew Gierth traced the behavior to shortest-name and alphabetical tie-breaking and proposed preferring canonical names, while Tom Lane questioned whether installed timezone data reliably identified such names #3 #5 #7.
- Andres Freund's tests showed that PostgreSQL 12 could preserve the spelling from an /etc/localtime symlink, whereas older branches searched equivalent zones and selected UCT; Christoph later confirmed that his reported PostgreSQL 12 discrepancy was probably a typo #10 #14 #25.
- Tom proposed backpatching the existing symlink implementation developed in thread #36500, Christoph confirmed its useful behavior on Debian, and Tom agreed to proceed; this reused earlier work rather than introducing a matching attachment in this thread #16 #17 #18.
- Andrew demonstrated that symlink detection did not fix systems with a missing or copied /etc/localtime, including a default FreeBSD installation, and submitted explicit UTC and Etc/UTC preferences in #24 after Tom initially opposed hard-coded preferences #19 #20.
- Andrew committed the UTC preference fix, prompting objections from Andres Freund, Robert Haas, and Tom about proceeding without consensus shortly before a release; Tom's buildfarm analysis nevertheless found the intended UCT-to-UTC changes without unrelated selection changes #26 #30 #31 #32.
- The buildfarm analysis also exposed localtime as an undesirable timezone name, leading Tom to propose negative preferences for localtime and posixrules because these pseudo-zones were uninformative, unstable, and potentially removable #32 #34.
- Andrew proposed a broader preference hierarchy using zone.tab or zone1970.tab and geographic names, but Tom opposed expanding the policy because of portability, behavioral compatibility, and political naming concerns #35 #51 #55.
- Reverting the tzdata update was rejected as a solution because installations using system timezone data would encounter the upstream change anyway, and Tom explicitly declined to seek a revert of the UTC preference fix #46 #55.
- Tom submitted an independent Factory cleanup in #56, proposing to remove exclusions from initdb and pg_timezone_names because modern tzdata used the short abbreviation -00 instead of a long diagnostic string.
- Stephen Frost and Robert Haas supported aspects of preferring upstream geographic names, Andrew supplied evidence that his proposal reduced naming discrepancies, and Thomas Munro reported FreeBSD's addition of zone1970.tab; the broader policy still lacked agreement #59 #64 #66 #67.
- Tom narrowed the immediate work to pseudo-zone preferences and Factory handling, then posted both patches in #68 after a user complaint in thread #40976; discovering that FreeBSD retained an old long Factory abbreviation caused him to replace blanket filter removal with a length limit and retain initdb's exclusion #60.
- The localtime/posixrules attachment from #68 landed unchanged as 3754113f33, and the Factory attachment landed unchanged as 8ab66081ca, allowing modern short abbreviations while rejecting those longer than 31 characters and simplifying zic output.
- Shay Rojansky subsequently emphasized that localtime prevented clients from identifying the server's timezone, and Tom confirmed that the committed fix addressed future initdb choices but did not repair existing configurations #69 #70.
- The thread ended with the two focused fixes landed and broader geographic-name preferences unresolved; Andres suggested resolving existing localtime settings through symlinks, but Tom questioned its usefulness across platforms and no implementation followed #72 #73 #74 #75.