model-writtenrun 5481 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Feb - Mar 2020 · 3 entries
- Kyotaro Horiguchi opened the thread with #1, proposing concurrent execution of postgres_fdw children of Append through three patches covering WaitEventSet ownership, executor infrastructure, and foreign scans, including scans sharing a connection.
- Thomas Munro questioned changes to the ExecProcNode contract and wait-event bookkeeping in #4, while Horiguchi explained in #5 that broader asynchronous joins would require substantially more executor machinery.
- Movead Li confirmed useful performance and successful tests in #10, and Andrei Lepikhov's cleanup and costing review in #14 led to revised startup-cost handling in #15.
Jun 2020 - Jun 2021 · 24 entries
- Horiguchi limited the initial scope to direct ForeignScan children after discussion of more complex nodes in #22, then added ordered-Append protection and revised total costing in #23.
- Following the summer pause, #28 rebased the series while retaining explicit resource ownership, despite Thomas Munro's suggestion in #25 to simplify ownership registration.
- Konstantin Knizhnik reported unfinished remote commands after LIMIT in #29 and supplied successive fixes in #30 and #31 for result draining and stale leader state; the subsequent full-series rebase #40 did not incorporate those deltas.
- Etsuro Fujita argued in #36 and #41 for preserving ExecProcNode semantics and combining Robert Haas's callback interfaces with ideas from Horiguchi and Munro, and Horiguchi accepted that direction in #44.
- Lepikhov supplied a mixed synchronous/asynchronous connection fix in #46 and a separate SubqueryScanPath extension in #47, the latter becoming the starting point for later planner work.
- Fujita took over the full implementation in #50, introducing per-Append event management, demand-time tuple requests, synchronous EvalPlanQual handling, and temporary restrictions on problematic shared-connection cases.
- Horiguchi's executor and result-handling review in #52 and #53, Movead Li's multi-server assertion report in #54, and Lepikhov's TPC-H review in #56 exposed problems in the revised implementation.
- Fujita substantially revised scheduling in #65, restoring requests to all needy subplans together, removing the extra begin callback and last-plan scheduling state, and returning responses from the requestee side.
- After initialization and pruning fixes in #70, Fujita tried executor-startup connection tracking and a ReconsiderAsyncForeignScan callback in #72, completing the attachment in #73.
- That approach was abandoned when Fujita realized in #76 that plan inspection could not account for foreign scans invoked inside functions; #79 instead completed pending commands before another operation used the same connection.
- The discussion in #78 left aggressive prefetching for future work because of LIMIT overhead and potentially growing tuple buffers, while same-server slowdown measurements in #80 led to the default-false async_capable server/table option in #81.
- Fujita explicitly left resource-owner changes and asynchronous cost estimation out in #82, because wait sets were allocated and freed per wait and shared connections invalidated the proposed maximum-subplan-cost assumption.
- Dedicated tests and callback refinements in #83, followed by Justin Pryzby's review in #84, produced the final attachment #87, which landed as 27e1f14563 with only comment edits.
- Tom Lane's post-landing Valgrind report in #91 led Fujita to bound the event count in #93; after Horiguchi's questions in #94, that patch landed unchanged as a8af856d32.
- Horiguchi's earlier concern about capability merging in #88 was addressed by the explanation and small executor cleanups in #97, committed unchanged as bb684c82f7.
- Lepikhov raised the unchanged optimizer costs in #99, but Fujita left connection-aware costing for future work in #102 to keep the initial implementation simple.
- Lepikhov reported empty asynchronous scans incorrectly labeled never executed in #100; Fujita replaced the proposed extra ExecProcNode call with full asynchronous callback instrumentation in #105.
- In parallel, Lepikhov proposed selecting asynchronous subplans during executor initialization in #101 to benefit from trivial-subquery removal, and reposted the enhancement in #115 after Fujita deferred it to the next development cycle in #106.
- A LIMIT-triggered instrumentation failure reported in #109 was fixed in #113, which landed unchanged as a363bc6da9.
- For the shared-connection performance caution proposed in #104, Stephen Frost supplied wording explaining consistency constraints in #107; Fujita added the user-mapping exception in #122 and committed it unchanged as 15fcd33e06.
- Fujita found incorrect results and assertions when rescanning without runtime pruning in #124; Horiguchi's initialization feedback in #125 produced #127, committed unchanged as f3baaf28a6.
- Lepikhov's request in #119 to explain mixed synchronous/asynchronous tuple delivery became the documentation patch #132, committed with only line wrapping as eab8195368.
- Lepikhov rewrote the executor-initialization proposal in #134, separating plan capability from actual execution mode, but Fujita demonstrated in #135 that it could break ordered Appends and argued against repeating planning work during execution.
- Fujita therefore returned to Lepikhov's older planner proposal #47, producing #135 with cached trivial-subquery checks and extending it in #137 to projections that require no separate Result node.
Aug - Sep 2021 · 1 entry
- Alexander Pyhalov reported successful review and testing in #138; after a long delay, Fujita refreshed the patch in #140, and Zhihong Yu's naming and documentation suggestions in #141 were incorporated into c2bb02bc2e.
Mar - Jun 2022 · 5 entries
- Justin Pryzby then reported a planner crash from a reduced sqlsmith query in #146, which Fujita traced in #149 to treating a gating Result node as the underlying SubqueryScan or ForeignScan.
- Pryzby confirmed the fix in #152, while discussion with Zhihong Yu concluded in #157 with agreement to retain Result checks inside individual path-type branches.
- Fujita supplied polished development and PostgreSQL 14 variants in #158; the development version landed with only two test-comment edits as 5c854e7a2c, and the same fix was backpatched.
- Finally, Fujita documented why early trivial_subqueryscan checks are safe in #160, adopted Zhihong Yu's formatting feedback in #162, and committed that attachment unchanged as 4a8a5dd7f5.
- The final message #163 confirms the comment patch was committed, completing the selected feature and follow-up fixes; the executor-time alternative was superseded in #135, while broader asynchronous node support, aggressive prefetching, and improved costing remained future work.