model-writtenrun 8875 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jul - Oct 2019 · 18 entries
- Fabien COELHO proposed options to create range- or hash-partitioned pgbench_accounts tables so partitioning overhead could be measured without manually writing partition DDL in #1.
- Simon Riggs suggested the names --partitions and --partition-method in #2, and Fabien adopted them in #3 before fixing Windows string-comparison portability in #5.
- Asif Rehman's testing prompted a requirement that --partition-method accompany positive --partitions in #6; Fabien implemented it in #7, which Asif retested successfully in #8.
- Amit Kapila requested partition information in benchmark results and several code cleanups in #9, leading Fabien to add catalog-based detection and reporting in #10.
- Fabien defended --partitions=0 as useful for scripted comparisons in #15, with Alvaro Herrera supporting that interface in #19; multi-level partitioning remained a possible extension because its parameter interface was unclear in #13.
- Dilip Kumar tested the feature and questioned open-ended range bounds in #11 and #14; Fabien explained their coverage and possible performance benefit in #16, and they were retained.
- Alvaro Herrera proposed more compact result output in #20, but Amit Kapila favored retaining one parameter per line and treating broader output changes separately in #22.
- Amit supplied an inner-join alternative as a delta in #23, but Fabien retained left joins in #24 to distinguish missing tables from unpartitioned tables and obtain a consistent normal result.
- Discussion of identically named tables in multiple schemas led to search_path-aware discovery in #30, and Amit Langote's review in #33 improved option diagnostics in #34.
- Amit Langote proposed explicit older-server checks in #39, while Fabien and Dilip preferred relying on server errors in #40 and #44, leaving version-specific initialization guards out of the patch.
- Amit identified incorrect handling of SQL NULL in #41; after accidentally reattaching the old patch in #45, Fabien supplied the actual fixes in #48, including correct empty-partition counting and rejection of unknown partition methods.
- Amit supplied a full revision extracting partition creation, restoring existing fillfactor behavior, and rejecting missing account tables in #50; Fabien independently implemented those requests in #51 before noticing Amit's attachment and incorporating selected comments in #53.
- Discussion of partitioned tables with no attached partitions concluded that benchmarking should remain allowed, and Fabien made creation and reporting checks consistently use partition_method in #59, then fixed the resulting compilation typo in #61.
- Alvaro requested moving discovery code out of main() in #62, which became GetTableInfo in #68; Fabien declined pg_partition_tree() because it would unnecessarily lose compatibility with older partition-capable servers.
- Removing the test tablespace exposed the restriction on explicitly assigning pg_default to partitioned tables in #68; Amit rejected the workaround in #69, and Fabien restored the dedicated tablespace in #70 before removing unnecessary path escaping in #72.
- Amit prepared the final formatted patch in #73, retaining separate error and hint lines after discussion and adding the hint's terminal period as announced in #79.
- The feature landed as b1c1aa5318 from #73 with only that punctuation edit, and Amit announced the push in #80.
- After the commit, Ashutosh Sharma demonstrated overflow-dependent partition counts from atoi in #82; Fabien classified integer-input validation as a broader existing problem to address separately in #83, and no fix was posted here.
Jan 2020 · 1 entry
- Peter Eisentraut later found the partition-method help wording ambiguous in #85 and clarified his interpretation in #87; Fabien proposed removing parentheses around the choices in #88, leaving that follow-up pending at the thread's end.