model-writtenrun 5800 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Oct - Nov 2021 · 4 entries
- Mark Dilger split this work from the broader superuser-delegation discussion, proposing regression tests, a role-ownership hierarchy, and reduced CREATEROLE powers to prevent privilege escalation #1.
- Nathan Bossart raised upgrade compatibility and broken ownership transfers, and Mark proposed initially placing upgraded roles under the bootstrap superuser while acknowledging the deliberate compatibility break; Andrew Dunstan favored ending the overpowered behavior rather than preserving it indefinitely #2 #3 #6.
- Shinya Kato requested visible owner information, which Mark added alongside working REASSIGN OWNED and ALTER ROLE OWNER TO support; DROP ROLE CASCADE remained outside the series because cluster-wide object removal posed unresolved problems #4 #5 #7.
- Shinya's membership tests prompted Mark and Tom Lane to clarify that existing WITH ADMIN OPTION behavior should remain compatible with the SQL standard, while Shinya separately identified inconsistent permission checks between creating and altering replication roles #8 #9 #10 #12.
Dec 2021 - Apr 2022 · 13 entries
- After Shinya rebased the series in December, Joshua Brindle tested and approved it, then discovered that assigning a role ownership of itself broke password changes #13 #15 #17 #20.
- Mark prevented direct and indirect ownership cycles, added regression coverage, and appended removal of the unreliable pg_auth_members.grantor field; Joshua approved the revision #21 #22.
- Andrew's review led to descriptive test-role names, clearer parameters and comments, and documentation improvements, followed by a rebase without substantive changes #23 #24 #25.
- Stephen Frost accepted the test patch but challenged automatic inheritance of owned roles' privileges, the interaction with ordinary membership, granting attributes merely because the creator possessed them, and removing grantor information needed for privilege revocation #26.
- Andrew extracted and committed the unchanged test patch from #25 as e9d4001ec5, following his announcement that this independent piece could proceed #27.
- Robert Haas and Andrew defended automatic owner privileges as necessary for a useful sandboxed-superuser experience, while Stephen argued for separately controlled administration and data access; Mark emphasized reducing CREATEROLE's scope and explored accommodating both uses #28 #29 #32 #34 #37 #39.
- Mark reposted the remainder with ownership creation and ownership privileges combined, fixed noncontroversial review issues, and withdrew grantor-column removal for possible separate work, leaving the disputed privilege semantics unchanged #41.
- Mark then floated an experimental work-in-progress attachment separating attribute administration from ownership through per-attribute ADMIN options; he explicitly described it as an experiment rather than a complete replacement for the ownership series #42 #47.
- Michael Banck's testing exposed different privileges depending on whether CREATEROLE was assigned during creation or later, missing syntax for some administration options, and difficulty granting newly created groups; Mark acknowledged the compatibility and usability tradeoffs remained unresolved #43 #44 #45 #51.
- Stephen advocated independently delegable role-management operations, and Joshua supplied a concrete bot use case requiring account creation and login disabling without access to employee data; Mark explored optional membership and administration rights that could support both restricted bots and sandbox administrators #49 #50 #52.
- Andrew and Robert questioned the complexity of per-attribute administration, while Mark proposed configurable creator privileges to accommodate both camps; Robert accepted that goal, and Maciek Sakrejda supplied a cloud-provider use case supporting sandboxed administration #53 #55 #56 #57 #58 #59 #60.
- Robert later outlined alternatives including a TENANT grouping with separate creation, deletion, and privilege-assumption rights; Joshua supported restricting the bot's powers, but their follow-up left ADMIN membership semantics and self-revocation behavior unresolved #61 #62 #63.
- At the thread's close, Greg Stark requested a refreshed main patch, but Robert concluded that no further work would land for PostgreSQL 15 and suggested deferral or return with feedback; the regression tests had landed, while the broader redesign remained unsettled #64 #65.