model-writtenrun 6293 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Dec 2018 - Apr 2019 · 5 entries
- Pascal Legrand opened the thread with a planner hook that records planning time through pgss_store(), following previously quoted advice from Tom Lane, while leaving an optional tracking GUC and final column naming unresolved #1.
- Sergei Kornilov requested correct extension versioning, documentation, separation of unrelated changes and performance measurements; Pascal's next full patch added the 1.8 upgrade, documentation and regression tests #2 #4 #8.
- Julien Rouhaud supplied a formatting fix-up and identified a correctness problem: an entry evicted between parse analysis and planning could be recreated without query text; he also proposed passing an operation enum to pgss_store() instead of inferring the operation from timing values #9 #11 #12.
- After Pascal requested help, Julien took over implementation and split the work into passing query text through planner interfaces and adding planning statistics, with a track_planning switch and explicit store-operation selection #18 #21.
- Pascal's initial benchmarks found small overhead, and after the July rebase Sergei Kornilov and Tomas Vondra independently reported acceptable performance across several execution modes and machines #22 #25 #26 #27.
Sep 2019 · 1 entry
- Yoshikazu Imai identified unused planning buffer measurements and demonstrated large planning-time variation between generic and custom plans; Julien added minimum, maximum, mean and standard-deviation planning times, shared array-based calculations, and actual accumulation of planning buffer usage #29 #32 #34.
Nov 2019 - May 2020 · 13 entries
- The naming discussion exposed a flaw in combining planning and execution distributions: their counts need not match, and standard deviations cannot simply be added, so separate planning and execution statistics remained necessary #39 #40 #41.
- After a gap, upstream changes required a January rebase, but a subsequent submission accidentally used the older version without the expanded statistics; Julien acknowledged the mistake and restored the newer implementation in March #42 #48 #50 #56.
- Pascal's IVM compatibility reports and Julien's CREATE TABLE AS investigation led to ignoring planner calls with zero queryId, preserving the existing treatment of underlying utility statements while broader utility-query accounting was deferred #43 #52 #55 #56.
- Imai initially requested a non-NULL query-text assertion, but Marco Slot explained that Citus constructs internal queries without retaining text and should not incur deparsing overhead; the revised series allowed NULL and skipped those planner calls in pg_stat_statements #57 #58 #61 #64 #67.
- Imai and Sergei then caught documentation omissions and checked compatibility with older installed extension versions, producing documentation revisions and a ready-for-committer assessment #68 #70 #71 #72 #73.
- Fujii Masao's review retained extension-aligned function versioning, proposed reusing BufferUsageAccumDiff(), and revived the concern about sharing planner and executor nesting depth; Julien implemented separate depths and the reusable helper #74 #76 #77 #78 #80.
- Fujii extracted the buffer refactoring for independent commitment and challenged the usefulness of a combined total_time column; Julien dropped that column and repackaged the work as three patches, leaving users to calculate totals appropriate to their needs #82 #84 #85.
- The buffer-helper member of the final three-patch series landed unchanged as 4a539a25eb, and its planner-query-string member landed with query_string naming edits as 6aba63ef3e #85 #86.
- Testing the remaining feature exposed unwanted tracking of a prepared query inside EXPLAIN EXECUTE with track=top; Julien corrected the planner check to use both planner and executor nesting depths and resubmitted only the uncommitted feature #88 #91 #92 #93.
- Fujii polished hook naming, exception cleanup, enum values and documentation examples, and that exact final attachment became the planning-statistics feature commit #94 #95 17e0328224.
- After landing, Fujii requested clearer documentation about successful planning followed by failed execution; Pascal and Julien supplied wording, and Fujii's resulting patch explaining failures and cached plans landed unchanged #97 #98 #99 #100 58ad961f19.
- Andy Fan subsequently proposed five query-shape fields for identifying tuning opportunities, but Julien Rouhaud and Michael Paquier questioned coupling them to pg_stat_statements and their semantics for nested queries, while Andy argued that optimizer transformations make raw query text insufficient #107 #108 #109 #112.
- That separate proposal shifted into discussion of storing complete plans, EXPLAIN overhead and a possible plan identifier, with no revised attachment or landing shown in the thread #113 #115 #116 #117.
Jul 2021 · 2 entries
- In July 2021, Tom Lane reported that forced cache invalidation made the prepared-statement planning count unstable, a problem previously hidden because the buildfarm had not run these tests; he and Julien agreed to check a bounded count only for the prepared statement #118 #119 #120.
- Julien tested the repair with and without forced invalidations, and Tom committed it with formatting, comment and output-column adjustments as 0806d08d46; the thread ends with the original feature, documentation and test repair landed, while Andy's query-shape proposal has no recorded resolution #124 #125 #127.