model-writtenrun 6896 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jul 2025 - Jan 2026 · 13 entries
- Alvaro Herrera opened the thread with Antonin Houska's pg_squeeze-derived plan to unify VACUUM FULL and CLUSTER under REPACK, first adding the ordinary command and later concurrent rewriting through logical decoding; the initial syntax retained optional index selection and clustered-index persistence #1.
- Robert Treat identified command-tag and documentation problems and requested ANALYZE support, while Fujii Masao caught a build problem and questioned WAL-level restrictions; the revised command added ANALYZE for a single nonpartitioned table and a pg_repackdb utility implemented through vacuumdb #2 #3 #5 #8.
- Antonin's Windows installation problems and Andres Freund's objections to executable symlinks led Alvaro to extract shared vacuuming.c/h routines and build a separate pg_repackdb executable, alongside the first combined concurrent-rewrite and visibility-preservation series #10 #13 #22.
- Mihail Nikalayeu began stress-testing lost updates, initially attaching the wrong concurrent-index test, then incorporating the initial-snapshot correction and a simpler visibility-preserving replay based on transaction waits and SnapshotSelf #24 #26 #29 #30.
- The reusable vacuuming code became the first landing: c4067383cb extracted the first component of #33, with further changes to object-filter representation and vacuum/analyze mode handling.
- Subsequent basic-command review consolidated the progress views, preserved compatibility for existing monitoring tools, tightened system-catalog behavior and filtered pg_repackdb's work by MAINTAIN privilege #37 #38 #39.
- When Alvaro restored the concurrent series, he identified three unresolved design concerns: decoding spill space, MVCC unsafety and the long-lived xmin horizon preventing vacuum cleanup #40.
- Reviews by jian he and Mihail corrected grammar, options, completion, index-copy flags and exclusion metadata; Antonin also reorganized decoding setup and removed target-table TRUNCATE handling because the command's lock already prevents it #41 #43 #48 #57.
- Mihail warned that pausing decoding during long index builds could retain excessive WAL and found locking and TOAST problems, leading Antonin to add a background decoding worker and correct TOAST swapping #62 #66.
- Worker review refined synchronization and shared-file communication, and Antonin added an experimental multiple-snapshot copying design intended to advance visibility horizons during long operations #68 #70.
- Stress tests and review then exposed timeout, snapshot-stack, record-filtering and locking mistakes, which were corrected across the worker and multiple-snapshot revisions #75 #79 #83 #84.
- Mihail's duplicate-key and incorrect-sum failures proved to be more than expected MVCC unsafety: Antonin diagnosed an initial logical-decoding snapshot racing transaction commit publication and moved its fix to a separate discussion #86 #88 #89 #91.
- Mihail explored rejecting old snapshots through relcheckxmin instead of preserving every historical tuple, while testing exposed unsafe PROC_IN_VACUUM use and prompted a separate PROC_IN_REPACK horizon experiment #95 #97 #100.
Feb 2026 · 3 entries
- Alvaro narrowed the release scope by removing pg_repackdb and the experimental multiple-snapshot component, prioritizing the SQL command and concurrent rewriting; stronger stress tests continued to exercise catalog vacuum interactions #109 #111.
- After further rebases, Alvaro identified the allocation ceiling and copying costs of flattened TOAST tuples, and Antonin folded the worker into the main feature while adding serialization of individual values #119 #120 #129.
- Mihail's next review corrected memory ownership, identity-index handling and INCLUDE-column assumptions, and Antonin reused the existing snapshot-building API and added BulkInsertState support #130 #131 #133.
Mar 2026 · 8 entries
- Mihail supplied an additional fixup for per-row expression-context cleanup and scan-key ordering, which Antonin integrated into the final pre-landing basic-command series #135 #136.
- The ordinary REPACK command landed as ac58465e06 from #136, with committer changes to lock release, ANALYZE transactions and error handling; Alvaro then rebased the concurrent remainder #137.
- Noriyoshi Shinoda noticed the missing progress-view command-column documentation, Antonin approved the patch, and Alvaro added compatibility-view explanations when committing a630ac5c20 from #139 #141 #143.
- Alvaro's next review led to a substantial conversion from direct heap operations and HeapTuple to table-AM calls and tuple slots, while Srinath Reddy Sadipiralla's testing exposed decoding of unrelated relations before filtering was initialized #145 #165 #167.
- Antonin provided a deterministic reproducer for the filtering crash, while Jim Jones's separate report about other sessions' temporary tables was redirected to another bugfix thread because the behavior also affected the older commands #174 #175 #176 #177.
- Discussion among Antonin, Matthias van de Meent, Robert Treat and Alvaro established that removing REPLICATION requirements needed protection against resource exhaustion, motivating replication slots reserved specifically for REPACK #148 #151 #152 #153 #160.
- Alvaro proposed protecting the final lock upgrade by aborting sessions that waited on REPACK; Antonin supplied comment corrections and fixed tuple-slot lifetime across decoded transactions, and the series gained a separate worker source file #184 #187 #189 #195.
- Amit Kapila challenged special treatment of REPACK transactions during decoding startup, leading Antonin to replace that mechanism with database-specific logical snapshots; further review moved control into output-plugin options and revised record filtering and cleanup #198 #205 #221 #233 #241.
