model-writtenrun 5917 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Oct 2019 · 9 entries
- Hans Buschmann reported crashes and corrupted-looking values after upgrading PostgreSQL 11 to 12, affecting a user table created from information_schema queries #1.
- Tomas Vondra reproduced the failure and traced it to sql_identifier changing its underlying type from varchar to name, making existing stored values binary-incompatible #3 #8.
- Tom Lane proposed rejecting affected upgrades; Tomas considered automatic conversion, but Tom argued that this would complicate pg_dump and silently change user column types, and Andres Freund supported requiring user intervention #9 #10 #11 #12.
- Tomas submitted the initial safety check, and Tom identified missing domain coverage and tested recursive-query planning on old releases; Tomas extended the design to handle arbitrarily nested domains and composite types #13 #14 #16 #17.
- The revised patch checked only objects with storage, and Tom's final review added defensive handling of dropped composite attributes; it landed with those small edits as 0ccfc28223 from #18 #19 #21.
- Tomas then submitted analogous repairs for line and unknown checks, fixing both missed indirect uses and false positives for unused composite types; the two parts landed separately with comment refinements as 8d48e6a724 and a524f50d0f from #23, as confirmed in #26.
- Justin Pryzby proposed preventing future binary-format regressions with a table containing every core type and a comparison across pg_upgrade, establishing the testing work that would resume later #24.
- The line-check backpatch exposed missing RELKIND definitions on older branches; Tom diagnosed the missing header, and Tomas fixed it without an attached patch in e09ab32a22 #27 #28 #30.
- Bruce Momjian supplied and reported committing a small follow-up that aligned the new sql_identifier error-list heading with the existing clearer format #31.
Dec 2020 - Jan 2021 · 3 entries
- After fourteen months, Justin returned with a many-types regression table and test.sh repairs, demonstrating that the test reproduced the original corruption when the upgrade guard was disabled #32.
- Justin excluded optional XML support, added OID-related coverage, and rebased the series; Peter Eisentraut and Bruce requested better explanations, and the next revision separated infrastructure adjustments and added snapshot types #33 #34 #35 #36 #37.
- Andrew Dunstan explained that buildfarm cross-version testing used its own machinery, and Tom subsequently proposed a shared cleanup SQL script in the PostgreSQL tree, objected to losing existing test coverage, and challenged the completeness query's type exclusions #41 #43.
Apr 2021 · 1 entry
- Justin added the shared-script proposal, range and multirange coverage, and explicit type exclusions; Jacob Champion then identified missing BRIN types and discrepancies between the script and buildfarm cleanup, prompting corrected predicates and clearer version-variable names #44 #45 #47 #49.
Sep - Dec 2021 · 8 entries
- A parallel buildfarm-client effort normalized dump differences and counted only changed lines; Andrew supplied the patch, verified Windows behavior, and reported committing it in the buildfarm repository #51 #52 #53.
- Michael Paquier extracted the immediate test.sh repairs, adopting catalog-driven OID removal and conditional dump options while preserving WAL-size and permission coverage; he deliberately left older initdb-option incompatibilities unresolved #55.
- Michael tested and committed that attachment unchanged as fa66b6dee0, and Justin resent the remaining type-table and SQL-extraction work #56 #57.
- Justin clarified that the completeness query would detect newly introduced core types; Michael accepted that design, improved its comments and formatting, and committed his final attachment unchanged as 835bcba8b8 #57 #58 #59 #61.
- The newly landed table immediately failed on a buildfarm animal whose username contained a hyphen; Tom reported the ACL parsing error, and Michael identified missing identifier quoting in the aclitem expression #63 #64.
- Andrew endorsed sharing cleanup SQL with the buildfarm, and Justin resent the remaining extraction with additional older-version cases; Michael requested an absolute script path for VPATH builds and closer alignment with buildfarm object recreation #60 #62 #65.
- Michael consolidated the cleanup into upgrade_adapt.sql with ordered version checks, removed unnecessary operations, and prepared it for backpatching; it landed from #66 with minor comments and removal of the putenv cleanup as 0df9641d39.
- At the thread's end, the core SQL extraction was committed and backpatched through PostgreSQL 10, while buildfarm adoption still needed validation and broader contrib-module upgrade testing remained open, including how to preserve separate test databases without excessive cost #67.