Dependability Assessment of NoSQL Engines

Abstract

NoSQL databases are the response to the sheer volume of data being generated, stored, and analyzed by modern users and applications. They see great usage in big data and real-time web applications as they are capable of both: extreme horizontal scaling, and storing data with less rigid structures than the ones used in relational DBMSes. NoSQL databases are known to sometimes compromise consistency in favor of availability, partition tolerance, and performance. Several studies have evaluated the performance of these databases, but their users also need to understand how they behave in the presence of faults and how to quantify the impact of those faults. As such, the main goal of this dissertation is to perform an experimental assessment of NoSQL databases. This work proposes an experimental methodology based on fault injection, an experimental procedure to implement the methodology, and three metrics based on dependability attributes, which can be used to evaluate the databases. A prototype for a tool, which implements the experimental procedure and automates the testing process, developed during the work is also presented. Lastly, an experimental campaign assessing three popular NoSQL engines was performed to demonstrate both the methodology and tool. The results of the experimental campaign show that many times, even in the presence of simple faults, the integrity of the data is a ected. They also show that each database handles the workloads and faults di erently, evidencing that users need to carefully select which solution to use in their systems.