Apr 2026 · 19 entries
- Srinath's testing found process-flag cleanup and Windows slot-initialization defects, while Alvaro and Antonin corrected WAL-flush boundaries and transient exclusion constraints and incorporated earlier transaction-block validation #215 #216 #222 #225.
- The index-progress prerequisite landed as caec9d9fad from #234, with a revised flag-based index_create_copy interface, leaving the rest of the series in #236.
- Alvaro's final two-patch submission supplied both the source-file rename c0b53ec063 and concurrent rewriting 28d534e2ae; the latter incorporated final memory-context and tuple-validation corrections after review #239 #245 #252.
- Tom Lane's immediate Valgrind report and Srinath's inline diagnosis produced 5bcc3fbd19, which initializes serialized-snapshot padding and adjusts cleanup ordering #257 #259.
- John Naylor's buildfarm report prompted the one-line pointer-cast correction a3b069ef90 in the newly committed tuple-processing code #266 #267.
- The pending series supplied database-specific snapshots in 0d3dba38c7 and reserved replication slots in e76d8c749c, both extracted from #252 with only light changes.
- Vignesh's request for a concurrent-mode example became the separate documentation commit 8fb95a8ab6, directly from the discussion rather than an attachment #242 #243.
- Hayato Kuroda and Andres soon challenged database-specific snapshots because serialized snapshots could be reused incorrectly and indirect shared-catalog access could corrupt caches, exposing concerns beyond the original startup optimization #271 #276.
- Alvaro added autovacuum exclusion to the blocker-abort prototype but withdrew the proposal immediately after Andres objected to failing production sessions merely because they waited on REPACK #279 #280.
- Tomas Vondra's 32-bit compiler warning and Tom Lane's proposed uint64 storage replacement produced 2cff363715, resolving tuple-header storage without a larger redesign #284 #286.
- Alexander Lakhin then exposed a parallel-worker snapshot assertion, for which Antonin supplied a patch delaying the active snapshot until copying, while noting that the database-specific optimization might be reverted #299 #301.
- Zhijie Hou reported that the long-running REPACK transaction pinned restart_lsn and retained all intervening WAL, proposing advancement of the temporary slot as decoding crossed WAL segments; Antonin approved the approach #300 #302.
- Alexander's sanitizer report exposed an uninitialized shared readiness flag, and Antonin's two-patch repair also removed an unnecessary shared DSM pointer; both landed with light edits as 05c401d578 from #310 #306.
- Mihail and Antonin explored several alternatives for safe lock upgrades: preferential deadlock victims, queue-local checks, declared future lock edges and explicit upgrader state; Andres rejected queue-local checks as insufficient for cycles involving other relations #305 #311 #312 #319 #321.
- Justin Pryzby's privilege report led Antonin to remove the REPLICATION check, Chao Li independently confirmed that fix, and Alvaro added LOGIN bypass and tests; the combined submission became 5dbb63fc82 with small test changes #320 #325 #326 #327.
- Tom Lane pointed out that the new successful concurrent tests broke core regression testing with wal_level=minimal, so Alvaro moved them into contrib/test_decoding and committed 4b2aa4b39c from #336 with comment edits #330.
- Mihail simplified the lock-protection work to PROCLOCK upgrade intent plus preferential victim selection, abandoning the more elaborate future-lock graph handling as too complex for the practical benefit #338 #339.
- Renewed stress tests showed duplicate keys and lost data with database-specific snapshots; Alvaro proposed strengthening a protective assertion, and Antonin proposed separating database-specific and cluster-wide running-transaction processing #342 #343 #348.
- Alexander's observation that background workers already used the desired signal handler produced the small discussion-driven cleanup 5d48d3b14e #323.
May - Aug 2026 · 7 entries
- Alvaro bundled the snapshot cleanup with configurable stress tests, but Amit and Masahiko Sawada identified broader problems involving WAL generation, stale decoding state, cross-database snapshot restoration and standby support #350 #349 #351 #352.
- Antonin's injection-point test demonstrated a lost INSERT, and Amit explained why database-specific snapshots required deeper architectural work; Alvaro agreed to defer that optimization and committed the discussion-driven revert 01a80f0621 #359 #362 #363 #364 #365.
- Hou resumed the WAL-retention fix, incorporated Chao's comments and Amit's safety questions, and added a TAP test; Alvaro split the implementation in #371 into global restart-LSN accounting 38470c2c1e and REPACK-worker slot advancement 45b02984e2, rewriting explanatory comments #366 #369 #372 #373.
- The uncommitted WAL-recycling TAP test was resent with preexisting slots removed so they could no longer prevent recycling; its landing is not established by the supplied candidates #374.
- Mihail reminded the thread that deadlock cancellation could still discard extensive REPACK work, but Alvaro deferred the risky lock-manager changes to PostgreSQL 20; Mihail suggested retaining the simpler victim-selection alternative, and Antonin separately submitted a stale TOAST-comment cleanup #375 #376 #377 #378.
- At the last message, Amit accepted deferring lock protection and removing it from PostgreSQL 19's open issues: the main command and concurrent mode remained landed, database-specific snapshots had been reverted, and lock protection, horizon improvements and the residual test work were not demonstrated as landed here #380.
- After the thread's last message, the snapshot/CLOG race diagnosed here was fixed by 070e21268b, which waits for relevant transactions before treating the initial decoding snapshot as an MVCC snapshot; the committed implementation belongs to thread #53118 through its September submission #1154063, following the redirection in #91.