DataFusion + Rust + Arrow as the data plane

Applies to: every native operator

The JVM stays the control plane — planning, coordination, checkpoint orchestration — while record processing moves to native code operating on Arrow column batches. Columnar batches turn per-record interpretation into vectorized kernels, Rust removes GC pressure from the hot path, and DataFusion supplies maintained, optimized compute (accumulators, physical expressions, hash joins, file scans) instead of hand-rolled kernels.

Every native operator either runs on DataFusion compute or is custom only where Flink parity forces it.