One usecase is analytics workflows. It can also be used to build more elaborate machine learning models (say over data that is too large to fit in one box, and requires distributed coordination)