model-writtenrun 7636 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Oct 2017 - Apr 2018 · 8 entries
- Tomas Vondra proposed multi-range BRIN summaries in #1 to preserve gaps and isolate outliers that make ordinary minmax summaries ineffective, initially supporting only a small set of types.
- After correcting an incomplete attachment in #2, Tomas combined Bloom and multi-range development in #3, separating multiple-scan-key dispatch and common NULL handling into prerequisite patches.
- Mark Dilger's regression testing in #6 exposed a wrong catalog reference, which Tomas fixed in #7; Mark then verified the fix and identified an allocation-type issue in #8 and #9.
- Tomas substantially rewrote multi-minmax in #10, introducing distance support procedures, additional data types, documentation, and memory-management fixes.
- Alvaro Herrera requested regression coverage and earlier placement of NULL cleanup in #11, leading to the reordered and tested series in #13.
- Tom Lane opposed extending obsolete abstime/reltime support in #16, and Tomas removed those opclasses in #18.
- Questions about distance calculations and exceptional numeric values in #19 and #20 led Tomas to explain the merging heuristic in #22 and normalize numeric distances in #23.
- The update in #24 introduced configurable Bloom sizing and minmax summary limits, initially as index-wide settings, while documenting sizing and validation limitations.
Jun 2018 · 1 entry
- Following the catalog-format rebase in #25, Thomas Munro's Windows build report in #26 prompted the floor()/ceil() portability fix in #28.
Feb - Mar 2019 · 3 entries
- After a long gap, Tomas restored a current series in #33, and Alexander Korotkov argued for per-opclass parameters rather than access-method-wide settings in #34.
- Nikita Glukhov supplied a revised series in #40 that preserved the old consistent-function signature, corrected validation and catalogs, and centralized NULL processing through oi_regular_nulls.
- Alexander emphasized extension and pg_upgrade compatibility in #42 and #44, establishing the rationale later used to restore the legacy interface.
Jun 2019 · 1 entry
- Tomas integrated the separate opclass-parameter infrastructure in #45 and updated it in #48, enabling different settings for individual indexed columns.
Apr - May 2020 · 1 entry
- Once opclass parameters landed separately, #52 contained the remaining BRIN work; the discussion in #54, #55 and #59 deferred the feature to the next development cycle.
Jul 2020 - Mar 2021 · 24 entries
- The rebase in #61 restored a minmax-multi attachment accidentally omitted from #60.
- Masahiko Sawada's testing in #62 exposed unreadable pageinspect output and a Bloom parameter assertion, while Alexander's summary-type proposal in #75 and #77 led to dedicated types in #79.
- Tomas also introduced experimental correlation-costing changes in #79, because physical correlation poorly predicts the usefulness of Bloom summaries.
- John Naylor's tests in #82 and #90 demonstrated resilience to outliers but exposed construction costs; Alvaro's expanded-summary proposal in #86 and #89 led to retained in-memory state and serialization callbacks in #91.
- The series was consolidated in #93 and #95, while benchmarks uncovered ineffective NULL pruning and allocation overhead that prompted the fixes in #96 and #99.
- John's detailed Bloom review in #97 and #104 questioned the hashing assumptions, and his experimental one-hash implementation in #108 was included in #116.
- Oversized-summary investigation exposed an existing TOAST-pointer corruption bug in #113; its fix was developed later in thread #43194 and landed as 7577dd8480.
- Large-page experiments produced prime-bound and integer-width fixes in #117 and #118, before the broader benchmark comparison in #122 favored the simpler two-independent-hash Bloom scheme.
- Tomas selected that hashing scheme in #127 and removed Bloom's initial sorted representation in #128, judging its complexity unjustified by the demonstrated benefit.
- John's profiling in #130 exposed repeated compaction on monotonic input, leading to simpler reduction and batch-building experiments in #132.
- Stress testing in #139 uncovered missing datum copies and incorrect distance catalog entries, while comparisons of three construction strategies favored a bounded larger buffer over aggressive compaction or unrestricted batching.
- Tomas rejected making minmax-multi the default and questioned the usefulness of types without meaningful distance in #139; the selected hybrid implementation was consolidated in #148.
- The portability round in #145 corrected 32-bit alignment and by-reference value handling and removed unsafe cross-type Bloom operators.
- The final preparation moved bsearch_arg into src/port in #155, corrected documentation in #156, and added Windows build integration in #157.
- John caught an expanded-range assertion error in #158; Tomas fixed it and added coverage in #160, while choosing to remove legacy consistent-signature support.
- John tested the late series in #161, and Tomas explicitly deferred correlation and bitmap-path costing work to a later major version in #162.
- The polished series in #164 supplied the scan-key, NULL-handling and allocation components committed as a1c649d889, 72ccf55cb9 and 8e4b332e88, with only a catalog-version adjustment to the first.
- Tomas also experimented with contrib packaging in #164 and resent the remaining opclasses in #166, exposing SQL opclass-storage and dynamic summary-type lookup limitations.
- Alvaro opposed requiring separate extensions without a compelling benefit in #167, and Tomas ultimately retained both new opclasses in core.
- The packaging experiment revealed redundant cross-type support-procedure records; Tomas posted an unindexed cleanup in #168, Alvaro reviewed it in #169, and it landed as a68dfa27d4.
- Tomas reconsidered extension-upgrade compatibility and restored the old consistent-function signature in a681e3c107, announcing the reversal in #171.
- The core components of #166 landed as 77b88cd1bb and ab596105b5, with light Bloom cleanup and minmax catalog corrections; the contrib conversions were omitted.
- Immediately after landing, buildfarm failures in #173 prompted Tom Lane's sanitizer advice in #174; Tomas identified unaligned deserialization in #176 and reported pushing alignment and big-endian fixes in #177.
- The thread ended with both opclasses landed and Tomas explaining their distinction from contrib/bloom in #179, while costing improvements remained deferred and the custom-opclass NULL-key forwarding concern raised in #171 remained unresolved in the supplied discussion.