Predictive Auto-scaler for Kubernetes Cloud

Abstract

One of the basic requirements with adapting to cloud technology is to find an optimal resource allocation based on the dynamic workload. The default functioning of Kubernetes Horizontal Pod Auto-scaling in cloud is scaling of its pods only when the threshold of the cluster/application is crossed in order to adapt to increasing workload. Rather we want to deploy a proactive provisioning framework based on machine learning based predictions. We have demonstrated a novel deep learning framework based on a transformer in the area of dynamic workload predictions and showed how to apply the results to a custom auto-scaler in cloud. Our Framework builds time-series predictive models in machine learning such as ARIMA, LSTM, Bi-LSTM and transformer models. The dynamic scaling framework applies machine learning algorithms and presents recommendations to make proactive and smart decisions. Though the transformer model has been used in NLP and Vision applications mostly, we showed that the transformer based model can produce the most effective results in cloud workload predictions.