Considering one ML application is deployed using Kubernetes, its output depends on the data which is constantly stored in the model, if needing to scale the system based on available CPUs, what feature should be enabled?