NoneDriver: Minikube start restarts containerd, hence it kills all my non k8s containers

[marcosdiez]

I am using version minikube 1.1.1

Whenever I do a

sudo -E minikube start --vm-driver=none --extra-config kubeadm.ignore-preflight-errors=SystemVerification ${EXTRA_DNS_PARAMETER} --logtostderr

I see this:

I0620 13:33:02.265561   22899 exec_runner.go:39] Run: systemctl is-active --quiet service containerd
I0620 13:33:02.279081   22899 exec_runner.go:39] Run: sudo systemctl stop containerd
I0620 13:33:02.361533   22899 exec_runner.go:39] Run: systemctl is-active --quiet service containerd
I0620 13:33:02.381293   22899 exec_runner.go:39] Run: systemctl is-active --quiet service crio
I0620 13:33:02.396272   22899 exec_runner.go:39] Run: sudo systemctl start docker

which I traced to

https://github.com/kubernetes/minikube/blob/master/pkg/minikube/cruntime/docker.go#L78

// Enable idempotently enables Docker on a host
func (r *Docker) Enable() error {
	if err := disableOthers(r, r.Runner); err != nil {
		glog.Warningf("disableOthers: %v", err)
	}
	return r.Runner.Run("sudo systemctl start docker")
}

disableOthers is here by the way: https://github.com/kubernetes/minikube/blob/master/pkg/minikube/cruntime/cruntime.go#L96

What's the rationale for that ?
Why does minikube bother messing with my container daemons in first place ?

Should we just remove disableOthers ? Should we remove it only for the none driver ?
Should we just not restart containerd if docker is > SOME_VERSION ?

Should we just not restart containerd because today everybody uses it, hence it shouldn't just be restarted ?

All the solutions are trivial to implement. Hence I am opening the discussion so I can create the right PR :)

[afbjorklund]

The Minikube ISO starts all three container runtimes (especially Docker, which is mandatory) at boot
So therefore the start script will stop the others again, in order to not use as many resources.

When you give Minikube your other dedicated VM, it doesn't expect there to be any other daemons. But we could avoid "disableOthers" for the none driver - similar to how we e.g. avoid the caching ?

---

Switching to containerd as the default is definitely something we discuss, but today people are using Minikube both for docker (via docker-env) and for kubectl. But we could start it on-demand ?

If the current restart code could be improved, that would be interesting. I know there were some issues when stopping the dockerd, and we need to look further into how dockerd and containerd play together.

[marcosdiez]

Now it makes sense. So I'll just send a PR that will not run disableOthers when the driver is none.
Thank you @afbjorklund !

[afbjorklund]

Also, we only recently started running containerd - before that it was docker-containerd

Right now the code still expects each container runtime to be separate (docker/rkt/containerd/crio/etc)
Since that is not really true anymore (and probably wasn't for docker 18.06 either), it needs to improve

https://docs.docker.com/engine/release-notes/#18090