model-writtenrun 2423 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Aug 2020 · 1 entry
- Dmitry Dolgov opened the thread to stop ORM-generated IN lists of different lengths from filling pg_stat_statements with separate entries, comparing early conversion to an array Const with suppressing repeated constants during hashing and favoring the latter #1.
Nov 2020 - Jan 2021 · 1 entry
- The first developed implementation suppressed hashing beyond a threshold for constants and VALUES, then added Params support and corrected normalized-text copying and multi-column VALUES handling with review from Zhihong Yu #2 #5 #6 #7.
Jun 2021 · 1 entry
- When query jumbling moved into core, Dmitry ported the implementation and explained that a separate hashing extension could not express collapsed query text without duplicating normalization machinery; this version also added expression evaluation, followed by a fix for missing constants #11 #13.
Jan 2022 · 1 entry
- Tom Lane objected to expression evaluation and conflating parameterized queries, while Robert Haas supported optional collapsing, prompting Dmitry to remove expression evaluation and Params handling and narrow the feature to constant ArrayExpr lists with merging disabled by default #16 #22 #24.
Mar 2022 · 1 entry
- Robert Haas and Tom Lane argued that retaining a prefix of constants added complexity without useful information, and Dmitry replaced that algorithm with whole-list eligibility checking and first/last location tracking #25 #29 #31 #33.
Sep 2022 · 1 entry
- After further rebases, Sergei Kornilov identified a missing sample configuration entry, and subsequent upstream changes required adapting the patch to generated query jumbling through ArrayExpr's custom_query_jumble attribute #35 #37 #44.
Jan - Mar 2023 · 4 entries
- Michael Paquier questioned the representation of merged locations and requested separate regression tests, while David Geier examined corner cases, coercions, prepared statements and simpler eligibility checks #45 #55 #58.
- Álvaro Herrera proposed preserving approximate cardinality through powers-of-ten buckets, and Dmitry implemented range text, magnitude hashing, reusable decimal-length helpers and a boolean setting #53 #60.
- David Geier found the treatment of NUMERIC values counterintuitive, but Dmitry deferred broader coercion support because general expression simplification risked reviving the earlier objections #61 #62.
- Dmitry adapted the tests to upstream restructuring, and later practical testing by Jakub Wartak, Maciek Sakrejda and Yasuo Honda supported the basic feature #65 #68 #69 #70.
Jul 2023 · 1 entry
- Nathan Bossart and Michael Paquier argued for separating simple merging from cardinality grouping, leading Dmitry to split the work into basic merging, decimal-length helpers, power-of-ten grouping and an optional threshold #66 #71 #74.
Sep - Oct 2023 · 1 entry
- Michael Paquier's implementation review led to a generic node-field annotation and _jumbleElements(), while configuration moved into pg_stat_statements with a setter controlling core jumbling #75 #77.
Mar - May 2024 · 2 entries
- Rails testing exposed that prepared parameters and bigint conversions still escaped merging; Dmitry acknowledged both limitations and kept them outside the initial scope #88 #89 #90 #91.
- Sutou Kouhei's review uncovered insufficient normalized-query buffer allocation for very large cardinality ranges, which Dmitry fixed alongside documentation, assertions and readability improvements #94 #96.
Aug 2024 · 1 entry
- Following more rebases, Dmitry summarized basic constant-array merging as the likely first landing, with grouping, thresholds and broader expression support still disputed or deferred #99 #101.
Feb - Mar 2025 · 13 entries
- Álvaro resumed work toward committing the basic feature, while Sami Imseih demonstrated inconsistent bigint and float behavior and argued that configuration must work without pg_stat_statements; Álvaro agreed that the setting belonged in core and advocated limited cast recognition #102 #103 #104 #107 #108.
- Sami supplied an alternative custom ArrayExpr jumbler that traversed function arguments, but Dmitry continued the existing node-attribute implementation, posted only its basic patch, restored the core setting and added simple cast-wrapper support #112 #114.
- Álvaro's review simplified the list scan, avoided repeated volatility lookups and prompted the name query_id_squash_values, which Dmitry adopted #115 #116.
- Sami's testing found missing query text when multiple squashed intervals occurred and renewed the request for prepared-statement support; Dmitry deferred Params while investigating the normalization problem #117 #118.
- Julien Rouhaud pointed out that bare ellipses broke EXPLAIN GENERIC_PLAN on normalized queries, and Álvaro proposed a parameter followed by a commented ellipsis, which Dmitry implemented #119 #120 #122.
- Dmitry then fixed multiple-interval text loss and parameter positioning, and Julien's follow-up prompted contiguous visible parameter numbers and a minimum two-element requirement for squashing #124 #125 #129.
- Sami supplied a delta experimenting with PARAM_EXTERN and looser function handling, but Dmitry resisted hiding arbitrary functions and warned that posting experimental patches displaced the main patch in CI #136 #137.
- Álvaro took over preparation by posting Dmitry's patch after pgindent, while Sami continued to argue that cast eligibility should not depend on function volatility #140 #142.
- Álvaro concluded that the built-in-function and cast-form restrictions made volatility checks unnecessary, then posted a revised patch with consistent squashing terminology, clearer documentation and reorganized normalization code #146 #147.
- After fixing SGML and receiving Dmitry's testing feedback, Álvaro committed 62d712ecfd from #147, additionally allowing explicit casts and updating the corresponding expected output #148 #149 #150.
- The landing provided optional, default-off constant-list squashing, while power-of-ten grouping and thresholds remained outside the committed patch; Álvaro explained that grouping had drawn objections over strange, invalid SQL output, although the new commented representation made reconsideration possible #150.
- Álvaro also noted that squashed lists actually contributed nothing to the hash despite comments describing first/last contributions, leaving possible empty-list ambiguity and further testing as open questions #150 #153.
- The thread ends with Sami's uncommitted follow-up adding PARAM_EXTERN recognition and tests, after Álvaro invited a patch and Dmitry offered collaboration; prepared-statement support therefore remained under review #151 #152 #154 #155.