The design of Wren, a Fast and Scalable Transactional Causally Consistent Geo-Replicated Key-Value Store

This paper presents the design of Wren, a new geo-replicated key-value store that achieves Transactional Causal Consistency. Wren leverages two design choices to achieve higher performance and better scalability than existing systems. First, Wren uses hybrid logical physical/clocks to timestamp data items. Hybrid clocks allow Wren to achieve low response times, by avoiding the latencies that existing systems based on physical clocks incur to cope with clock skew. Second, Wren relies on a novel dependency tracking and stabilization protocol, called Hybrid Stable Time (HST). HST uses only two scalar values per update regardless of the number of data centers and nodes within a data center. HST achieves high resource efficiency and scalability at the cost of a slight increase in remote update visibility latency. We discuss why Wren achieves higher performance and better scalability than state-of-the-art approaches.