Practical Experience Report: Cassandra+: Trading-Off Consistency, Latency, and Fault-tolerance in Cassandra

Abstract

The exponential growth of data has pushed the industry towards new types of data management systems such as NoSQL database. Our goal is to augment NoSQL, with the focus on its application as a distributed key-value store (KV-store). Existing production-ready systems often provide only probabilistic guarantees on consistency and fault-tolerance, and may violate their correctness properties if a cluster has severe clock drift. Our system Cassandra+ addresses these issues by providing more choices for consistency and fault-tolerance. We build Cassandra+ by implementing theoretical distributed shared memory (DSM) in Cassandra, one of the most popular NoSQLs in the industry. In this paper, we share our experience in adapting and implementing DSM algorithms into a real-world system. We hope our experience and results allow a better understanding of the DSM and the tradeoffs involved.