model-writtenrun 7351 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tom Lane opened the discussion in #1 with a reproducible concurrent-DROP failure: getTables() called server-side inspection functions before acquiring table locks, while expensive ACL subqueries widened the vulnerable interval; he initially proposed separating minimal discovery and locking from detailed collection.
- Justin Pryzby confirmed production failures in #2, and Tom clarified in #3 that the proposed correction would make failures occur cleanly at LOCK TABLE rather than guarantee success during concurrent drops; automatic retries remained only a possibility.
- Hans Buschmann reported that parallel dumps underestimated tables with large TOAST data in #4, and Tom incorporated TOAST relpages and overflow protection into the six-patch series posted in #5.
- The series in #5 separated requested and existing object components, moved ACL merging to the client, postponed unsafe inspection calls, batched table-related queries using Andres Freund's idea, corrected TOAST scheduling, and prepared repetitive queries; Tom reported major speedups but left the bulk-query patch dependent on resolving support for servers without unnest().
- Justin measured a sixfold reduction in single-table ACL overhead and questioned an array-length check in #6; Tom explained the initial opening brace in #7, and Hans reported successful, faster Windows parallel dumping of a TOAST-heavy database in #8, while separately suggesting optional clustered output that was not implemented in this series.
- Andres benchmarked an older-server filtering alternative and reported a large prepared-query speedup in #9, while warning that static preparation flags could fail with parallel workers; Tom raised maintenance concerns in #10, Alvaro Herrera proposed a separate state structure in #11, and Andres suggested a centralized query helper in #12.
- Tom rejected the helper as excessive machinery and posted #13 with an enum-indexed boolean array attached to Archive and initialized per connection; the other five patches were unchanged.
- After roughly five weeks, Tom posted #14 with rebased patches, explanatory array-length comments, and a minimum source-server version of 8.4, choosing that restriction over compatibility code because a separate discussion favored dropping still more old-server support.
- On December 6, the final patches in #14 landed unchanged as component tracking 5209c0ba0b, client-side ACL handling 0c9d84427f, and postponed unsafe inspection e3fcbbd623; these exact final matches supersede the older source attribution in thread #44991.
- The remaining final patches in #14 also landed unchanged on December 6 as bulk queries 9895961529, prepared queries be85727a3d, and TOAST-aware scheduling 65aaed22a8; preparation of less-common object queries remained omitted because setup costs could outweigh savings.
- With all six patches committed, Hans requested a TOAST scheduling backpatch in #15, but Tom declined in the final thread message #16 because older getTables() implementations required substantial rewriting and retesting for a scheduling problem he considered uncommon.
- On December 31, the related follow-up 3e6e86abca extended the same safety principle to getPolicies(), filtering interesting tables on the server before evaluating pg_get_expr(); it has no matching attachment or review exchange here and is attributed to thread #44991 under the cross-thread rule.