Tunable Causal Consistency: Specification and Implementation

TL;DR

This paper introduces tunable causal consistency (TCC), a flexible model allowing clients to specify session guarantees per operation, with a protocol implementation demonstrating low latency and high throughput in a distributed key-value store.

Contribution

It formally defines TCC, designs a protocol, and implements TCCSTORE, enabling customizable consistency levels with efficient performance in distributed systems.

Findings

Latency under 38ms for all workloads

Throughput up to 2800 operations/sec

Better performance than causal consistency with negligible overhead

Abstract

To achieve high availability and low latency, distributed data stores often geographically replicate data at multiple sites called replicas. However, this introduces the data consistency problem. Due to the fundamental tradeoffs among consistency, availability, and latency in the presence of network partition, no a one-size-fits-all consistency model exists. To meet the needs of different applications, many popular data stores provide tunable consistency, allowing clients to specify the consistency level per individual operation. In this paper, we propose tunable causal consistency (TCC). It allows clients to choose the desired session guarantee for each operation, from the well-known four session guarantees, i.e., read your writes, monotonic reads, monotonic writes, and writes follow reads. Specifically, we first propose a formal specification of TCC in an extended (vis,ar) framework originally proposed by Burckhardt et al. Then we design a TCC protocol and develop a prototype distributed key-value store called TCCSTORE. We evaluate TCCSTORE on Aliyun. The latency is less than 38ms for all workloads and the throughput is up to about 2800 operations per second. We also show that TCC achieves better performance than causal consistency and requires a negligible overhead when compared with eventual consistency.