Scalable fault tolerance for high-performance streaming dataflow

Abstract

Streaming dataflow systems offer an appealing alternative to classic MySQL / memcached web backend stacks. But websites must not go down, and current fault tolerance techniques for dataflow systems either come with long downtimes during recovery, or fail to scale to large deployments due to the overhead of global coordination. For example, in the failure of a single dataflow node, existing lineage-based techniques take a long time to recompute all lost and downstream state, while checkpointing techniques require costly global coordination for rollback recovery. This thesis presents a causal logging approach to fault tolerance that rolls back and replays the execution of only the failed node, without any global coordination. The key to knowing how to replay a valid execution while ensuring exactly-once semantics is a small, constant-size tree clock piggybacked onto each message, incurring runtime overheads that are low and scalable. After recovery, the state of the system is indistinguishable from one that never failed at all. We implement and evaluate the protocol on Noria, a streaming dataflow backend for read-heavy web applications. Compared to Noria's original protocol of lineage-based recovery, tree clock recovery time is constant in relation to state size and graph size. Experimental results show sub-second recovery times with 1.5ms runtime overheads, which translates to a 290x improvement in recovery time.