Trade-offs in Replicated Systems

Replicated systems provide the foundation for most of today’s large-scale services. Engineering such replicated system is an onerous task. The first—and often foremost—step in this task is to establish an appropriate set of design goals, such as availability or performance, which should synthesize all the underlying system properties. Mixing design goals, however, is fraught with dangers, given that many properties are antagonistic and fundamental trade-offs exist among them. Navigating the harsh landscape of trade-offs is difficult because these formulations use different notations and system models, so it is hard to get an all-encompassing understanding of the state of the art in this area. In this paper, we address this difficulty by providing a systematic overview of the most relevant tradeoffs involved in building replicated systems. Starting from the well-known FLP result, we follow a long line of research and investigate different trade-offs, assembling a coherent perspective of these results. Among others, we consider trade-offs which examine the complex interactions between properties such as consistency, availability, low latency, partition-tolerance, churn, scalability, and visibility latency.