model-writtenrun 8908 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- At Magnus Hagander's request, Andres Freund posted a global-barrier prototype that lets checksum state changes wait until participating processes have handled interrupts, avoiding a server restart and ensuring that in-progress page writes have finished #1.
- After a gap, Magnus examined the prototype and added wait-event reporting so stalled barrier waits would be observable #5; Andres accepted that addition and explained that long-running checkpoint work also needed better interrupt handling #6.
- Robert Haas took up the work in November, questioned the shared-memory-only background-worker change, identified problems with PID checks and the placement of BufferSync interrupt processing, and proposed callback-based extensibility #7.
- Andres welcomed Robert's takeover but argued that arbitrary callbacks were unsafe in low-level interrupt contexts and raised the need to defer acknowledgement of some barrier types #8.
- Robert separately demonstrated converting the autovacuum launcher to CHECK_FOR_INTERRUPTS() #9, but the later series preserved process-specific behavior while extracting and sharing interrupt-handling code #12.
- Robert committed the auxiliary-process signal-handler member of Andres's original series as 0d3c3aae33, reporting completion in #11; the original shared-memory-worker change was not included.
- Robert's replacement series moved all barrier state into ProcSignal, introduced latch waits with increasing timeouts, retained wait-event reporting, and separated three interrupt-handling cleanups from the barrier implementation and temporary test code #12.
- Andres's synchronization review identified a missing ordering guarantee after waiting #13, leading Robert to add a full memory barrier and adjust loop ordering in #14, while deferring selective acknowledgement because of its complexity and potentially long waits.
- Daniel Gustafsson read and tested the revised series, successfully ran online-checksum tests on top of it, and exercised the API with experimental GUC barriers #15.
- Robert and Andres continued debating extension callbacks and how to land infrastructure before its first client #16 #17 #18, ultimately choosing the placeholder barrier suggested by Andres #19.
- The three cleanup members of #14 landed separately without substantive diff changes: interrupt extraction as 5910d6c7e3, SIGHUP-handler reuse as 1e53fe0e70, and broader handler deduplication as 7dbfea3c45.
- Robert then posted the remaining barrier patch with a documented no-op placeholder, leaving out the temporary test function so the infrastructure could land independently of a real client #20.
- Sergei Kornilov reported compiler errors in the newly committed interrupt-extraction code and supplied prototype fixes #21; Robert explained that he had already pushed two equivalent hunks following Andrew Gierth's report, and Sergei confirmed a clean build #22 #23.
- The remaining barrier patch #20 landed as 16a4e4aecd with only an additional documentation-table correction, and Robert announced completion #24.
- At the thread's close, the infrastructure had landed, but Robert revisited deferred acknowledgement after identifying possible time-of-check/time-of-use problems in ALTER SYSTEM READ ONLY and online checksums; safe interaction with critical sections and cached checksum state remained unresolved #26.
- The later online-checksum feature commit f19c0eccae builds on this infrastructure, but its attached-patch lineage belongs to the newer work in thread #41134, rather than to a checksum implementation submitted here.