model-writtenrun 9509 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Nov 2013 · 2 entries
- Michael Paquier opened the thread in November 2013 with comment cleanup, snapshot-wait refactoring, and a tested REINDEX CONCURRENTLY implementation that preserved index OIDs by swapping relfilenodes #1.
- After Michael expanded the design explanation in #4, Andres Freund demonstrated unsafe concurrent index-state transitions and questioned whether relfilenode swapping could expose inconsistent storage mappings #5.
Jan 2014 · 1 entry
- Álvaro Herrera committed the comment cleanup #6, and Michael resent the remaining snapshot helper and core feature with rebasing and comment corrections #8.
Aug 2014 - Feb 2015 · 4 entries
- Michael revived the work in August 2014 #9, then changed the swap to use AccessExclusiveLock and added psql autocommit-off handling to address correctness concerns #12.
- Jim Nasby's regression testing and detailed review #13 led Michael to fix a missing include, copy existing tuple descriptors, and rewrite the phase documentation #14; subsequent isolation testing exposed deadlocks addressed with WaitForLockers() #17.
- Peter Eisentraut objected that exclusive locking undermined the advertised concurrent behavior #30, while Robert Haas's dual-relfilenode proposal #24 and Álvaro's relmapper alternative #26 prompted further discussion of storage and cache correctness.
- In February 2015, Andres proposed abandoning relfilenode swapping entirely: create an index with a new OID, transfer its dependencies and name, then drop the old index concurrently #39.
Feb - Apr 2017 · 5 entries
- After a two-year gap, Andreas Karlsson implemented that design on top of Michael's work, making index validity changes transactional #41; system-catalog support was excluded because fixed catalog-index OIDs could not safely change #42.
- Michael's testing exposed schema-reindex crashes and snapshot leaks #42, and discussion settled on grouping short phases across a table's indexes while keeping expensive build and validation phases in separate transactions #46 #53.
- Jim raised concerns about stored index OIDs and requested event-trigger support #54, but Andres opposed invoking event triggers inside the operation #56, and Jim explicitly declined to make his unusual object-tracking use case a blocker #66.
- Thomas Munro checked predicate-lock handling and confirmed the existing transfer to heap-level locks covered it #67 #68, while Andreas completed foreign-key and trigger-reference transfer in the next revision #69.
- Michael Banck's testing and review #71 and Michael Paquier's snapshot and code-reuse concerns #73 informed Andreas's August 2017 revision #77, which retained the per-table processing tradeoff.
Nov 2017 - Feb 2018 · 1 entry
- Further discussion favored shared index-creation infrastructure over constructing parser nodes internally #82, but limited development time stalled the work again #85 #86.
Dec 2018 - Apr 2019 · 22 entries
- Peter revived Andreas's latest patch in December 2018, updating APIs and reviewing the new-OID design while preserving its architecture #87.
- Sergei Kornilov found inconsistent partitioned-table handling, duplicate verbose output, incorrect documentation fields, and replacement-index name collisions #88 #90, which Peter addressed in #107.
- Sergei then reported lost index comments and a separate failure mode where interrupted concurrent rebuilding followed by ordinary REINDEX could leave an undroppable duplicate TOAST index #108; the latter remained unresolved in this thread.
- A syntax debate retained CONCURRENTLY as a top-level keyword, with Andres arguing that consistency with CREATE/DROP INDEX and the command's distinct transaction behavior outweighed moving it into parenthesized options #125.
- Michael proposed splitting reusable snapshot-wait and index-build routines from the feature #114, but Peter kept the remaining work together and fixed comment preservation with regression coverage #126.
- Michael supplied both a full revised patch and a review delta simplifying index copying and correcting table APIs #127; Peter incorporated most changes while retaining his helper naming convention #128.
- Sergei reported successful installcheck-world testing and examined concurrent-operation deadlocks #130, while Michael recommended controlled reindexdb TAP tests for database-wide coverage #132.
- Peter added reindexdb --concurrently and whole-database tests #134, after which Sergei reviewed transaction-grouping documentation, recovery advice, and client-version handling #137.
- Peter's final full patch corrected excessive locking during deletion by separating concurrent lock selection from concurrent-drop processing, and completed the documentation and client checks #138.
- The final full patch #138 landed unchanged apart from rebasing as 5dc92b844e, announced by Peter in #141.
- Immediately after landing, Peter reported clang-specific buildfarm failures near the final deletion phase #144, Michael supplied a short reproducer #146, and Peter subsequently reported the issue fixed #152.
- Noriyoshi Shinoda's partition-index testing exposed dependency errors and leftover invalid indexes #147, starting a separate post-commit repair.
- Robert Treat proposed making concurrent rebuilding the default #148, but Andres cited substantially greater I/O and additional failure modes #149, and Michael emphasized its incompatibility with transaction blocks #155.
- Nathan Bossart supplied an inline documentation typo correction #150, Justin Pryzby expanded it into a documentation cleanup #151, and Peter reported those corrections fixed #153.
- Peter's first partition repair transferred outgoing dependencies as well as incoming ones #156, but Michael demonstrated that the replacement index disappeared from the partition tree #157.
- Michael expanded the repair to preserve the partition flag and switch inheritance links at swap time, avoiding premature attachment that would prevent cleanup after failure, and added partition-tree tests #159.
- Dagfinn Ilmari Mannsåker proposed correcting documentation about recovery from invalid indexes #161, prompting Álvaro Herrera to argue that explicitly rebuilding one invalid index should be allowed even if table-wide rebuilding skips invalid indexes #163.
- Michael replaced that documentation-only remedy with code, documentation, and tests allowing REINDEX INDEX CONCURRENTLY on invalid indexes while retaining table-level skipping #164.
- Peter committed the expanded partition repair from #159 with light cleanup as ef6f30fe77 and announced it in #165.
- Dagfinn caught a missing documentation parenthesis in the invalid-index patch #167, Peter endorsed the fix #169, and Michael committed #164 with that correction and additional failure tests as a6dcf9df4d.
- At the final message, the feature and both follow-up repairs had landed, and Michael asked that further feature discussion move to a new thread #170.
- The unresolved TOAST-index report later informed work in thread #41947, whose newer patch #444539 produced 61d7c7bce3; that commit belongs to the later thread rather than to an attachment here.