model-writtenrun 5607 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jun 2020 - Feb 2021 · 13 entries
- Amul Sul proposed persistent, reversible WAL prohibition for controlled switchovers in #1, using checkpointer-coordinated process barriers, terminating transactions with XIDs, and checking WAL permission before critical sections; he credited initial work by Amit Khandekar and design by Robert Haas.
- Tom Lane questioned the complexity and ALTER SYSTEM interface in #6, while Robert defended preserving read queries and WAL streaming during switchover in #7, and Andres Freund emphasized avoiding checkpoint and restart delays in #12.
- Robert clarified in #4 that the guarantee concerned new WAL generation rather than all disk writes, and that prepared transactions could remain pending while the recovery checkpoint was deferred until writes resumed.
- Tushar exposed the inability to change state in single-user mode in #3, leading Amul to add standalone support and startup-checkpoint deferral in #26.
- Read-only shutdown caused checksum-tool and upgrade difficulties reported in #27 and #34; Robert explained in #32 and #35 that these followed from deliberately suppressing the shutdown checkpoint.
- Soumyadeep Chakraborty's suggestions for broader file-freezing modes in #36 and #52 were kept outside the immediate feature because Robert argued that checkpoint-dependent transitions would undermine fast switching in #48 and #56.
- Reviews by Soumyadeep and Andres in #44 and #45 led to consolidated checkpointer code, node serialization, conflicting-transition checks, and cleaner WAL guards in #54.
- Robert's race analysis in #58 prompted separate shared state and atomic updates in #59; Andres's further review in #61 led to a grantable SQL interface and revised locking in #63, followed by generation-based transition tracking in #65.
- Robert independently revised the barrier prerequisite and supplied a test-only stress module in #67; Amul identified redundant retry restoration in #68, corrected it in #72, and Andres approved the revised processing loop in #74.
- Robert and Andres argued that WAL prohibition must not wait for a lengthy checkpoint in #73 and #75, so #78 added prompt transition processing and temporarily disabled CheckpointLock pending separate investigation.
- Robert committed the barrier component of #78 as a3ed4d1efe, leaving its test module and the actual WAL-prohibition feature uncommitted, as announced in #79.
- After Robert's interface and correctness review in #80, Amul removed the committed prerequisites and ALTER SYSTEM syntax in #81, then adopted the pg_prohibit_wal function name and clearer state representation in #83.
- Robert and Andres agreed that archive recovery should clear prohibition and that startup WAL-writing operations should be isolated for later checkpointer execution in #84 and #85; Amul implemented the first XLogAcceptWrites design in #86.
Mar - Oct 2021 · 18 entries
- Dilip Kumar questioned the grouping of recovery and promotion work in #88, and Robert insisted in #94 that recovery should finish normally while only necessary WAL writes remain deferred; #95 and #98 moved that work into the checkpointer and guarded against recursive transitions.
- Ibrar Ahmed supplied a conflict-resolving rebase in #99, Amul continued it in #100, and #101 completed documentation and revised WAL guards.
- Prabhat Sahu reported a standby crash in #102, leading #103 to distinguish pending, skipped, and completed startup work; Bharath Rupireddy's wording review in #105 was subsequently incorporated in #106.
- Robert's review in #109 requested independently compilable startup refactoring, while discussion with Dilip in #113, #115, and #121 exposed the risk of new backends writing before deferred initialization finished.
- Although Robert proposed five transition states in #123, Amul used asymmetric publication and barrier ordering in #125, publishing writable state only after initialization and prohibited state only after barrier absorption, while splitting out the startup prerequisite.
- Dilip's reviews in #128 and #131 produced comment and terminology fixes in #129 and #132, and Amul added a separate control-file-state helper refactor in #134.
- Robert found that the startup extraction prolonged InRecovery and required unsuitable dummy arguments from the checkpointer, so #135 replaced it with staged helper extraction, delayed WAL operations, and a separate XLogAcceptWrites extraction.
- Robert identified startup-local state dependencies in #136 and demanded stronger recovery testing in #141; Prabhat supplied tests atop the earlier full series in #142, and Amul integrated Robert's refactors and new TAP coverage in #143 and #144.
- After Robert identified test portability issues in #145, Amul expanded feature tests and moved recovery_end_command testing to separate work in #148.
- Mark Dilger reviewed critical-section safety, wait events, and interface behavior in #149, #151, and #154, then supplied isolationtester improvements and concurrent cursor tests in #158.
- Amul attempted to remove startup-only dependencies in #159, separated functional changes from mechanical refactoring in #162, and restored the full feature with Mark's feedback in #163, also reporting an intermittent WALWriteLock hang.
- Robert challenged the use of current control-file state and mutable WAL positions in #169 and #172; subsequent refactor-only submissions #171 and #174 reordered the work and used lastReplayedEndRecPtr to recognize recovery.
- Jaime Casanova reported an overwrite-contrecord test failure in #175, which Amul traced to the uncommitted dependency-removal patch in #176 and addressed in #177.
- Robert committed the first two components of #177 as 6df1543abf and 811051c2e7, preserving the helper extraction and adding only explanatory comments to the postponed-operation patch; #178 explicitly left the remaining refactors for further work.
- Robert's additional cleanup proposals in #178 led to Amul's standalone proof of concept in #179; Robert then proposed saving and restoring local WAL permission and used diagnostic patches in #180 to demonstrate that delaying timeline initialization would break existing assumptions.
- Amul incorporated the save-and-restore approach in #181, and Robert split and reordered its first two components into a75dbf7f9e and 18e0913a42, removing unnecessary StartupXLOG save-and-restore logic as confirmed in #182.
- Nathan Bossart reported a compiler warning and proposed initializing oldXLogAllowed in #183; Robert applied that discussion-only fix as a030a0c5cc and confirmed it in #184.
- Amul posted the two remaining refactors in #186 and restored the full feature in #187, explicitly omitting Mark's isolation tests because their expected error ordering was unstable.
Nov - Dec 2021 · 2 entries
- Robert showed why replay-state fields could not reliably replace the exact end-of-log values in #188, and Amul demonstrated the timeline mismatch in #191, leading #192 to export EndOfLog and EndOfLogTLI into shared memory.
- The restored feature in #193 avoided the checkpointer waiting on itself by invoking CreateCheckPoint directly and guarded against recursive initialization; #194 further changed when the checkpointer halted, with thorough testing still outstanding.
Apr 2022 · 1 entry
- After four months, #195 adapted the remaining series to the xlog.c and xlogrecovery.c split and revised reentrancy handling; Bharath's questions in #196 and Amul's response in #197 left interface naming and broader test coverage unresolved.
Jul 2022 · 1 entry
- Jacob Champion noted the earlier withdrawal in #198, and Amul confirmed in #199 that re-registration would wait for a design rework with no near-term schedule, leaving the prerequisite commits landed but the WAL-prohibition feature withdrawn and uncommitted.