Generic Checkpointing Support for Stream-based State-Machine Replication

Abstract

Stream-based replication facilitates the deployment and operation of state-machine replication protocols by running them as applications on top of data-stream processing frameworks. Taking advantage of platform-provided features, this approach makes it possible to significantly minimize implementation complexity at the protocol level. To further extend the associated benefits, in this paper we examine how the concept can be used to provide generic support for creating, storing, and applying checkpoints of replica states, both in the use case for catch up and garbage collection as well as to recover failed replicas. Specifically, we present three checkpointing-mechanism designs with different degrees of platform involvement and evaluate them in the context of Twitter's stream-processing engine Heron.