Workload-Aware Machine Learning for Microservice Scaling in Kubernetes

Abstract

Modern cloud platforms rely on Kubernetes as their main technology to manage microservices with fluctuating and unpredictable workloads, but current autoscaling solutions mainly respond to workload changes and find it challenging to sustain optimal service quality during sudden shifts in demand. The upsizing and downsizing of the resources to be done mostly depend on the threshold of resources and do not consider the actual workload thereby causing scaling, instability and wastage of resources which are not efficient. This paper addresses the gap in current autoscaling systems by examining their inability to forecast workload demands and proactively implement measures to prevent service performance issues. A key objective of this research is to create and test a machine learning framework that is aware of the work-load and therefore, the microservice scaling is going to be both proactive and stable in the Kubernetes environments. The proposed approach gathers various workload metrics, transforms them into structured features and ap-plies short-term prediction techniques to estimate future resource needs and potential service level issues. A policy-driven decision engine confirms these predictions and initiates scaling operations that are carried out through Kubernetes’s built-in mechanisms. A constant feedback loop makes it possible for the system to be in sync with the changing workload patterns over a period of time. The results of the experiment show that the proposed method is able to reduce the error in the prediction of the workload by more than forty percent and also improves the time taken for reaction to the scaling from almost ninety seconds to under thirty seconds. The service is made much more stable, as the scaling fluctuations are brought down by more than sixty