Following up on this project, we're already seeing 2x performance over flink sql in NEXMark, even when reading from row-wise data sources like kafka!