model-writtenrun 9211 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jul 2018 - Jan 2019 · 15 entries
- Andres Freund opened the thread with a reworking of Haribabu Kommi’s pluggable-storage design, using typed tuple slots and moving buffer ownership and visibility into access methods while keeping index insertion outside them #1.
- Alexander Korotkov argued for more AM control over index updates, but Andres deferred richer indexing until concrete zheap needs emerged and preferred an AM-level CLUSTER callback #3 #6.
- Haribabu’s early attachments merged the table-AM headers, converted COPY and FDW refetching to slots, and repaired isolation-test failures; Andres’s feedback led to input-byte COPY accounting and removal of a premature bulk-insert abstraction #8 #9 #13 #17 #19.
- Experience porting zheap drove explicit bitmap/sample, vacuum, and rewrite interfaces, static handlers, and a default-AM setting, while projection-list support was deferred to avoid coupling access methods to plan nodes #12 #14 #23 #24.
- The storage-lifecycle work exposed an invalid assumption that indexes had table AMs, while Alexander’s separate costing series needed missing implementation files added after Haribabu’s review #32 #34 #33 #37 #38.
- Haribabu fixed ANALYZE offsets and CREATE AS materialization, then replaced Dmitry Dolgov’s concurrent-update workaround with explicit invisible-tuple handling that Dmitry tested successfully #40 #41 #43 #44 #45.
- The cumulative fixes gained initial documentation, which Amit Langote and Kyotaro Horiguchi reviewed, while Asim R P highlighted remaining block-storage assumptions and Andres prepared a more deeply split implementation #46 #47 #51 #56 #48 #54.
- Dmitry added psql access-method display and pg_dump support, then replaced per-table USING output with tracked default_table_access_method state; Amit Khandekar’s reviews caught archive-version, memory-management, and display-suppression issues #59 #61 #63 #66 #79.
- Two unattached preparatory changes separated heap vacuum code and reused parallel-scan memory sizing: 285d8e1205 and 90525d7b4e, clearing ground for the full series later posted in #78.
- Andres’s first full attachment split the work into 42 patches, restored parallel scans with reusable block helpers, overhauled relation creation and truncation, and moved XID-horizon handling toward AM callbacks #78.
- The snapshot and visibility prerequisites from #78 landed as e7cc78ad43, 63746189b2, b7eda3e0e3, c91560defc, ebcc7bf949, and 346ed70b0a, with include cleanup, expanded snapshot documentation, combined relocation patches, and light naming/formatting edits.
- Haribabu reduced heap-tuple OID dependence and revised snapshot callback removal, but complete t_tableOid removal remained blocked by trigger consumers that still accessed HeapTuples #77 #80 #82 #99.
- The slot groundwork from #78 landed as 8aa02b52db, 5408e233f0, and b8d71745ea, adding pin transfer, embedded tuple workspace, and combined table-OID/TID propagation with local refinements.
- Trigger slotification and EPQ slot storage from #78 became ff11e7f4b9 and ad0bda5d24 after local restructuring; Andres retained the external HeapTuple trigger interface for compatibility and deferred its redesign #101.
- The buffer-slot virtualization and simple virtual-slot changes from #78 became f414abd62d and 70b9bda65f, with the former expanded to handle materialization safely after an external redis_fdw failure thread #212443.
Feb - Jun 2019 · 25 entries
- Amit added bounded-lookahead prefetching to primary-side XID-horizon computation, and Robert Haas reviewed its state and heuristic before publishing a combined, polished patch #88 #89 #92 #95 #96.
- With the prerequisites landed, Andres posted the remaining series, incorporated psql/pg_dump, and disallowed AM selection on partitioned roots; that exposed inappropriate slot sharing and led to per-partition ON CONFLICT state #105.
- The infrastructure and dump components of #105 landed as 8586bf7ed8 and 3b925e905d, with only a catalog-version adjustment in the former, and the cleaned ON CONFLICT patch #106 landed unchanged as 277cb78983 after Robert Haas and David Rowley reviewed it #107.
- Andres consolidated scan interfaces, documented callbacks, and added mandatory-callback assertions following Ashwin Agrawal’s testing, while leaving the proposed trigger/EPQ snapshot change out because its design needed more thought #102 #114 #105.
- The scan component of #119 landed with light cleanup as c2fe139c20; a buildfarm failure then exposed insufficient 32-bit alignment, repaired without an attached patch in 8cacea7a72 #120 #121.
