Bulk entry writes for INT arrays¶
INT arrays with more than one element reserve child-vector capacity once before writing their values. Each element then uses the reserved range, and the child writer count advances once per array. Empty and singleton arrays retain the existing per-element loop. This reduces writer dispatch and capacity checks; it does not eliminate payload copies or claim lower allocation.
Parent list offsets, child NULL masks, nested arrays and fresh-batch ownership retain their existing behavior. Exported buffers remain retained and a fresh batch never overwrites them. Comet's Arrow writer pattern was consulted before this change. Ownership, transpose, failure-cleanup and runtime ARRAY_DISTINCT SQL checks pass 41 cases without skips on both released Flink 2.2.1 and 1.18.1, using the exact delivery Java sources and canonical release/mimalloc library.
Whole-job evidence¶
The final five comparisons use already-supported INT ARRAY_DISTINCT and a native library matching canonical production Rust, with no experimental array-cast implementation. Each candidate/original/original/candidate block starts a fresh JVM, performs two warmups and five alternating stock/native repetitions. Runs within a block share the JVM. Both row/Arrow transposes and the rowwise blackhole remain measured: Java 17, Flink 2.2.1, two CPUs, a 2 GiB heap, release/mimalloc, local zero-copy exchange disabled. No heavy work overlaps timed trials. Only the three entry-writer files vary. Times are pooled medians in seconds; negative percentages mean faster. NULL patterns give parent/element periods, with zero meaning no NULLs. The empty fixture explicitly verifies zero-length arrays; its domain and element-NULL settings have no effect on empty contents.
| Width | Rows | Domain | NULL pattern | New native | Prior native | Stock | Versus prior | Versus stock |
|---|---|---|---|---|---|---|---|---|
| 0 | 10,000,000 | 8 | 8/7 | 1.406787 | 1.415357 | 1.342771 | -0.61% | +4.77% |
| 1 | 10,000,000 | 65536 | 0/0 | 1.510447 | 1.504536 | 1.534712 | +0.39% | -1.58% |
| 8 | 10,000,000 | 65536 | 8/7 | 2.059181 | 2.116566 | 2.785315 | -2.71% | -26.07% |
| 64 | 2,000,000 | 65536 | 8/7 | 1.842782 | 1.972906 | 9.184806 | -6.60% | -79.94% |
| 64 | 2,000,000 | 8 | 8/7 | 1.098850 | 1.338459 | 2.703348 | -17.90% | -59.35% |
The width-eight and both width-64 native timing ranges do not overlap their previous native controls. Singleton results are near level with the previous writer (overlapping ranges), rather than an established improvement. Empty arrays also remain near level with the previous path and still lose to stock; this change preserves their existing loop and does not claim to accelerate them.
All 560 trials and all 56 engine summaries retain the original unbounded prototype, the revised writer with the experimental library, and the final canonical-library comparisons, including all unfavorable trials and timing ranges. The initial unbounded writer regressed singleton ARRAY_DISTINCT by 1.97%, prompting the small-array dispatch rule. Measurements on the experimental library are labeled separately and do not replace the final canonical-library evidence.
This improves an existing native operation; it does not change collection
function admission. ARRAY