model-writtenrun 5617 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Nov 2021 - Apr 2022 ยท 17 entries
- Mark Dilger proposed independently grantable SET and ALTER SYSTEM privileges to reduce dependence on superuser access, replacing an earlier hard-coded-role approach with per-parameter ACLs and explicit parameter-registration commands #1.
- Robert Haas and Tom Lane objected to maintaining an exhaustive parameter catalog and requiring extension registration scripts, leading to demand-created ACL entries, text parameter names, and ordinary role dependencies; Andrew Dunstan also helped examine catalog and upgrade behavior #2 #12 #14 #19 #21.
- Mark implemented grants on unknown names and dump support, then followed Robert's review by moving the privileges view out of information_schema, adopting setting terminology, and adding privilege inquiry functions #30 #31 #32.
- After Andrew found an accidentally retained local-testing change, Mark cleaned up and rebased the patch while expanding its scope to revocable PUBLIC access through USERSET-to-SUSET conversions and initial grants, subsequently extending that treatment to contrib modules #33 #35 #36.
- Joshua Brindle tested the feature, reported a locale-dependent failure, and requested security hooks; Mark fixed the tests and added a GRANT/REVOKE hook, while further discussion established that controlling SET itself required additional support #37 #38 #39 #42.
- Following a January rebase, Mark added SET and ALTER SYSTEM hooks using the existing OID interface after off-list testing with Joshua, despite parameters often lacking catalog OIDs #44 #45.
- Tom revived review in March with a rebase and challenged the extension upgrade burden and mass context changes, proposing default ACLs derived from existing USERSET/SUSET contexts instead #46 #48.
- Mark adopted context-derived defaults and removed the initialization and extension changes, and Andrew reported that the resulting behavior matched his expectations in testing #51 #52.
- Tom rejected using unstable parameter ACL OIDs for SET hooks and identified snapshot constraints; Joshua supplied a string-hook delta and corrected its drop-hook macro, providing a separate foundation for the security integration #53 #58 #71 #74.
- Peter Eisentraut's terminology and privilege-name feedback helped settle GRANT SET or ALTER SYSTEM ON PARAMETER, and Mark adopted that syntax with pg_parameter_acl, expanded completion, and a new psql \dcp command on a base containing string hooks #63 #64 #65 #75.
- Tom found that the regression tests modified shared server configuration and failed on repeated runs, and that \dcp lacked sufficient documentation and consistent pattern handling; Mark moved the tests to unsafe_tests, repaired cleanup, and deferred \dcp to later work #78 #80 #82.
- Tom demonstrated that unloaded custom parameters broke grants and pg_dumpall, prompting Mark to try SUSET placeholders and extension-script PUBLIC grants while converting parameter hooks to the string interface #87 #89.
- David G. Johnston reviewed the resulting revocation semantics and documentation, while Tom proposed abandoning revocable USERSET access to avoid extension ambiguity and routine catalog lookups; Mark, Andrew, and David accepted that simplification #90 #92 #94 #95 #96.
- Mark posted the simplified implementation, after which Tom identified mixed-case custom-name hazards and proposed removing OID-based privilege inquiries because catalog OIDs represented ACL entries rather than stable parameter identities #101 #102.
- Tom's revised patch normalized names, simplified default ACL handling, removed the privileges view and GRANT hook, and reduced completion code, then received a grammar-conflict rebase #107 #108.
- Tom committed the final posted revision with expanded documentation, a catalog-version bump, test cleanup, and additional SET privilege handling for connection-time SU_BACKEND parameters as a0ffa885e4, crediting Mark as the patch author #108 #109.
- The thread ended with the core privilege feature committed, USERSET access remaining irrevocable, and Tom planning separate follow-up discussions for a redesigned \dcp command and another idea #109.