model-writtenrun 8184 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Thomas Munro opened the thread with a revived proof of concept that lets Append consume whichever immediate foreign-scan child becomes ready, using a small readiness API and postgres_fdw sockets; parallel execution was disabled in this version #1.
- Munro explained that remote shards could work concurrently while the local executor remained serial, and that integrating Parallel Append would require solving worker ownership of scans and connections; Bruce Momjian considered the restricted feature useful for aggregation and selective queries #3 #4.
- Kyotaro Horiguchi proposed reviving his broader asynchronous-execution work, while Munro and Momjian supported pursuing a useful implementation without waiting for an executor rewrite; Robert Haas made avoiding regressions for ordinary Append queries a central requirement #5 #6 #7 #8.
- Horiguchi posted his independent three-patch series with WaitEventSet resource ownership, planner-selected asynchronous subplans, a separate ExecAppendAsync path, and postgres_fdw support; his subsequent benchmarks reported substantial foreign-scan gains without a consistent slowdown for local scans #10 #12.
- Ahsan Hadi forwarded Movead Li’s distributed benchmarks showing strong gains, especially when few rows returned to the parent; later discussion identified a throughput plateau as shard counts increased, while Munro requested confirmation of which implementation had been tested #13 #16 #17 #18 #19.
- After several months, Etsuro Fujita rebased Munro’s implementation and added regression-test ordering fixes, while identifying shared-connection failures and the need to exclude ordered Appends from asynchronous execution; he favored planner-time eligibility checks and continued comparing the two approaches #21.
- Fujita withdrew his initial claim that Horiguchi’s implementation failed the same shared-connection query, and Horiguchi explained how vacate_connection() and its waiting queue serialize competing requests #22 #23.
- The thread ended with unresolved correctness and performance concerns: Fujita reported that adding LIMIT produced an outstanding-command error during CLOSE and questioned the cost of waiting for competing requests on shared connections #24.
- The eventual asynchronous-execution commit credits contributions from this discussion, but its final matching patch was posted later in thread #42002 as #473706; 27e1f14563 therefore belongs to that later integration rather than a direct landing of an attachment here.