A Time Series-Based Approach to Elastic Kubernetes Scaling

With the increasing popularity of cloud-native architectures and containerized applications, Kubernetes has become a critical platform for managing these applications. However, Kubernetes still faces challenges when it comes to resource management. Specifically, the platform cannot achieve timely scaling of the resources of applications when their workloads fluctuate, leading to insufficient resource allocation and potential service disruptions. To address this challenge, this study proposes a predictive auto-scaling Kubernetes Operator based on time series forecasting algorithms, aiming to dynamically adjust the number of running instances in the cluster to optimize resource management. In this study, the Holt-Winter forecasting method and the Gated Recurrent Unit (GRU) neural network, two robust time series forecasting algorithms, are employed and dynamically managed. To evaluate the effectiveness, we collected workload metrics from a deployed RESTful HTTP application, implemented predictive auto-scaling, and assessed the differences in service quality before and after the implementation. The experimental results demonstrate that the predictive auto-scaling component can accurately predict the future trend of the metrics and intelligently scale resources based on the prediction results, with a Mean Squared Error (MSE) of 0.00166. Compared to the deployment using a single algorithm, the cold start time is reduced by 1 h and 41 min, and the fluctuation in service quality is reduced by 83.3%. This process effectively enhances the quality of service and offers a novel solution for resource management in Kubernetes clusters.