Load prediction-based auto-scaling in a cloud-native environment

Abstract

To optimally benefit from the elasticity and scalability of the cloud, applications need to be able to adapt to workload changes by provisioning and de-provisioning resources in an autonomic manner. This process is usually referred to as auto-scaling. A lot of research has already been conducted in the field of auto-scaling. However, auto-scaling in a cloud-native environment is still an open issue about which only very little literature exists.

The first part of this thesis provides an extensive overview of the available scientific literature in the field of auto-scaling. It compares the state-of-the-art techniques and discusses their major shortcomings. The second part of this thesis reports on a series of conducted experiments in which the benefits of using machine learning-based prediction models in auto-scaling are being validated in a cloud-native environment.

The conclusion of this thesis is that the cloud-native paradigm increases the burden of auto-scaling due to the large number of critical components and the high dimensional interdependencies of microservices. As a result, there are still numerous challenges that need to be solved to lower the complexity and improve the efficiency of auto-scaling in this context. Machine learning-based prediction techniques show a lot of potential to solve several of these challenges. Especially online machine learning models have a few interesting properties that set them apart from many other techniques.