For that matter, you could just take away the "server" abstraction altogether and run isolated processes in production instead, each having an ephemeral filesystem in the form of a container.

SSH access would be more like "heroku run bash" where you don't log into an existing instance, you log into a fresh ephemeral instance that just runs sshd and gets destroyed on logout.

Debugging running instances is more interesting though, and I think is possible if each container has an SSH daemon with ephemeral keys, and a central authority they register with that you can hop through to get shell on any instance.

Ops is a really fun job to be in when you stop thinking of servers as "things that people log into".