How We Learned to Stop Worrying and Read from Replicas

A standard pattern in scaling relational data access is to use replicas for reads, but this compromises read-after-write consistency. Unfortunately, since many use-cases require stronger consistency guarantees, the pattern is of limited utility. At Box much of our relational data access requires strong consistency for the purposes of enforcing enterprise permissions, but we need to scale too! To address this challenge, we have introduced a novel design pattern and a framework implementing it. This framework, currently used at scale to serve production traffic at Box, utilizes real-time replication position analysis to safely direct traffic that requires strong read-after-write consistency to read-only replicas.