model-writtenrun 497 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Noah Misch opened the thread with AI-assisted findings and TAP reproducers showing that server-based subscriptions could fail during dump/restore because user mappings and foreign-server privileges were checked before restoration finished; cross-database ownership reassignment was being handled separately #1.
- Jeff Davis proposed deferring connection validation until use and supplied SQL regression tests instead of adding a postgres_fdw dependency to pg_dump tests; he also found that retain_dead_tuples could force a publisher connection during binary upgrade #2.
- Jeff argued that restore should recreate catalog state without connecting and that connection-time validation must be authoritative, while Hayato Kuroda highlighted missing foreign-server permission checks during REFRESH PUBLICATION #3 #6.
- Jeff supplied an independent three-patch series fixing role-membership lock release, DROP SERVER dependency handling, and unsupported SCRAM pass-through for subscription workers #5.
- The design evolved toward constructing conninfo only when needed, preserving server capability and relevant privilege checks, and reducing missing-user-mapping checks to warnings; changing dump/restore ordering was deferred for possible consideration in PostgreSQL 20 #7 #8.
- Jeff consolidated eleven patches, incorporating Shlok Kyal's ALTER and ACL work, separating conninfo generation from subscription loading, and proposing a revert of owner-change validation that could break restore #8.
- Amit Kapila initially retained DDL-time retain_dead_tuples validation with a binary-upgrade exception, then offered an alternative that removed the check from ENABLE while retaining authoritative worker checks; Jeff preferred and approved that alternative #9 #10 #12 #14.
- Jeff and Amit agreed that dropping and recreating a user mapping in separate transactions could legitimately trigger disable_on_error, so that behavior was left unchanged; Amit also supplied an explanatory comment for separating subscription loading from conninfo construction #11 #13.
- The report's conninfo-logging finding produced a separate commit removing the DEBUG1 connection-string message without an attached matching patch 448c4b3bfd, as later confirmed in Jeff's findings summary #23.
- The first three components of the consolidated series landed unchanged as the lock-release fix b48caa7dfd, DROP SERVER fix b66ea17a3a, and SCRAM pass-through rejection d63f214d68, all from #8.
- Jeff resent the eight remaining patches with Amit's explanatory comment and a corrected revert that retained the DROP SERVER tests #15.
- Hayato tested the revised retain_dead_tuples patch on PostgreSQL 19 and master, and Amit committed it with only comment wrapping changes as 98bcb0ab8a before announcing the push #16 #18.
- Hayato investigated whether separate subscription and conninfo reads could see inconsistent catalog state, then concluded that invalidation processing avoided his suspected inconsistency; Shlok confirmed the reported permission and ownership problems were fixed, and Amit endorsed the overall series #17 #19 #20 #21 #22.
- The conninfo refactor from #15 landed with a local allocation-order adjustment as 8f7617ff7a, followed by the unchanged disabled-worker ordering fix 8b73ceb78f.
- The ALTER decision logic from #15 landed with ENABLE removed from conninfo-requiring cases and new ENABLE/DISABLE tests as e90e198e3f, followed by the unchanged unconditional USAGE checks dcfb02deed.
- The remaining restore fixes from #15 landed unchanged as mapping warnings 5c39be9e6d, conditional CREATE conninfo construction 254beb174e, the owner-validation revert ac230cd3ce, and relaxed owner-change checks 7f4fb05807.
- Jeff closed the restore-related open item after the serious fixes landed, while recording remaining documentation and psql omissions, application_name semantics, startup invalidation questions, session-state limitations, and separately discussed DROP SUBSCRIPTION behavior #23.
- Tom Lane then relayed a Coverity concern about CREATE SUBSCRIPTION's conninfo branch; Jeff explained why the existing branch was safe and supplied a clearer condition with assertions, which Tom approved #24 #25 #26.
- At the last message, Jeff planned to commit and backport the approved cleanup after the freeze ended; the supplied commit subsequently landed the patch unchanged as 3d00537feb #27.