You do it roughly based on the deployment requests and the average HPA values throughout a time slice. Most k8s workloads run on a homogenous set of node types

You do it roughly based on the deployment requests and the average HPA values throughout a time slice. Most k8s workloads run on a homogenous set of node types, so you can have an hourly cost per gb and per vcpu without digging too much into it.