model-writtenrun 6147 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Jacob Champion opened the thread with the announcement that libpq would exercise protocol negotiation by default during PostgreSQL 19 development, with the experiment scheduled for removal before RC1; crake soon exposed failures connecting to old servers #1 #2.
- Tom Lane considered dropping very old upgrade support, but Jelte Fennema-Nio pointed out that future default protocol changes would have the same problem, and Jacob established that only PostgreSQL 11 supported negotiation throughout its release line, motivating a permanent pg_upgrade fix #3 #4 #5 #6.
- Jacob submitted a fix that forced protocol 3.0 for old servers, but Tom's testing uncovered a separate failure in the buildfarm's comparison dump before pg_upgrade ran; Tom supplied a complementary buildfarm patch and reported a clean run with both fixes #7 #8 #9 #11.
- Tom also identified that libpq versions before 18 reject max_protocol_version, prompting Jacob to add a runtime library-version guard in the PostgreSQL 14 backport and test it against both old and new libpq libraries; the backpatch rationale covered distributions linking older utilities against newer libpq #8 #11 #14.
- Nathan Bossart proposed raising pg_upgrade's minimum supported server version in a future release, while Jacob explained that fixing protocol negotiation now would remain useful until all pre-11 servers aged out of support #13 #15.
- Jacob discovered that ClusterInfo retained only major-version information, making the detailed minor-release cutoffs misleading; Tom endorsed simply forcing protocol 3.0 for every server before PostgreSQL 11, avoiding additional version-detection complexity #16 #17 #18.
- Jacob submitted revised patches for all six supported branches after testing upgrades from PostgreSQL 9.2, 10.1 and 11.0 with normal and greased libpq, and Tom approved them with a minor comment correction for older branches #19 #20.
- Jacob announced that the patches had been pushed after release tags appeared, landing the development-branch fix as a60a103386 from #19 and backpatching through PostgreSQL 14; the announced final wording correction affected only older-branch variants #21.
- Andrew Dunstan supplied an alternative buildfarm fix using PGMAXPROTOCOLVERSION, and Jacob and Tom requested limiting that setting to the comparison dump so it would not conceal pg_upgrade regressions #22 #23 #24.
- Tom posted the narrower environment-variable scope and reported clean cross-version tests on every branch, which Andrew independently confirmed; the thread ends with pg_upgrade committed and the buildfarm adjustment tested but without an explicit landing announcement, followed by Álvaro Herrera's advice on propagating small comment corrections across branches #25 #26 #27.