Asynchronous replication of eventually consistent updatable views

Abstract

Users of software applications expect fast response times and high availability. This is despite several applications moving from local devices and into the cloud. A cloud-based application that could function locally will now be unavailable if a network partition occurs. A fundamental challenge in distributed systems is maintaining the right tradeoffs between strong consistency, high availability, and tolerance to network partitions. The impossibility of achieving all three properties is described in the CAP theorem. To guarantee the highest degree of responsiveness and availability, applications could be run entirely locally on a device without directly relying on cloud services.

Software that can be run locally without a direct dependency on cloud services are called local-first software. Being local-first means that consistency guarantees may need to be relaxed. Weaker consistency, such as eventual consistency, can be used instead of strong consistency. Implementing conflict-free replicated data types (CRDTs) is a provably correct way to achieve eventual consistency. These data types guarantee that the state of different replicas will converge towards a common state when a system becomes connected and quiescent.

The drawback of using CRDTs is that they are unbounded in their growth. This means they can quickly become too large to handle using less capable devices like smartphones, tablets, or other edge devices. To mitigate this, partial replication can be implemented to replicate only the data each device needs. This comes with the added benefit of limiting the information users obtain, thus possibly improving security and privacy.

The main contribution of this thesis is a new approach to partial replication. It is based on an existing asynchronously replicated relational database to support local-first software and guarantees eventual consistency. The new approach uses database views to define partial replicas. The database views are made updatable by drawing inspiration from the large body of research on updatable views. We differentiate ourselves from earlier work on non-distributed updatable views by guaranteeing that the views are eventually consistent.

The approach is evaluated to ensure it can be used for real scenarios. The approach has proved to be usable in the scenarios. The replication of database views has also been experimentally tested to ensure that our approach to partial replication is viable for less capable devices.