Abstract Many “big data” applications must act on data in real time. Running these applications at ever-larger scales re- quires parallel platforms that automatically handle faults and stragglers. Un