model-writtenrun 7253 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Oct - Nov 2021 · 2 entries
- Joshua Brindle opened the thread with #1, replacing membership tests with has_privs_of_role() so predefined roles would respect NOINHERIT consistently with ordinary roles, while retaining is_member_of_role() with a warning against using it for privilege checks.
- Stephen Frost supported the consistency fix and found a stale comment in #6; Nathan Bossart then identified missing documentation in #8, leading to the expanded corrections in #9 and his approval in #10.
Jan - Apr 2022 · 7 entries
- After the rebase in #12, Joe Conway offered to commit and backpatch the change, but Tom Lane and Robert Haas opposed changing stable-branch security behavior in #14 and #15, so the eventual fix targeted the development branch only.
- Robert Haas argued in #15 that role-wide NOINHERIT prevents useful combinations of personal privileges and predefined capabilities, while Joe Conway demonstrated an intermediate-role workaround in #16 and Joshua Brindle documented existing inconsistencies among predefined roles in #21.
- Robert Haas accepted consistency among predefined roles as the strongest argument in #24, while broader changes to role attributes and inheritance were deferred to a later development cycle in #27.
- Joe Conway took over patch preparation in #32, rebasing Joshua Brindle's work, correcting an error message, and adding two newly introduced basebackup call sites; Joshua and Stephen questioned how NOINHERIT should behave on replication connections that cannot SET ROLE in #33 and #36.
- Joe Conway withheld the basebackup additions and committed the remaining 23 files from #32 as 6198420ad8, explaining the split in #38.
- Robert Haas accepted the basebackup check changes in #39; the missing required_role documentation reported by Joe Conway in #32 was separately supplied in #503689 in thread #45655 and committed unchanged as 26a0c025e2, so that commit belongs to the other thread.
- Joe Conway then committed the three withheld files from #32 unchanged as 9752436f04, confirming the second landing and spotting awkward wording in the new documentation in #41; Robert proposed clearer wording in #42.
Jun - Aug 2022 · 10 entries
- Robert Haas reopened the architectural question in #43, proposing per-grant inheritance so a role could automatically inherit one membership while requiring SET ROLE for another; Tom Lane warned about breaking existing dumps and suggested retaining rolinherit as a default for future grants in #44.
- Robert Haas and Stephen Frost explored extensible GRANT options and the distinction between implicit privileges and SET ROLE in #50, #51, and #52, while additional controls over SET ROLE and conversion of role attributes remained outside the implemented inheritance feature.
- Peter Eisentraut opposed removing role-level inheritance and favored per-grant overrides in #60, but Robert's first implementation in #62 still removed rolinherit and included a temporary version stamp for testing.
- Nathan Bossart's compatibility review in #63 helped lead to #72, which retained role-level inheritance and introduced TRUE, FALSE, and DEFAULT grant settings, with DEFAULT consulting the role attribute at use time.
- Robert Haas then adopted Tom Lane's alternative in #79: each grant stores a Boolean, initialized from the recipient's role attribute when unspecified, and later ALTER ROLE changes affect only future grants; Nathan Bossart highlighted the resulting compatibility change in #80.
- Tushar Ahuja found malformed pg_dumpall output in #82 and an upgrade failure caused by an empty WITH clause in #85; Robert fixed these in #83 and #86, and Tushar reported successful broader testing in #90.
- After substantial integration with intervening grantor changes in #91, Tushar retested successfully in #92, and Robert committed e3ce2de09d with only a stale-comment correction and catalog-version bump, as reported in #93.
- Robert discovered overlooked inheritance documentation and posted #94; Nathan Bossart requested one punctuation correction in #95, and the corrected patch landed as 620ac28548, confirmed in #96.
- Robert found that grantor selection incorrectly ignored ADMIN TRUE memberships when INHERIT was false and posted #97, moving ADMIN detection before the inheritance filter and adding regression tests; Nathan's portability report in #98 led to corrected test queries and expected output in #100 and #101.
- Nathan Bossart examined the grantor semantics in #102, Robert explained that intermediate links need INHERIT while the final link needs ADMIN in #103, and Nathan approved the final patch in #105 before it landed unchanged as 0101f770a0.
Oct 2022 · 1 entry
- The submitted implementations and follow-up fixes had landed when Pavel Luzanov proposed displaying membership-level INHERIT and ADMIN options in psql in #108, but this thread contains no implementation of that proposal.
Apr 2023 · 1 entry
- The final message, #109, leaves follow-up observations from Noah Misch about an obsolete AUTHOID cache callback, inaccurate inheritance documentation and comments, and unconditional inheritance of implicit pg_database_owner membership in the new behavior; no resolving patch appears in the supplied thread.