model-writtenrun 8092 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Mar - Apr 2020 · 8 entries
- Andres Freund opened the thread with #1, identifying frequently modified xmin cache lines and indirect PGXACT access as major snapshot bottlenecks, and proposing deferred visibility horizons plus dense arrays for transaction state.
- Andres added experimental snapshot caching in #4, allowing reuse when no transaction had completed since the preceding snapshot.
- David Rowley reviewed the initial patches in #6, while Thomas Munro challenged unlocked counter reads, 64-bit atomicity and horizon conversion in #7; Andres accepted the naming and atomicity concerns in #8.
- Old-snapshot handling required a larger redesign: Andres proposed separating removal thresholds from cancellation decisions in #13, then posted #24 with combined horizon computation, conditional old-snapshot processing, FullTransactionId tracking and cache reuse under ProcArrayLock.
- Benchmarking by David in #20 and Andres in #21 exercised read, write, connection and prepared-transaction workloads, while the separate pgbench attachment in #25 was explicitly supplied for testing rather than commitment.
- Robert Haas's reviews in #27, #29, #35 and #39 drove clearer horizon terminology, removal of Copy suffixes, stronger locking documentation and revised patch boundaries.
- The independently useful delayChkpt relocation from #24 landed unchanged as 75848bc744, reducing unnecessary writes to cache lines used for snapshot acquisition.
- Andres posted the reduced and polished series in #42 but declined to push the large change while exhausted at feature freeze; discussion in #45 and #46 weighed its benefits against limited stabilization time.
Jun - Oct 2020 · 16 entries
- After the release-cycle gap, Andres revived the work with the rebased series in #56, and Alvaro Herrera requested the FullTransactionId prerequisite split and identified a duplicate declaration in #57.
- The revision in #58 extracted that prerequisite and corrected a vacuum failure Thomas had highlighted by keeping pruning's approximate horizons consistent with accurately computed vacuum horizons.
- Andres's proposal in #58 to remove redundant Full naming produced fea10a6434, followed by the extracted tracking patch as 3bd7f9969a with light naming and expression changes after Thomas's review in #64.
- Andres posted the remaining six patches in #69 and planned staged commits separated by buildfarm cycles, leaving further optimizations for later work.
- The deferred-horizon component of #69 landed as dc7420c2c9, including a final conversion of HeapTupleSatisfiesNonVacuumable(), and the xmin relocation followed as 1f51c17c68 with documentation cleanup.
- The dense-XID, vacuumFlags and subtransaction components of #69 landed as 941697c3c1, 5788e258bb and 73487a60fc, with final edits retaining an invalid-XID guard and correcting the prepared-transaction cached subtransaction count.
- Tom Lane reported freeze-the-dead buildfarm failures in #70, prompting Andres's temporary vacuum diagnostics in 49967da65a after the instrumentation plan in #72.
- Tom's delayed test in #75 made the failure reproducible, allowing Andres to identify a dense-array index error and post the one-line correction in #78, committed as f6661d3df2.
- The snapshot-cache component of #69 then landed as 623a9ba79b with only a comment correction, retaining ProcArrayLock during cache validation.
- Andres closed the main commitfest entry in #86, explicitly deferring lock-free cache reuse to separate work; performance investigation in #87, #90 and #101 confirmed idle-connection gains while exposing other workload bottlenecks.
- Ian Lawrence Barwick reported in #97 that persistent standby sessions never saw newly replicated rows because replay did not advance xactCompletionCount, and Andres explained in #98 that existing one-shot psql tests had missed the problem.
- Ian supplied a dblink-based reproducer in #104 and replay-site counter increments in #105; Andres combined and substantially revised them in #106, moving invalidation under the existing recovery lock and testing persistent psql sessions, ordinary transactions and prepared transactions.
- The test-infrastructure discussion in #107, #108 and #109 considered interactive psql helpers and platform dependencies, but Andres retained IPC::Run and a local helper, polishing its matching and timeout behavior in #111.
- The combined standby fix and test in #111 landed unchanged as 7b28913bca, crediting Andres and Ian as authors, and Ian later confirmed his original application tests passed in #119.
- MSYS test timeouts reported in #113 led Andrew Dunstan to propose CRLF normalization in #114 and verify it on jacana in #118; Andres committed that inline fix as 1df2b50dbe.
- The main series and standby fix were committed, but Andrew reported an unresolved MSVC test hang in #121, and Andres's broader PerlIO cleanup proposal in #120 remained a discussion item.
Feb - Mar 2021 · 1 entry
- The thread ended with connection-pooling questions in #122 and Konstantin Knizhnik's explanation in #124 that these snapshot improvements do not eliminate contention on other resources or the usefulness of pooling.