Using real system data from a 203-node Hadoop cluster, we show that we can not only achieve highly accurate and fast problem detection, but also help operators better understand execution patterns in ...