model-writtenrun 6102 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Dilip Kumar opened the thread with measurements suggesting that shared-memory tuple transfer substantially limited Gather performance, proposed batching counter updates, and discussed dynamically growing queues for future worker redistribution in #1, then supplied a 4kB-threshold prototype in #2.
- Robert Haas observed that the measurements could justify revisiting parallel_tuple_cost in #3, and Dilip supplied broader benchmarks plus a redesigned patch using pending-byte accounting, a quarter-ring batching threshold, and per-call forced flushing in #4.
- Zhihong Yu suggested aligning field order with documentation and checking force_flush before the batching threshold in #5 and #6, while Tomas Vondra raised startup-latency concerns and requested clarification about the patch versions in #7.
- Andres Freund questioned profiling unoptimized builds and suggested avoiding division and separating frequently modified queue data from read-mostly fields in #8; Dilip clarified that benchmark timings already used optimized builds and agreed to improve profiling in #10 and #12.
- Dilip resent the unchanged refined batching patch alongside an experimental queue-size GUC in #9, explicitly explaining that configurable queue size was a testing aid rather than a proposed production feature.
- Tomas reported 20–30% improvements in larger optimized-build benchmarks in #13, and Dilip reported reduced cache misses, no consistent regression in a targeted latency experiment, no additional gain from rearranging shared fields, and a non-power-of-two obstacle to replacing division with masking in #15.
- Dilip incorporated Zhihong's small ordering and condition changes in the final attached patch #16, and Tomas's queue-size experiments supported retaining the existing 64kB queue and deferring size tuning to separate work in #17.
- Robert initially favored committing in #18, but Tomas reported suspicious results on a dual-socket machine in #19, prompting discussion of batching latency, tuple size, and possible alternative thresholds in #20 and #21.
- Dilip could not reproduce the slowdown and found the quarter-ring threshold generally outperformed smaller batches in #29 and #30, while Robert explained how delayed publication could occasionally make receivers sleep unnecessarily in #31.
- Dilip identified an unusually fast unpatched baseline in #32, Amit Kapila encouraged repeating the test in #33, and Tomas's expanded rerun brought the anomalous case into line with the other results in #37.
- Robert committed #16 as 46846433a0 with only comment edits, then announced the landing and identified zero-byte publication checks and reassessment of parallel_tuple_cost as follow-up questions in #39.
- The thread ended with the CommitFest entry marked committed in #41; the batching optimization had landed, while queue-size tuning and the broader future optimizations remained outside this change.