A short answer: It's a recent approach to incremental computation geared towards data-science/db people. Imagine you want to maintain the connected components of a graph where edges are being added and removed. It turns out you can express this using dataflow and get reasonable looking runtimes (30,000 edge ins/del per second) [1]. You could probably do much better with a careful implementation of parallel union-find for just edge-insertions but part of the appeal of data-flow is that you express your computation using functional primitives like joins, maps and filters and then the incremental stuff 'just works'.

[1] http://www.frankmcsherry.org/differential/dataflow/2015/05/1...