Model to estimate the CPU usage for a Pod to be scheduled

[knarayan]

Considering the historical data about container images, resource requests, and their corresponding average resource utilizations, it may be possible to train a ML model to predict the resource utilization for a pod to be scheduled. E.g., if five pods are already running and Horizontal Pod Autoscaler decides to create a sixth pod, then the average CPU utilization across the existing five pods can be considered as the average CPU utilization for the sixth pod.

[knarayan]

To begin with, we predict utilization for a container based on its requests/limits similar to the way it is done in the k8s scheduler-plugin Trimaran https://github.com/kubernetes-sigs/scheduler-plugins/blob/master/pkg/trimaran/targetloadpacking/targetloadpacking.go#L197