model-writtenrun 2980 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tom Lane traced copperhead's upgrade-test failures to pg_init_privs ACLs referencing dropped roles, concluding that missing pg_shdepend records allowed DROP ROLE to leave dangling references #1.
- Noah Misch connected the failure to an older discussion, and Tom distinguished genuine grants issued by extension scripts from the disputed recording of default privileges #2 #3.
- Tom discovered that Meson disabled test_pg_dump in installcheck mode, explaining why other upgrade-testing animals had not exposed the problem, and proposed removing that restriction alongside the backend fix #5.
- Daniel Gustafsson supplied a proof of concept excluding default privileges from recorded initial ACLs, but Tom explained that the buildfarm failure required tracking legitimate initial grants' role dependencies instead #7 #8.
- Tom's replacement approach introduced SHARED_DEPENDENCY_INITACL and direct pg_init_privs cleanup during DROP OWNED; subsequent revisions added regression coverage, used aclexplode for stable role-name substitution, and improved comments after Daniel tested and reviewed the patch #9 #10 #11 #12 #13.
- The dependency fix landed from the final April attachment with only an added catalog-version bump as 5342874039, followed by the separately discussed Meson restriction removal as b0c5b215da #12 #14 #15.
- David G. Johnston suggested identifying the bootstrap superuser by its fixed OID, while Tom defended using current_user because the test might run under another superuser, a distinction David accepted #16 #17 #18.
- Hannu Krosing reopened the issue with a reproducer showing that REASSIGN OWNED followed by DROP ROLE still left dangling initial privileges and generated unrestorable dumps #19.
- Daniel and Tom reconsidered omitting INITACL dependencies for object owners, favoring explicit dependencies over repeated catalog scans, while questioning how ownership changes should interact with pg_init_privs as a historical record #21 #22 #23 #24.
- Robert Haas proposed replacing old-owner references with the new owner, prompting discussion of extension ownership, member-object ownership, and pg_dump's lack of information needed to reproduce those changes reliably #25 #26 #27 #28.
- Hannu supplied a separate patch making REVOKE tolerate missing roles, but Tom and Daniel rejected changing general REVOKE behavior to conceal invalid ACLs; Hannu then proposed filtering invalid initial privileges in pg_dump, which Robert supported #32 #33 #34 #35 #36.
- Robert emphasized that preventing dangling role references was essential, while fully preserving extension-member ownership would require broader infrastructure that was difficult to introduce as a late-cycle or backpatched fix #37.
- Tom posted a follow-up that recorded INITACL dependencies for every non-pinned referenced role and made REASSIGN OWNED substitute role references, while leaving ALTER OWNER unchanged and deferring pg_dump's broader ownership limitations #38 #39.
- Daniel endorsed the follow-up, requested consistent rejection of null initial ACLs, and recommended removing unused owner-ID lookups; Tom accepted those changes and chose a separate commit for the latter cleanup #40 #41 #42.
- The follow-up landed as 35dd40d34c from #38, with additional null assertions, empty-ACL handling, dependency-helper cleanup, and a catalog-version bump.
- The separately agreed removal of recordExtensionInitPriv[Worker]'s ownerId argument landed as ba26d15663, avoiding unnecessary catalog lookups while keeping that cleanup easy to reverse if original-owner storage became necessary #40 #41.
- Tom ruled out backpatching because older clusters lacked the required dependency records; Hannu requested cleanup guidance or pg_dump mitigation for existing corruption, and Robert supported documenting the problem, but no such remedy was settled in the thread #44 #45 #46 #47 #48 #49 #50.
- In September, Egor Chindyaskin attached failing regression output from running test_pg_dump as a non-bootstrap superuser; Tom initially rejected the scenario as misuse of TAP infrastructure, and Marina Polyakova clarified that the failing test was the SQL regression test #51 #52 #53.
- At the thread's end, the four candidate changes had landed, the REVOKE workaround remained rejected, broader ownership and existing-corruption handling remained unresolved, and the newly reported regression-test failure had no attached fix #33 #39 #47 #53.