John Wilkes of Google talks about this problem with Jeff Meyerson [1] and how it relates to the choice to use or not use containers. The spoiler is that container management tooling allows separation of infrastructure builds from deployment: a configuration problem when building a container happens on the build server instead of while a script is running on machine in production. His argument is that when a container deployment to production fails, the state of the machine is readily known (new bad container) versus an more complex state when a scripted build fails part way to completion.

And a container management tool can facilitate handling a failed distribution automatically via rollback to a previously deployed working container.

http://www.se-radio.net/2016/01/se-radio-show-246-john-wilke...