model-writtenrun 8760 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Aug - Sep 2015 · 3 entries
- Anastasia Lubennikova proposed storing duplicate B-tree keys with GIN-style posting lists and possibly posting trees, raising compatibility and dead-item cleanup questions in #1.
- Tomas Vondra questioned the distinction from btree_gin and the concurrency cost, while Alexander Korotkov explained the importance of ordinary scans, composite keys, and efficient range searches in #2 and #3.
- Peter Geoghegan raised prefix compression and page-space accounting alternatives in #5, while Anastasia chose posting lists as the initial scope and deferred posting trees in #6.
Jan - Mar 2016 · 4 entries
- The first implementation in #9 demonstrated substantial index-size savings, but testing by Thom Brown and Aleksander Alekseev exposed shared-memory warnings and slow insertions in #15 and #17; #19 removed unrelated changes accidentally included during rebasing.
- Peter's review in #26 required binary equality rather than operator-class equality to preserve values returned by index-only scans, and asked for explicit locking and representation arguments; Anastasia responded with refactoring and design documentation in #27.
- Anastasia added WAL support in #31, and Alexandr Popov reported favorable size and performance tests in #32, but Robert Haas and Peter considered the work too late and insufficiently reviewed for the release in #33 and #35.
- After Anastasia accepted postponement in #37, Heikki Linnakangas questioned the on-disk encoding and identified the unsafe high BlockNumber flag bit in #39.
Jul 2019 - Mar 2020 · 31 entries
- Following a multiyear gap, Anastasia revived the project in #41 around PostgreSQL 12's B-tree changes, replacing eager posting updates with lazy deduplication immediately before a page split.
- Peter's real-data testing found large space savings and improved insertion behavior in #42 and #43, establishing that the lazy design addressed important weaknesses of the earlier implementation.
- Peter's detailed review and formatted patch in #45 made preservation of logical tuple contents and heap-TID order central requirements; Anastasia added amcheck support and insertion inside posting lists in #56 and #57, also incorporating Rafia Sabih's feedback from #50.
- Peter fixed ordering and truncation failures and strengthened amcheck in #60, then supplied the separate split-accounting correction in #61, which improved space utilization by accounting for posting data removed from high keys.
- The next development round tightened page-space assumptions and dead-item handling, while Peter's experimental unique-index implementation in #65 initially performed poorly and was left out of Anastasia's main revision in #66.
- Anastasia reworked atomic posting replacement and WAL handling in #67 and #70, but Peter found amcheck failures and poorer packing in #71.
- Peter proposed keeping a posting list's size unchanged and moving its former maximum TID into a new plain tuple in #73; Anastasia restarted from an earlier version to implement that simpler split design in #74.
- Peter's subsequent revision #78 fixed Valgrind findings, simplified searches and insertion, made backward scans return posting entries in descending order, and documented the design.
- Posting lists gained kill_prior_tuple support in #79, allowing LP_DEAD marking when every represented tuple was dead and adding the recovery-conflict handling needed to remove such lists safely.
- Anastasia introduced logical posting-split and deduplication WAL in #80, but Peter found that replay lacked the original incoming TID and supplied a diagnostic fix in #83.
- Development briefly split into Peter's space-accounting branch #84 and Anastasia's WAL-correction branch #86, each intentionally excluding the other's latest changes.
- Peter traced excessive WAL volume to a missing temporary-page LSN and supplied #88; Anastasia merged both branches and repaired the replay bug exposed by that fix in #89.
- The merged implementation was simplified and hardened in #92 and #93, with fewer allocations, shared index-build helpers, recovery memory management, and corrected VACUUM statistics.
- Improved WAL efficiency made unique-index deduplication promising again, so Peter reintroduced it as a separate delta in #94; Anastasia then changed deduplication to update pages directly and log individual posting tuples in #95.
- Anastasia added a per-index switch and integrated unique-index support in #97, and Peter developed specialized heuristics intended to delay version-driven page splits in #99.
- After further benchmarking and a gap in submissions, Peter moved the implementation into nbtdedup.c in #101 and coordinated deduplication with index-build fillfactor and single-value page splits in #102.
- The remaining series in #103 temporarily disabled unique-index deduplication by default and excluded system indexes, while #113 promoted the pageinspect instrumentation into a documented companion feature.
- Peter identified WAL overhead even for indexes that did not benefit from deduplication in #108, then addressed it and introduced a separate obsolete pin-scan cleanup patch in #115.
- A more selective unique-index trigger, based on finding an existing duplicate of the incoming value, replaced the earlier default-setting compromise in #118, alongside backward-scan and VACUUM memory fixes.
- The pin-scan prerequisite from #118 landed as 9f83468b35 with additional documentation and WAL-version maintenance, and #124 carried the remaining series forward.
- Peter revised the posting-split explanation and preserved LP_DEAD bits during VACUUM updates in #125, after which Heikki's detailed review in #126 challenged the per-posting WAL approach, unique-index complexity, and VACUUM record size.
- Peter initially incorporated smaller review changes in #128 and experimented with Sokolov Yura's sorting work to address a page-compaction bottleneck in #130.
- Following Heikki's feedback, #131 restored temporary-page deduplication and one WAL record per page, eliminating the compaction bottleneck, the sorting prerequisites, and the custom unique-index deduplication strategy while retaining selective triggering.
- The next revision #132 encoded VACUUM deletions as compact posting-list offsets instead of entire replacement tuples, and removed the system-wide GUC in favor of the per-index deduplicate_items parameter.
- Peter added equalimage operator-class support in #133, based on Anastasia's privately supplied work, to prevent deduplication where equal values can have semantically different representations.
- Anastasia questioned redundant support functions and user-facing notices in #134; the final submission #140 consolidated the API around a type-OID argument, lowered the notice to DEBUG1, and revised the documentation.
- The main deduplication component of #140 landed as 0d861bbb70 with light final edits, enabling lazy deduplication by default where safe and requiring REINDEX for upgraded indexes to acquire the safety flag.
- Peter announced the main landing in #141, reported that a compiler-warning issue had already been addressed, and delayed pageinspect for another inspection after the buildfarm stabilized.
- Fujii Masao supplied the inline documentation index-term correction in #142, which Peter applied as e537aed61d and acknowledged in #143.
- The pageinspect component of #140 subsequently landed as 93ee38eade, with additional validation, tuple-descriptor handling, and documentation edits; the separate debug-value instrumentation was excluded.
- By the closing congratulations in #144 and reply in #145, the feature, pin-scan cleanup, pageinspect support, and documentation correction had landed without a reported revert; the default-enablement decision remained subject to beta feedback, as discussed in #138 and recorded in 0d861bbb70.