- Kyotaro’s review improved the pending syntax/API documentation, while Haribabu supplied a separate psql access-method label fix and additional typo repairs that Andres reported pushing #122 #124 #126 #133 #134.
- Andres reworked modification/EPQ handling, added missing concurrency tests, distinguished updates from deletes, and moved latest-version locking into the AM; the final component in #130 became 5db6df0c01 after cosmetic edits #127 #131.
- The remaining series #130 supplied CREATE AS/materialized-view insertion, row-version fetching, index existence checks, and latest-TID lookup in 919e48b943, 9a8ee1dc65, 71bdc99d0d, and 2e3da03e9e, with direct input-slot use, heap_fetch cleanup, and wrapper refinements.
- Primary-side XID-horizon computation landed as 558a9165e0, combining the prefetch work in #96 with the later table-AM integration in #130, removing obsolete WAL relfilenodes and avoiding recovery-time catalog lookups #132; this is later than the source credited by thread #39814.
- Index-build scans and relation lifecycle/rewrite operations from #130 landed as 2a96909a4a and d25f519107, with scan-interface polish and CLUSTER progress integration.
- VACUUM/ANALYZE and planner-size callbacks from #130 became 737a292b5d and 696d78469f, including block-rejection handling for analysis and cleanup of the planner wrapper.
- Sample and bitmap scans from #130 landed as 73c954d248 and bfbcad478f, with visibility and scan-loop refinements; bitmap prefetch/visibility-map abstraction and more general sampling remained future work.
- Bulk-insert finishing and heap-only contrib guards from #130 landed as d45e401586 and 4b82664156, with flag naming cleanup and broader guard coverage, leaving COPY as the main outstanding core piece #135.
- Andres rejected a merged index/table documentation chapter and duplicated per-callback reference text, choosing a separate conceptual table-AM chapter with authoritative API comments in tableam.h #135 #137 #138.
- Andres’s revision #140 became b73c3a1196 with added storage, WAL, and transaction guidance; Justin Pryzby’s subsequent wording repairs #141 became 86cc06d1cf with light editorial changes.
- COPY’s table-AM conversion from #130 was combined with buffering work in thread #39488 to form 86b85044e8, avoiding repeated slot recreation and batching inserts across partitions.
- Heikki Linnakangas’s toy AM exposed index-only slot confusion, invalid vacuum horizons, and assumptions about physical files; its comment-fix component #143 landed as 6421011ea2, while further toy testing revealed an index reorder-queue slot bug #146.
- Heikki’s next comment pass prompted questions about scan timing, flags, and interface contracts; Andres and Tom Lane discussed replacing numerous scan booleans with flags, while broader API redesign remained outside the immediate release scope #147 #148 #150.
- The invalid-horizon fix #151 expanded beyond assertions to let CLUSTER/VACUUM FULL preserve AM-supplied invalid horizons and landed as fdc7efcc30.
- Andres proposed routing relation sizes through the AM, Ashwin examined implications for Zedstore and sampling, and the solution grew a separate TID-validity callback; Rafia Sabih’s toy-AM failure also prompted Ashwin’s callback-signature adaptation #155 #156 #166 #157 #159 #161.
- Andres rejected passing a fixed slot at scan startup because a scan may use multiple slots, leaving Ashwin’s column-store initialization concern to a first-getnextslot workaround #164 #165.
- The two patches in #166 landed as 7f44ede594 and 147e3722f7 after Ashwin’s review, with light relation-size edits and additional TID scan-lifecycle changes; block-oriented ANALYZE generalization remained deferred #169 #171 #173.
- The scan-flags discussion also contributed to c3b23ae457, which fixed ANALYZE’s inappropriate predicate locking after a crash report in thread #40612 and incorporated sample-locking considerations from thread #40615, without a matching attached patch here.
- Heikki reported the index-slot problems resolved in June, while the toy access method remained a testing aid rather than a proposed production AM #174 #143.
- By the final exchange, the table-AM infrastructure and listed follow-up fixes had landed; larger storage/sampling and trigger-interface improvements remained future work, and Michael’s last message proposed regrouping configuration sections without a further patch #173 #101 #179.
Aug 2019 · 1 entry
- Heikki reopened the missing sample setting in August, Michael Paquier supplied #177, and Andres committed it unchanged as f7db0ac7d5 #176 #178.