Efficient read monotonic data aggregation across shards on the cloud

Abstract

Client-centric consistency models define the view of the data storage expected by a client in relation to the operations done by a client within a session. Monotonic reads is a client-centric consistency model which ensures that if a process has seen a particular value for the object, any subsequent accesses will never return any previous values. Monotonic reads are used in several applications like news feeds and social networks to ensure that the user always has a forward moving view of the data.
The idea of Monotonic reads over multiple copies of the data and for lightly loaded systems is intuitive and easy to implement. For example, ensuring that a client session always fetches data from the same server automatically ensures that the user will never view old data.
However, such a simplistic setup will not work for large deployments on the cloud, where the data is sharded across multiple high availability setups and there are several million clients accessing data at the same time. In such a setup it becomes necessary to ensure that the data fetched from multiple shards are logically consistent with each other. The use of trivial implementations, like sticky sessions, causes severe performance degradation during peak loads.
This paper explores the challenges surrounding consistent monotonic reads over a sharded setup on the cloud and proposes an efficient architecture for the same. Performance of the proposed architecture is measured by implementing it on a cloud setup and measuring the response times for different shard counts. We show that the proposed solution scales with almost no change in performance as the number of shards increases.