Closing The Performance Gap between Causal Consistency and Eventual Consistency

It is well known that causal consistency is more expensive to implement than eventual consistency due to its requirement of dependency tracking and checking for causality. To close the performance gap between the two consistency models, we propose a new protocol that implements causal consistency for both partitioned and replicated data stores. Our protocol trades the visibility latency of updates across different client sessions for higher throughput. An update, either from a local client or a remote replica, is only visible to other clients after it is replicated by all replicas. As a result, a read operation never introduces dependencies to its client session. Only update operations introduce dependencies. By exploiting the transitive property of causality and total order update propagation, an update always has at most one dependency. By reducing the number of tracked dependencies and the number of messages for dependency checking down to one, we believe our protocol can provide causal consistency with similar cost to eventual consistency.