ML serving is about optimization and portability