model-writtenrun 8345 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Feb - Mar 2017 · 4 entries
- Kyotaro Horiguchi proposed a checkpoint-time cap on WAL retained by replication slots to protect primary-server disk space, initially preferring a small retention change over automatic slot deletion or invalidation #1.
- Michael Paquier preferred external monitoring and slot removal #2, while Robert Haas argued that bounded retention should let primary availability take priority over a failed standby #6; Horiguchi retained a cluster-wide setting rather than the per-slot option Andres Freund proposed #7 #10.
- Peter Eisentraut requested size-based units and visible failure status #8 #9, leading first to max_slot_wal_keep_size #14 and then to megabyte-based arithmetic and more selective warnings #16.
- Petr Jelinek raised possible logical-replication data loss #4, and Craig Ringer clarified that decoding starts at restart_lsn and errors when required WAL is missing #15.
Aug - Sep 2017 · 1 entry
- After Eisentraut renewed the monitoring request #17, Horiguchi added live and distance columns as a supplementary patch #19, combined the patches #20, and supplied missing regression expectations #21.
Oct 2017 - Jan 2018 · 3 entries
- The monitoring implementation changed from predicting the next checkpoint to checking actual WAL availability #22; Freund then argued for an available LSN instead of a constantly changing distance #26, producing textual status and min_secure_lsn #27.
- Horiguchi fixed default-setting crashes #28, shared retention arithmetic between checkpointing and monitoring, and introduced recovery TAP coverage #30.
- Sergei Kornilov's testing prompted clearer state names and documentation #31 #32, while his and Thomas Munro's test_decoding reports led to corrected expectations and more stable TAP tests #33 #40 #41.
Mar 2018 · 1 entry
- After rebases across PostgreSQL development changes #43 #44, CI exposed an unclosed WAL directory that Horiguchi fixed #45.
Jun 2018 - Feb 2019 · 8 entries
- Masahiko Sawada's detailed review corrected the interaction between wal_keep_segments and the new cap, simplified arithmetic, and reduced repeated checkpoint warnings #46 #47.
- Sawada requested useful per-slot headroom #48 and supplied a remaining-segments delta after discovering identical values across slots #52; Horiguchi adopted per-slot calculations but exposed bigint bytes instead #54.
- Further Sawada testing uncovered headroom errors involving actual retained segments and wal_keep_segments #55 #64 #74, leading to arithmetic corrections and additional TAP cases #56 #65 #75.
- The discussion kept checkpoint scheduling and slot retention as separate controls and retained wal_keep_segments for slotless uses #58 #60; converting that older setting to size units was deferred to separate work #62 #63.
- Eisentraut requested an explicit unlimited value and documentation of exceeded limits #69, so Horiguchi changed the default to minus one and clarified that checkpoint-time removal makes the cap a soft limit #73.
- A slot could appear lost while decoding continued from an already-open segment #74 #75, prompting an optional per-record removal check #76; Sawada considered that issue separable from the main feature #77.
- Paquier identified the reader's frontend/backend restrictions and reviewed WAL discovery, diagnostics, test cost, and test output #78 #79; Horiguchi responded with corrected reader handling, smaller WAL segments in tests, named-slot warnings, expanded output, and a TAP setup helper #80.
- After another rebase and language cleanup #81 #84, Jehan-Guillaume de Rorthais rebased the series and used network-delay tests to show slots moving misleadingly between lost and recoverable states #85.
Jun - Jul 2019 · 1 entry
- Horiguchi revised monitoring around normal retention, extra retention, recoverable loss, and definitive loss, with nullable headroom and delayed warnings #86 #88; de Rorthais verified the changed behavior but questioned the unstable state and repeated warnings #90.
Nov 2019 - Jan 2020 · 1 entry
- Following a gap, Horiguchi corrected reporting after configuration changes and restored oldest-file discovery #92, then addressed Herrera's API concerns by introducing an enum, separating distance arithmetic, and simplifying affected-slot reporting #89 #94 #95.
Mar - Jun 2020 · 15 entries
- Herrera resumed integration by rebasing and squashing the series #97; Horiguchi generalized TAP configuration support #99, but Herrera deferred that infrastructure change in favor of append_conf() and simplified oldest-WAL discovery #100 #101.
- Herrera extracted the expanded-output cleanup from #99 into 69360b3458 and posted the remaining feature and reader-check patches #102.
- Herrera challenged leaving lagging walsenders alive and relying on recycled-but-open files #103 #104, leading Horiguchi to remove the optional reader patch and actively terminate affected walsenders #108.
- Herrera then separated retention arithmetic from side effects through checkpoint-time InvalidateObsoleteReplicationSlots() #109, and Horiguchi completed that design, replaced remain with min_safe_lsn, and removed the losing state #110.
- Review exposed races from changing a slot without owning it #112, so Herrera implemented kill-and-acquire invalidation #113 and an inquiry-mode acquisition loop #114, then combined those changes with Horiguchi's restored test file #115 #116.
- Horiguchi proposed an alternative invalidated flag against his earlier branch #118, but Herrera found that its acquisition check prevented invalidation and dropping slots, so that alternative was left out #119.
- The combined patch #116 landed as c655077639, with light naming, API, documentation, formatting, and catalog-version edits; Herrera announced the landing in #119.
- Immediately afterward, Horiguchi reproduced a logical-walsender assertion caused by invalid restart_lsn #120 and supplied a rejection check #121, which Herrera expanded to SQL decoding and slot advancement #125.
- The invalid-LSN fix landed as d0abe78d84 from #125, with diagnostic wording adjusted and the SQL check moved after CreateDecodingContext() to preserve more specific errors #131.
- Horiguchi also found that slot-invalidation waits consumed checkpoint wakeups #121 and proposed a pending-request flag #122; Herrera's attempted simplification #126 introduced a busy loop that Horiguchi identified while supplying a TAP reproducer #128.
- Herrera combined the reproducer with a shared-flag check #130, then replaced SetLatch() with continue when committing 1816a1c6ff, explaining the final choice in #132.
- Andres Freund later found that the checkpoint fix still allowed shutdown signals to be missed and proposed moving WaitLatch() earlier in the loop #136; Herrera reviewed that restructuring, while their discussion also considered delayed sync-request processing #137 #138 #139.
- Herrera separately proposed making CreateCheckPoint() return bool #140, but Paquier questioned the benefit #141; the supplied thread shows no landing for that cleanup or the later latch restructuring.
- The last message supplies Justin Pryzby's documentation wording fix #143; the core feature and two runtime fixes had landed, while this documentation patch and the remaining checkpointer proposals had no confirmed resolution in the supplied messages.
- The later invalidated_at change, 0188bb8253, belongs to the newer matching patch #452583 in thread #42516, although its discussion trailers also reference the invalidation concerns raised here.