Improving Raft When There Are Failures

Christian Fluri, Darya Melnyk, Roger Wattenhofer

Abstract: This paper investigates the Raft consensus algorithm in the presence of failures. We are especially interested in how the single failures - link failures, isolation and partition-affect the running time of the leader election, which is an important building block of Raft. Our tests show that such failures are non-negligible. We therefore propose new timeout policies which can improve the performance of Raft.