Service that collects resource usage data from the Kubernetes cluster. This data is taken by the Iron-bank and stored in proper places.
The downtime is Bad, but not critical. Resource usage accounting may lose data during downtime. Generally not critical to stop for minute or two, but resource usage accuracy will be affected.
One pod per machine with flows. Should be implemented as daemon set in terms of the Kubernetes cluster.
Use rolling release. Update pods one by one.
Cadvisor strongly depends on the Kubernetes API service. It will not start without it.