Towards Seamless Resynchronization for Active-Active Database Clustering

Database clustering is a well-established technology to improve the availability and scalability of a database service. Query replication and log replication are two popular methods for propagating database updates across a cluster of database servers. An important issue that received relatively scant attention and is the focus of this paper is how to add a new server to an active-active database cluster with minimum service disruption while maximizing the cluster's request processing concurrency during run time. Even though log replication allows attaching a new database server without stopping the attached database cluster's service, the asynchronous nature of its operation is incompatible with active-active clusters, where every cluster node is designed to share the load of servicing incoming read query queries. In contrast, query replication is synchronous and thus could readily support active-active clustering, but most existing query replication implementations need to halt a cluster's service when attaching a new server. In this paper, we present a database resynchronization scheme that achieves the best of both worlds: leveraging log replication to minimize the service disruption time associated with addition of new servers, while applying query replication during run time to maximize the parallelism of read query processing, and demonstrate its effectiveness in a product-grade database engine, PostgreSQL.