Designing a causally consistent protocol for geo-distributed partial replication

Abstract

Modern internet applications require scalability to millions of clients, response times in the tens of milliseconds, and availability in the presence of partitions, hardware faults and even disasters. To obtain these requirements, applications are usually geo-replicated across several data centres (DCs) spread throughout the world, providing clients with fast access to nearby DCs and fault-tolerance in case of a DC out-age. Using multiple replicas also has disadvantages, not only does this incur extra storage, bandwidth and hardware costs, but programming these systems becomes more difficult. To address the additional hardware costs, data is often partially replicated, meaning that only certain DCs will keep a copy of certain data, for example in a key-value store it may only store values corresponding to a portion of the keys. Additionally, to address the issue of programming these systems, consistency protocols are run on top ensuring different guarantees for the data, but as shown by the CAP theorem, strong consistency, availability, and partition tolerance cannot be ensured at the same time. For many applications availability is paramout, thus strong consistency is exchanged for weaker consistencies allowing concurrent writes like causal consistency. Unfortunately these protocols are not designed with partial replication in mind and either end up not supporting it or do so in an inefficient manner. In this work we will look at why this happens and propose a protocol designed to support partial replication under causal consistency more efficiently.