model-writtenrun 9324 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tom Lane opened the discussion after finding widespread regression-test role names that could collide with existing users during make installcheck; he proposed systematic renaming and considered reducing the special keyword-role tests to a rolled-back transaction or removing them #1.
- Alvaro Herrera proposed sharing a small pool of test roles, while Michael Paquier opposed dropping coverage and suggested compile-time or hook-based enforcement; Tom rejected shared roles because of concurrency hazards and the continued need to enforce naming rules #2 #3 #7.
- Greg Stark suggested a regression-test safety GUC, and Stephen Frost raised failures when the operating-system user is named user; Tom treated bootstrap-user selection as a separate issue and outlined a GUC-based enforcement scheme with transactional protection for exceptional names #4 #5 #6.
- Tom reported committing the mechanical global-object renamings in 2016 while leaving the exceptional rolenames cases unresolved #8.
- Robert Haas preferred a test module and opposed coupling a general regression-test GUC to parallel-query behavior; Michael and Tom raised installation and hook complications, while Robert argued that enforcement only needed to run during make check #9 #10 #11 #14 #15.
- Peter Eisentraut questioned protecting existing installations when temporary instances are cheap, but Tom defended installcheck as a way to verify the actual installed environment, including path-sensitive SELinux policies and upgraded databases #12 #13.
- After nearly three years, Tom revived the work with a compile-time warning patch because naming violations had returned, extending enforcement to subscriptions and replication origins; warnings deliberately allowed TAP tests using isolated servers to continue #16.
- Robert reviewed the enforcement patch and proposed softer warning wording, which Tom accepted; Tom then discovered that rolenames also reset application_name settings for all users and proposed preserving its coverage in an unsafe_tests module excluded from top-level installcheck #17 #18.
- Alvaro and Stephen supported isolating the unsafe tests, and Tom posted a completed enforcement-and-cleanup series with database checks, expanded documentation, TAP adjustments and a new prohibition on user-created pg_-prefixed replication origins, explicitly leaving the rolenames move separate #19 #20 #21.
- The reserved replication-origin check from #21 landed unchanged as the separate commit a1e61badf9.
- The test-cleanup attachment from #21 landed unchanged as ca129e58c0, followed by its enforcement and documentation portion as 54100f5c60; the submission was split into three commits without substantive edits.
- Tom then posted the complete rolenames relocation, declining to split safe cases back into the core suite and asking whether older releases needed the move or removal of ALTER USER ALL tests; the posted patch landed unchanged as c91504b958 #22.
- By the final exchange all attached work had landed, and Tom explained to Peter that placing unsafe_tests under src/test/modules provided immediate buildfarm coverage without requiring owners to update their scripts; the thread records no resolution of the earlier backpatch or subscription-name reservation questions #23 #24 #22 #21.