Benchmarking Replication and Consistency Strategies in Cloud Serving Databases: HBase and Cassandra

Databases serving OLTP operations generated by cloud applications have been widely researched and deployed nowadays. Such cloud serving databases like BigTable, HBase, Cassandra, Azure and many others are designed to handle a large number of concurrent requests performed on the cloud end. Such systems can elastically scale out to thousands of commodity hardware by using a shared nothing distributed architecture. This implies a strong need of data replication to guarantee service availability and data access performance. Data replication can improve system availability by redirecting operations against failed data blocks to their replicas and improve performance by rebalancing load across multiple replicas. However, according to the PACELC model, as soon as a distributed database replicates data, another tradeoff between consistency and latency arises. This tradeoff motivates us to figure out how the latency changes when we adjust the replication factor and the consistency level. The replication factor determines how many replicas a data block should maintain, and the consistency level specifies how to deal with read and write requests performed on replicas. We use YCSB to conduct several benchmarking efforts to do this job. We report benchmark results for two widely used systems: HBase and Cassandra.