High Throughput Replication with Integrated Membership Management

This paper introduces ChainPaxos, a new distributed consensus algorithm for high throughput replication. ChainPaxos organizes nodes in a chain, allowing for a pipeline communication pattern that maximizes throughput, by minimizing the number of messages transmitted. While other proposals have explored such patterns, ChainPaxos is the first that can execute linearizable reads in any replica with no communication overhead, relying only on information used to process updates. These techniques build on a fully specified integrated membership management solution, allowing ChainPaxos's fault-tolerance to be independent of an external coordination service, often used in other solutions, which can lead to possible safety violations in the presence of network partitions. Our evaluation shows that, when compared with alternative Paxos variants, ChainPaxos exhibits significantly higher throughput and scalability with negligible latency impact. Compared to other solutions with similar communication patterns, besides avoiding the costs of an external coordination service, ChainPaxos's high throughput tends to increase with the ratio of read-only operations.