I'm writing a relational compiler. It makes Spark jobs run up to 10x faster on the same hardware.