It’s too bad that there have been no studies on the costs of containers. At least none that I know of an none that are not biased in some way.
Currently I’m working on a reporting cluster and while the usual etcd cluster with arbitrary worker worker nodes is good enough I realized that if my domain were evenly slightly more complicated that events like SDN (software defined networks) might require a complete redeploy; and while that is ok for a reporting system that has wide maintenance windows a 24x7 webapp might not have the same outcome.
Even so, as an operations person I would hate to dump the existing deployment without having deployed and tested the new environment. While this overlap makes perfect sense when using AWS or GCE, if I wanted to accomplish the same in a private environment I would need 4x or 3x the hardware regardless of how the hardware was partitioned.
Consider this: If you were going to perform green/blue deploy strategies in production just how much of your infrastructure would you need to deploy?
UPDATE: something I need to be clear about is that there is a thin layer between the hardware network and the SDN above it. That same network has an analog in the baremetal deployment as well as the VM, vSphere, environments.