model-writtenrun 7504 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tom Lane reported that dumping and restoring the regression database emitted an invalid SET DEFAULT for an inherited generated column, affecting both PostgreSQL 12 and development HEAD #1.
- Peter Eisentraut initially suppressed separate expression dumping #2, but Tom identified missing dependency handling and proposed avoiding child expression objects #3; Peter adopted that approach and then added a flagInhAttrs() check after Tom found a spurious DEFAULT NULL #4 #5 #6.
- Peter initially favored forbidding generated child columns beneath ordinary parent columns because of restoration difficulties #6, while Tom challenged that restriction and argued for supporting the case #7.
- Masahiko Sawada demonstrated several inconsistent CREATE TABLE inheritance cases and supplied a generated-flag propagation fix #8; Peter expanded it to allow generated children of ordinary parents while rejecting conflicting expressions, defaults and identity specifications beneath generated parents #9.
- The expanded DDL patch landed as 086ffddf36, with only error-hint wording changed from #9, and Peter announced its application to master and PostgreSQL 12 #10.
- Tom subsequently confirmed that dump/restore still failed #11, and Masahiko supplied an alternative pg_dump fix that suppressed generated-column defaults during output, later refreshing its comments #12 #13.
- Peter isolated another unrestorable state caused by dropping a generation expression only from the parent and proposed forbidding that operation when children exist, avoiding complicated updates to children's attislocal flags #14; Daniel Gustafsson questioned the restriction, while Tom favored the simpler backpatchable remedy #15 #16.
- Tom showed that attislocal could not reliably identify inherited expressions and rejected suppressing ordinary defaults on that basis #18; he also demonstrated that ALTER TABLE INHERIT could attach a child with a conflicting generation expression #19.
- An inherited flag in pg_attrdef was discussed as a possible future improvement, but Tom noted that catalog changes would not provide a backpatchable solution #21, and Peter reaffirmed that generated children of ordinary parents were intentionally supported #22.
- The DROP EXPRESSION restriction from #14 landed unchanged in code and tests as bf797a8d97, with Peter explicitly leaving the pg_dump work outstanding #23.
- Peter compared the information-gathering and output-stage pg_dump alternatives #24, Masahiko preferred the former because it avoided intermediate objects #25, and newly discovered binary-upgrade problems delayed completion while Peter developed tests #26.
- Peter's next pg_dump patch used actual parent generation expressions instead of attislocal, retained expressions for binary upgrade, fixed dependency handling and added tests #28; Tom reviewed it and supported landing while identifying adjacent concerns #29.
- The pg_dump patch #28 landed unchanged as 0bf83648a5, and Peter committed to investigating the remaining issues #30.
- Peter could not reproduce the alleged loss of an ordinary child default #31, and Tom withdrew that report after finding the separate ALTER TABLE statement in the dump #32.
- Peter then posted ALTER TABLE INHERIT validation for generated status and matching expressions #33; Zhihong Yu suggested simplifying a redundant condition #34, but Peter retained it for maintainability and revised expression comparison to handle different column orders #35.
- After refreshing the ALTER TABLE INHERIT patch for the next minor releases #36, Peter committed it unchanged as a970edbed3 and announced completion in the final message, leaving all four concrete fixes landed #37.