model-writtenrun 8557 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jul - Aug 2024 · 2 entries
- Alexander Pyhalov proposed asynchronous MergeAppend in #1 to avoid losing remote-query concurrency when the planner switched from Append to MergeAppend, initially borrowing Append machinery while preserving per-subplan ordering.
- Alena Rybakina requested broader tests and clearer design explanations in #2; Pyhalov added join and pruning tests and reported benchmark improvements in #3.
Jul 2025 · 1 entry
- After nearly a year, Pyhalov returned with #4, restricting execution to valid async subplans and adding interrupt handling while waiting.
Oct 2025 - Apr 2026 · 9 entries
- Álvaro Herrera rebased the work for the new GUC representation in #5, and Matheus Alcantara supplied build and expected-output fixes together with favorable benchmark results in #6.
- Alcantara's review in #8 exposed duplicated classification logic, unused accessors, and missing rescan coverage; Pyhalov addressed these in #9, explaining why MergeAppend must wait for a specific subplan rather than accept any available tuple.
- Further review in #10 and #12 refined node-type checks and comments through #13, while Pyhalov explained why results must remain indexed by subplan and why cached-result checks cannot be skipped.
- Alcantara introduced shared Appender/AppenderState types and execAppend.c in #14, expanding the series into a substantial executor refactoring.
- Pyhalov reviewed and tested that infrastructure in #15 and #17, leading to corrected bitmap cleanup and allocations; the refinement concluded with moving the classification helper into execAppend.h in #20.
- Pyhalov maintained the series through rebases ending at #24, after which Alexander Korotkov reordered it so infrastructure preceded the feature in #25.
- Korotkov substantially simplified execution in #27, splitting the infrastructure into three patches, using AppendBase types and existing merge slots, introducing a distinct wait event, and switching steady-state tuple fetching to ExecProcNode().
- Richard Guo and Etsuro Fujita objected to the short pre-freeze review window in #28 and #29; Korotkov acknowledged postponement as an option in #32, and the work continued beyond that cycle.
- Alcantara tested the simplified implementation in #30 and #34, while Pyhalov reported additional performance gains in #36.
Jul - Aug 2026 · 3 entries
- Korotkov subsequently found that ExecProcNode() on async postgres_fdw scans truncated results at the first batch, so #37 restored async fetching, added batch and rescan tests, cleared stale slots, and bundled an independently discussed Append rescan fix.
- After further testing and discussion, #41 replaced lazy pending-request handling with explicit draining shared between Append and MergeAppend.
- Fujita questioned whether MergeAppend could have pending requests at rescan in #43, and Pyhalov confirmed it could not in #44; Korotkov consequently removed unnecessary draining machinery and the bundled Append fix in #46.
Oct - Nov 2026 · 6 entries
- Pyhalov rebased the surviving series in #48, and solai v verified async plan selection and complete retrieval of 10,000 rows in #51.
- Fujita rejected the broad shared-base refactoring in #52, arguing that Append and MergeAppend have different async requirements and that sharing only the event-loop machinery would better support future maintenance.
- Fujita's replacement in #53 used specialized MergeAppend pending-request state, shared ExecAsyncEventWait(), and revised FDW wait configuration; it retained the latest tests while dropping the broad infrastructure and making the separate planner safeguard unnecessary.
- Pyhalov's review in #54 prompted allocation cleanup and documentation of the changed FDW callback contract, while discussion in #56 clarified rescan-test intent and supported removing obsolete cases; Fujita completed these refinements in #57.
- Pyhalov separately tried enabling async ordered Append scans in #59, but Kirill Reshke demonstrated an incorrect ORDER BY result in #63, and Pyhalov withdrew that approach in #64 because async Append does not preserve order.
- Fujita committed #57 as 11591f273e with light editorial changes, rebasing, and removal of a redundant runtime-pruning test, then announced closure of the commitfest item in #65; the specialized MergeAppend feature landed, while the broad refactoring and ordered-Append experiment were left out.