as others mentioned you cannot really compare a concept (differential dataflow) with a product like Flink. I did not look in to differential dataflow so far, but my impression is that it would not handle all the problems Flink can handle. Such as handling late data, complex event processing with state and still getting correct results. E.g. if you scale and some of your node fails, how do you ensure correctness? This is a rather hard problem which for example Flink solves. If you do not solve that problem then basically you cannot scale to as many nodes (might not be necessary for some use cases) as for example Flink does