Time Series Forecast on Kubernetes Metrics

This study focuses on time series forecasting of central processing unit (CPU) usage from metric values collected via Prometheus from Kubernetes clusters. Kubernetes is a widely used platform for container orchestration, and CPU usage is a critical indicator of system performance and health. In this study, time series forecasting models were tested on CPU usage data collected from a ten-node Kubernetes cluster. CPU usage percentage predictions were made using statistical models such as ARIMA, Seasonal Naive, as well as the DeepAR deep learning model and the pre-trained large-scale time series forecasting model Chronos-Bolt. Model performances were evaluated with Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Mean Squared Error (MSE) metrics. While DeepAR and Chronos-Bolt models generally performed better, differences were observed in capturing instantaneous changes. While the ARIMA model yielded successful results in some nodes, the Seasonal Naive model was successful in capturing instantaneous changes despite its low prediction accuracy. This study provides guidance in selecting the most suitable model when estimating CPU usage of Kubernetes environments.