Toward Optimal Load Prediction and Customizable Autoscaling Scheme for Kubernetes

Most enterprise customers now choose to divide a large monolithic service into numbers of loosely-coupled, specialized microservices, which can be developed and deployed separately. Docker, as light-weight virtualization technology, has been widely adopted support diverse microservices. At the moment, Kubernetes is portable, extensible, open-source orchestration platform for managing these containerized microservice applications. To adapt frequently changing user requests, it offers an automated scaling method, Horizontal Pod Autoscaler (HPA), that scale itself based on system’s current workload. The native reactive auto-scaling however, unable foresee system workload scenario in future complete proactive scaling, leading QoS (quality service) violations, long tail latency, insufficient server resource usage. In this paper, we suggest new scheme deep learning approaches make up HPA’s inadequacies default autoscaler Kubernetes. After meticulous experimental evaluation comparative analysis, use Gated Recurrent Unit (GRU) model with higher prediction accuracy efficiency model, supplemented by stability window mechanism improve model. Finally, third-party custom autoscaling framework, Custom (CPA), packaged our algorithm framework real cluster. Comprehensive experiment results prove feasibility scheme, significantly outperforms existing (HPA) approach.