model-writtenrun 6022 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tom Lane proposed replacing pg_dump's binary-search lookup arrays with one simplehash table keyed by CatalogId, finding shorter code but no measurable performance improvement and some executable growth #1.
- Andres Freund measured compiler-dependent size changes and found libpq and object sorting dominated client time, while Nathan Bossart found no speedup even with 200,000 tables #3 #4 #6.
- Tom identified server-side planning as a major cost and began experimenting with prepared queries; Andres proposed batching per-object queries, prompting Tom to explain the locking hazards of catalog-inspection functions #5 #7 #8.
- Andres supplied a separate getTableAttrs() prototype using unnest() to batch attribute and default retrieval, reducing his regression-database schema dump from approximately 0.675 to 0.477 seconds #9.
- Tom raised compatibility concerns about unnest() on old servers, and Andres suggested an ANY() alternative; they favored batching table and index queries while retaining prepared per-object queries where many built-in objects would otherwise be fetched #10 #11 #12.
- After testing several compilers, Tom favored landing the hash-table patch for maintainability despite its unmeasurable speed benefit and small executable-size changes #13.
- Andres suggested single-row retrieval and batching triggers, but lacked time to develop the prototype further; Tom agreed to take it forward alongside his other pg_dump work #14 #15.
- Andres discovered that frontend simplehash error handling could return, and Tom agreed to make it fatal; Tom declined unconditional hash-table initialization because it would require extra initialization plumbing without a measurable benefit #16 #17.
- An attempt to prevent inlining of simplehash growth functions encountered conflicting inline declarations and unused-function warnings, leaving that optimization unresolved #18.
- Tom committed the discussion-derived fatal-error fix with unlikely() annotations and copyright text, backpatching it to supported branches beginning with PostgreSQL 13 974aedcea4.
- Tom then committed the original hash-table patch with additional object-type assertions in seven lookup helpers, retaining its shared lookup structure and maintenance rationale #1 92316a4582.
- After landing, Andres proposed moving SH_GROW into the implementation section to control inlining, Tom supported the approach, and Andres reported compiler checks supporting its warning behavior; the thread ends with this refinement unresolved and no recorded landing of the bulk-query prototype #19 #20 #21.