This is maybe a silly question, but what's the difference between timely dataflow and Spark's execution engine? From my understanding they're doing very similar

This is maybe a silly question, but what's the difference between timely dataflow and Spark's execution engine? From my understanding they're doing very similar things - break down a sequence of functions on a stream of data, parallelize them on several machines, and then gather the results. I understand that the feature set of timely dataflow is more flexible than Spark - I just don't understand why (I couldn't figure it out from the paper, academic papers really go over my head).