Note that cAdvisor acts as a container resource utilization monitoring service, which is running inside kubelet on each node. The CPU utilizations we just monitored are collected by cAdvisor and aggregated by Heapster. Heapster is a service running in the cluster that monitors and aggregates the metrics. It queries the metrics from each cAdvisor. When HPA is deployed, the controller will keep observing the metrics which are reported by Heapster, and scale up and down accordingly. An illustration of the process is as follows: