model-writtenrun 545 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Chao Li reported that losing a foreign-server user mapping prevented ALTER SUBSCRIPTION from switching to a working server or connection string, and proposed lazy connection-string resolution with regression tests in #1.
- Hayato Kuroda reproduced the problem and found that DROP SUBSCRIPTION also failed after removing the user mapping in #2; Jeff Davis later explained in #17 that dependencies on individual mappings are inappropriate because lookup can fall back to a public mapping.
- Ajin Cherian and Zsolt Parragi questioned comments, lookup placement, and accessor usage in #4, #6, and #7, leading Chao Li to clarify the interface and extend accessor use through logical replication workers in #9.
- Jeff Davis considered the lazy accessor too easy to misuse and replaced it with explicit control over connection-string construction in #10, adapting Chao Li's tests and adding a separate DROP fix for Hayato Kuroda's report.
- Chao Li endorsed the new interface but showed that checking only the main slot name could skip cleanup of tablesync slots in #11, prompting Jeff Davis to move option parsing earlier and split DROP handling into delayed connection construction and subtransaction-based error recovery in #12.
- Jeff Davis also added a retain_dead_tuples fix in #12 so switching servers validates the new publisher using the new connection information; that component landed unchanged in 8eba2edb80.
- Chao Li caught an assertion that incorrectly required old connection information in #13, and Jeff Davis corrected it while making option handling more conservative in the remaining series #14; Chao Li approved the correction and supported keeping error capture local to DROP in #15.
- The ALTER component of #14 landed as e5c40584a7, with Jeff Davis adding DISABLE handling before committing and resending the two remaining DROP patches in #18.
- The delayed-connection DROP component of #18 landed unchanged as 702e9dfd6c, allowing DROP to finish without resolving a broken server when neither a main slot nor unfinished table synchronization requires cleanup.
- The remaining subtransaction patch stayed uncommitted: Jeff Davis questioned which errors could safely be absorbed in #19, Hayato Kuroda raised error-code and NULL-return concerns in #21, and Amit Kapila warned against swallowing cancellation or interrupt errors in #23.
- Jeff Davis identified a mismatch between the documentation and existing tablesync cleanup behavior in #20, then argued in #24 and #25 for an explicit way to drop a subscription without attempting any publisher connection.
- Amit Kapila pointed out that slot_name=NONE can also express manual management of the main slot while tablesync cleanup remains desirable in #26, leading Jeff Davis to suggest FORCE or NOCONNECT syntax in #27.
- Amit Kapila considered new syntax too late for PostgreSQL 19 and proposed documenting existing behavior in #28; Jeff Davis accepted the wording in #29, and Amit deferred syntax design to separate work in #30.
- Amit Kapila posted documentation patches for the supported branch groups in #31, Nisha Moond verified application and documentation builds in #32, and d7b62e3c22 clarified that DROP may still contact the publisher for tablesync cleanup after slot_name=NONE, with backpatching through PostgreSQL 14.
- Shlok Kyal opened a separate follow-up in #34 after finding that publication-changing ALTER SUBSCRIPTION commands could contact a foreign server without checking the subscription owner's permission.
- Hayato Kuroda's reviews in #35 and #37 narrowed permission checks to commands that refresh publications and requested separate statement-kind branches; Shlok Kyal incorporated both changes while preserving checks for the default refresh=true behavior in #38.
- At the final message #39, Hayato Kuroda considered the permission patch acceptable but required coordination with overlapping work; it had no landing shown here, while the general DROP error-catching patch and explicit no-connect syntax also remained unlanded.