Comparison with previous StreamFusion Iceberg I/O

Both legs execute native query operators. The baseline uses stock Iceberg I/O, including row/Arrow transposes around native query islands; the native leg keeps data columnar through Rust Parquet reads and data-file writes. q0 has no query island, so its baseline is the stock identity pipeline rather than invented transpose operators. Graph assertions cover the other three queries.

This separate release+mimalloc cohort uses the final Nexmark completion protocol, 2M events, four input/output buckets, parallelism four, 4 GiB job heaps, one warmup and three measured pairs for q0/q3/q4/q9 in memory/off and disk/off. All 32 paired validations pass. It is excluded from the stock-Flink matrix.

Every median/range and raw observation is retained. q0 improves 1.80× memory/off and 1.78× disk/off. The four-query geometric means are 1.90× memory/off and 0.95× disk/off. Slower results remain visible: q4 memory/off is 0.79×, q4 disk/off 0.95×, and q9 disk/off 0.49×.

Phase metrics separate corpus progress, checkpoint and drain. In q4 memory/off, native median corpus progress is faster (1.573 versus 2.373 s), but checkpoint/drain waits erase that gain. In q9 disk/off, native corpus progress is slower (29.560 versus 13.139 s), as well as its settling checkpoint; this needs state/I/O investigation rather than claiming every regression is final commit. The additional q9 profile keeps query execution fixed and retains the much slower profiled native observation.

The I/O paths therefore do not establish a universal end-to-end speedup over previous StreamFusion behavior. Supported contracts now accelerate by default when the optional module is installed. Further state/checkpoint and input scheduling work is tracked in issue #391.