Autoscaling Preditivo de Microsserviços usando Kubernetes, Keda e Séries Temporais

Abstract

The widespread availability of computing devices has paved the way for the development of a new generation of containerized applications that can run in a distributed cloud environment. Additionally, the dynamic nature of the workload demands an elastic deployment of applications that can adapt to any scenario. One of the most popular existing container orchestration systems, Kubernetes, employs a scale strategy based on limits that can be application-dependent and challenging to modify. Furthermore, its vertical scaling approach is disruptive, limiting deployment availability. Scale decisions, rather than being proactive, are reactive in nature. In this work, our goal is to dynamically collect pod resource usage and predict future utilization over a period of time. We aim to use the maximum utilization within this time window for proactive scaling to enhance overall resource utilization. We also contrast Kubernetes' built-in threshold-based scaling policy with a reinforcement learning-based model and a Time Series-based prediction model, utilizing the Facebook framework: Prophet. We demonstrate the benefits of data-driven rules, which can be combined with event-driven automatic scaling, Keda, and the Kubernetes container orchestrator.