model-writtenrun 8126 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jun - Sep 2020 · 13 entries
- Robins Tharakan reported roughly 45% lower throughput in PostgreSQL 13 beta under highly concurrent, cached SELECT workloads, with performance recovering when pg_stat_statements.track_planning was disabled #1.
- Julien Rouhaud initially argued that pooling or prepared statements could avoid the overhead, while Fujii Masao identified per-query spinlock contention and favored a safer default; splitting planning and execution counters or using atomics raised memory and consistency concerns #2 #3 #4 #6.
- Fujii posted a patch disabling planning tracking by default, retaining explicit test coverage and explaining zero planning counters, which Julien reviewed favorably #8 #9.
- Ants Aasma supplied a Linux futex prototype to test whether better contention handling would help, while recommending LWLocks; Andres Freund also favored replacing the pg_stat_statements spinlocks with LWLocks #10 #14.
- Fujii posted a separate experiment using hashed LWLock partitions for counters, and Ants clarified that his futex prototype was a diagnostic experiment rather than a proposed inclusion #15 #16.
- Andres questioned the hashed lock allocation and suggested initializing an embedded LWLock where each entry's spinlock was initialized; Fujii explained his concerns about allocating and recycling locks for dynamically created hash entries, but the posted implementation retained hashing #19 #22 #23 #24.
- Fujii and Peter Geoghegan judged the locking experiment too late for PostgreSQL 13 because its effects on other workloads needed measurement; the unchanged default-setting patch #8 landed as d1763ea8c9, and the LWLock experiment was registered for a later commitfest #18 #21 #25 #27.
- Pavel Stehule requested a warning about enabling planning tracking and selected wording describing contention from a small variety of concurrent queries; Fujii attached that warning in #31 and committed it unchanged as 321fa6a4a2 #28 #30 #33.
- After confusion over which patch had landed, Fujii clarified that the LWLock experiment remained pending and rebased it without substantive changes; Hamid Akhtar reported passing installcheck-world and requested comments plus consistent use of the partition-count macro #37 #38 #40.
- Hibiki Tanaka tested three workloads and found no improvement, including more than 50% degradation for repeated SELECT 1 with the LWLock patch, prompting Fujii to reconsider the approach #41 #42.
- Fujii revisited separate planning and execution spinlocks, Julien warned that this required duplicating additional counters and widening the output, and Andres clarified how a shared lock could protect entry lifetimes alongside atomic counters; no replacement patch followed #42 #43 #44 #45.
- Robins requested credit for Sean Massey's internal investigation and posted a release-note attribution patch, but Tom Lane objected because community credit generally required direct participation; the supplied record shows no landing for that patch #46 #47 #48.
- Michael Paquier marked the stalled locking proposal Returned with Feedback, describing the proposed patch as incorrect #49.
Apr 2021 · 1 entry
- In April 2021, Justin Pryzby reopened the documentation wording, questioning the phrase 'a fewer kinds' and proposing inline replacements that described competing updates to a small number of entries; Fujii helped refine the terminology #50 #52 #54 #55.
Jun - Jul 2021 · 2 entries
- Justin resumed the wording fix in June, and Julien questioned the ambiguity of identical structure before proposing 'statements with identical query structure'; Fujii incorporated that phrase and SGML markup into the final inline diff #57 #58 #59 #60 #61.
- After Julien approved the final wording, Fujii committed the inline correction as 9d2a775734, crediting Justin as author; the thread ended with the default change and both documentation changes landed, while the locking experiment remained returned with feedback #62 #63 #64 #49.