model-writtenrun 6258 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Thomas Munro proposed running bgwriter and checkpointer during crash recovery in #1, aiming to move buffer eviction and sync work out of startup while making crash recovery resemble streaming recovery; he treated adaptive bgwriter tuning as separate work and raised optional checkpoint waiting as a possible extension.
- Tom Lane argued in #2 that reliability mattered more than recovery speed and explained that waiting for a checkpoint limits repeated WAL replay after another crash, while acknowledging that running the usual background processes could make recovery better tested.
- Simon Riggs supplied independent buffer-statistics instrumentation in #3 to measure whether recovery exceeded shared buffers, and Thomas reported external testing showing a 1.64-fold speedup with aggressive tuning in #4.
- Robert Haas supported unifying sync handling and suggested using RequestCheckpoint() for standalone recovery too in #5; Thomas adopted that simplification in #6, combined the core and statistics patches, and added an experimental checkpoint-wait GUC whose double-crash safety implications remained unresolved.
- Aleksander Alekseev reported passing tests and supported the core and statistics patches in #7, but recommended separating the GUC; Robert endorsed the core patch with one explanatory comment and argued that asynchronous recovery should follow promotion's end-of-recovery-record approach in #8.
- Robert proposed folding buffer statistics into the existing recovery log and removing potentially confusing percentages in #8, but Andres Freund defended CPU-time reporting as useful diagnostic evidence in #9; Robert accepted that argument in #11, and Jakub Wartak advocated richer statistics and periodic progress reporting in #14.
- Thomas committed only the core component of #6 as 7ff23c6d27, adding Robert's requested comment, and stated in #10 that the statistics and optional-checkpoint patches needed further consideration.
- Robert supplied a follow-up patch collapsing duplicate checkpoint branches in #12; Amul Sul endorsed it and proposed removing the promoted flag in #13, while Robert preferred retaining the flag's role and considering a clearer name in #15.
- Thomas approved the cleanup in #17 and committed #12 unchanged as 8f7c8e2bef, announcing the landing in #18.
- A year later, Justin Pryzby reported in #19 that handing checkpoint work to checkpointer left startup's process title incorrectly displaying the last WAL file being recovered, and supplied an inline patch clearing that title in #20.
- Michael Paquier suggested displaying end-of-recovery activity instead in #21, but Justin argued in #22 for clearing stale information immediately and leaving richer status reporting to future work, then requested a PostgreSQL 15 fix in #23.
- Michael implemented Justin's inline fix with an explanatory comment as fbb5f54b67, backpatched it to PostgreSQL 15, and confirmed the repair in the final message #24; the core change, cleanup, and status regression fix had landed, while this thread records no landing of the statistics or optional-checkpoint proposals.