Apollo: Learning Query Correlations for Predictive Caching in Geo-Distributed Systems

The performance of modern geo-distributed database applications is increasingly dependent on remote access latencies. Systems that cache query results to bring data closer to clients are gaining popularity, but they do not dynamically learn and exploit access patterns in client workloads. We present a novel prediction framework that identifies and makes use of workload characteristics obtained from data access patterns to exploit query relationships within an application’s database workload. We have designed and implemented this framework as Apollo, a system that learns query patterns and adaptively uses them to predict future queries and cache their results. Through extensive experimentation with two different benchmarks, we show that Apollo provides significant performance gains over popular caching solutions through reduced query response time. Our experiments demonstrate Apollo’s robustness and scalability as a predictive cache for geo-distributed database applications.