Using Redundant Reads for Improving Performance and Availability of Distributed Replicated Systems with Relaxed Consistency

Distributed NoSQL databases is a new time of data storages which offer configurable levels of consistency so that data can scale across many geographically distributed nodes. In order to achieve high system availability and allow a quick response the architects of modern large-scale web applications such as Facebook, Twitter, etc. often decide to weak consistency requirements. In this paper, we put forward the idea of using redundant read requests to further reduce response time and improve availability of replicated distributed systems operating at a relaxed consistency level. The proposed approach was implemented on a testbed Cassandra NoSQL cluster. Our evaluation results show that redundant reads can be considered as an effective means of reducing the probability of extreme delays that regularly occur in distributed systems. In some scenarios, the proposed mechanism can not only improve system availability and minimize the worst-case execution time, but also reduce the average response time despite the increase in system workload.