Streaming Iceberg Nexmark observations

2,000,000 deterministic events; parallelism 4; four-bucket input/output tables. All times include streaming completion and final drain/commit.

Query	State / mini-batch	Variant	Flink median [range], s	StreamFusion median [range], s	Median ratio
q3	disk-off	exact	0.644 [0.644–0.644]	0.651 [0.651–0.651]	0.99×
q3	memory-off	exact	0.737 [0.737–0.737]	0.713 [0.713–0.713]	1.03×
q3	memory-on	exact	0.604 [0.604–0.604]	0.605 [0.605–0.605]	1.00×
q4	disk-off	exact	72.896 [72.896–72.896]	16.294 [16.294–16.294]	4.47×
q4	memory-off	exact	3.431 [3.431–3.431]	16.779 [16.779–16.779]	0.20×
q4	memory-on	exact	10.680 [10.680–10.680]	28.032 [28.032–28.032]	0.38×

Exact-query aggregates

Geometric means use every exact-query cell in each mode; labelled expression variants remain separate above. Throughput is nominal corpus events divided by median end-to-end time, including startup, checkpoints and drain.

State / mini-batch	Queries	Geomean ratio	Flink corpus events/s, range	StreamFusion corpus events/s, range
disk-off	2	2.10×	27,436–3,103,596	122,746–3,071,904
memory-off	2	0.46×	582,950–2,714,470	119,200–2,803,333
memory-on	2	0.62×	187,259–3,311,161	71,347–3,308,051

Raw samples, warmups, parity checks, plans, committed table locations and native/fallback task/file witnesses are retained in this artifact directory. PROCTIME q12 validates per-bidder count conservation and includes waiting for the final processing-time window.