model-writtenrun 4826 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Robert Haas proposed a per-membership SET option in #1, allowing privilege inheritance without allowing SET ROLE, particularly to prevent members from creating objects as predefined roles; he also explained that ADMIN holders could still grant themselves access.
- Nathan Bossart supported the proposal while examining ADMIN's ability to bypass restrictions in #2, and Wolfgang Walther proposed treating SET and INHERIT as separately grantable role privileges in #3; Robert rejected that larger redesign because the privileges overlap and separate administration would add complexity in #4.
- Jeff Davis demonstrated that disabling SET ROLE still allowed ALTER TABLE OWNER TO in #7, exposing the need to define which capabilities SET permission should control rather than restricting only one command.
- Peter Eisentraut questioned the relationship between role and user identities in #9, and Robert explained PostgreSQL's distinction between current_user and session_user in #10.
- Robert's revised patch in #11 made SET permission govern ownership transfers and creation of objects owned by another role, crediting Wolfgang's proposal from a related discussion and choosing this rule for backward compatibility and usefulness with predefined roles.
- Nathan questioned the surprising ownership semantics in #12, while Stephen Frost advocated splitting ADMIN rights in #13; Robert retained indivisible ADMIN authority in #14 and #15, leaving finer-grained administration outside this patch.
- Nathan accepted proceeding in #17, and Jeff accepted the general SET permission approach over a narrower virtual-role alternative in #20; Robert posted the rebase in #21 and announced commitment in #23, with 3d14e171e9 matching that attachment.
- After the main commit, Alvaro Herrera identified a leftover pg_dumpall comment in #22 and Erik Rijkers reported documentation misspellings in #24; Robert reported corrections and a forgotten catalog-version bump in #25, while Erik and Justin Pryzby noted a remaining spelling correction in #26 and #28.
- Michael Paquier supplied missing psql completion for the new SET clauses in #27; Robert approved the substance but questioned unrelated condition reordering in #29, and Michael removed that reordering before landing 9d0cf57492, as reported in #31.
- Noah Misch reopened the documentation question in #32, pointing out that SET FALSE's effect on ownership operations was undocumented and that inherited access to existing objects could still enable escalation through functions or indexes.
- Robert agreed to improve the documentation in #33 and surveyed the distinct membership checks in #35; Noah recommended documenting the SET option's broader scope in #34 and prioritized the ownership-rule gap in #36.
- Jeff broadened the discussion to SECURITY DEFINER and SECURITY INVOKER trust risks in #37; Robert discussed safer defaults and suggested moving broader trust-control proposals to a separate thread in #39, which Jeff agreed to do in #40.
- Robert posted ownership-documentation corrections in #38; Noah found three omitted reference pages and requested an explanation alongside the SET option itself in #41, leading to the expanded documentation patch in #42.
- Noah reviewed the expanded patch favorably but suggested removing an inheritance qualifier in #43; Robert defended it using administration-only and pg_hba.conf grouping use cases in #44, and Noah declined to insist on removal in #45.
- Robert landed the documentation attachment unchanged as 3cdf7502f8 and confirmed that decision in #46, leaving all submitted work committed while allowing that the explanatory wording might be improved later.