model-writtenrun 8967 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Peter Eisentraut opened the thread with runtime superuser control of allow_system_table_mods, tests documenting its behavior, and a proposal to disable apparently redundant index checks; he also identified the surprising non-superuser DML exception and ineffective SET STATISTICS allowance #1.
- Robert Haas and Stephen Frost advocated stronger barriers against dangerous catalog operations, while Andres Freund supported restricting catalog writes even for superusers and opposed removing checks that could provide separate layers of defense #2 #3 #6.
- Tom Lane warned that extension upgrade scripts depended on direct catalog updates and demonstrated cases without practical DDL replacements, arguing against trying to classify every catalog operation as safe or unsafe #4 #9.
- Chapman Flack raised compatibility mechanisms for extension updates, and Tom proposed separate DML and DDL switches with a transition period for existing scripts; Andres suggested a warning mode, while Robert ultimately accepted that finer safety distinctions might be impractical #11 #12 #14 #16 #17.
- Peter's next patch renamed the setting to allow_system_table_ddl, retained runtime superuser control, removed its non-superuser DML exception, and improved the regression tests; it omitted the index-check disabling and explicitly left a separate DML setting unimplemented #18.
- Bruce Momjian subsequently preferred one three-state setting, eventually suggesting none, dml, and all, but that alternative did not enter the submitted implementation #19 #20.
- After several months, Michael Paquier reported schedule-file conflicts and requested renewed attention; Peter responded with a three-patch series that preserved the original setting name and moved the tests into unsafe_tests #22 #23.
- Tom read the final series, accepted retaining the original name to avoid invasive renaming, and requested replacing “Inconsiderate” with “Ill-advised” and reconsidering the documentation's initdb sentence #24.
- The non-superuser DML restriction landed unchanged from the first patch in #23 as 508bf95b76.
- Runtime superuser control landed from the second patch in #23 as c4a7a392ec, with Tom's wording change and removal of the initdb sentence.
- The regression tests landed unchanged from the third patch in #23 as 7fc380f83d, using unsafe_tests to avoid running potentially disruptive catalog operations in the normal regression suite.
- Peter confirmed that the series was committed with the requested adjustments, closing the thread with all three final patches landed; broader DML controls and automatic distinctions between safe and unsafe catalog operations remained outside this work #25.