Improving the Scalability of Geo-replication with Reservations

Geo-replicated systems improve performance and fault tolerance by replicating data on sites in different physical locations. Because of performance and scalability costs of strong consistency, these systems rely on eventual consistency that improves performance but might violate application invariants. In this work, we exploit reservation techniques to strengthen eventual consistency, by protecting specific invariants without adding synchronization in the critical path. We define a consistency model called RPB that retains the advantages of eventual consistency, while enforcing stronger guarantees, including causality and safety properties.