model-writtenrun 670 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andres Freund opened the thread with evidence that growing regression-test runtimes, especially Windows licensing costs and assertion-heavy execution, threatened cfbot's compute budget #1.
- Profiling led Andres to psql's internal queries, which accounted for about 13% of query duration; he identified an unexpected sequential scan for exact table-name lookups and also highlighted publication-query planning, expensive tests, temporary files, and assertion overhead #2.
- Julien Rouhaud suggested compute_query_id = regress to avoid profiling-related test-output differences #3, and Jelte Fennema-Nio supplied a fix restoring index scans for exact table-name lookups, requesting backpatching to PostgreSQL 18 #4.
- Tomas Vondra offered to reduce statistics-test and BRIN validation costs without losing coverage #5, while Andres suggested cheaper query conditions and reusing function-call setup #6.
- David Rowley proposed moving compact-attribute validation into TupleDescFinalize #7, Andres discussed lighter checks and identified costly callers #8, and Andres separately suggested making pg_regress select the profiling-compatible query-ID setting automatically #9; these proposals have no attached implementation or landing in the supplied thread.
- After a gap until July, Tom Lane reviewed Jelte's fix, clarified that it concerns exact-match LIKE and regex patterns rather than general prefixes, and proposed requiring both collations to be deterministic #10.
- Jelte explained that a nondeterministic index can safely return a superset because the original expression is rechecked, restored that allowance, and added an ICU test; he also raised removing unnecessary rechecks as a separate optimization unsuitable for backpatching #11.
- Tom accepted the superset argument and landed #11 as 2d7808e6fc, refining comments and replacing one-off test tables with existing tables to avoid extra test overhead, as reported in #12.
- In the same message, Tom supplied a separate development-branch draft to remove redundant exact-match LIKE rechecks while retaining rechecks for regexes and malformed patterns #12.
- Jelte revised that draft to preserve rechecks for bpchar padding and indeterminate expression collations, but questioned whether the latter needed a different approach and explicitly declined to pursue it further; the regression fix had landed, while the recheck optimization remained an unfinished draft #13.