openrc: kubeadm is starting kubelet before complete config is writte to disk

[neolit123]

reported by @kvaster here:
kubernetes/kubernetes#86538

What happened:
kubeadm is timing out during cluster initialization on openrc systems.
Kubelet is failing to start cause there is no complete configuration on the disk.

What you expected to happen:
It is expected to have complete config already written to disk before attempt to start kubelet.

How to reproduce it (as minimally and precisely as possible):
Bug is reproducable with no configuration just by running kubeadm init

Anything else we need to know?:
Bug is not reproducable on systemd cause kubelet is constantly restarting until it can read all configuration files. And on openrc by default kubelet is not restarted once it fails with config. My proposition is to change order of configuration and launching:

diff --git a/cmd/kubeadm/app/cmd/init.go b/cmd/kubeadm/app/cmd/init.go
index 57e87d6603e..942b2057909 100644
--- a/cmd/kubeadm/app/cmd/init.go
+++ b/cmd/kubeadm/app/cmd/init.go
@@ -172,9 +172,9 @@ func NewCmdInit(out io.Writer, initOptions *initOptions) *cobra.Command {

        // initialize the workflow runner with the list of phases
        initRunner.AppendPhase(phases.NewPreflightPhase())
-       initRunner.AppendPhase(phases.NewKubeletStartPhase())
        initRunner.AppendPhase(phases.NewCertsPhase())
        initRunner.AppendPhase(phases.NewKubeConfigPhase())
+       initRunner.AppendPhase(phases.NewKubeletStartPhase())
        initRunner.AppendPhase(phases.NewControlPlanePhase())
        initRunner.AppendPhase(phases.NewEtcdPhase())
        initRunner.AppendPhase(phases.NewWaitControlPlanePhase())

Environment:
kubeadm version: &version.Info{Major:"1", Minor:"17", GitVersion:"v1.17.0", GitCommit:"70132b0f130acc0bed193d9ba59dd186f0e634cf", GitTreeState:"archive", BuildDate:"2019-12-18T09:14:17Z", GoVersion:"go1.13.5", Compiler:"gc", Platform:"linux/amd64"}

[neolit123]

@kvaster

My proposition is to change order of configuration and launching:

this is something that we discussed at some point, but the problem is that at this point it could be a breaking change to all the users that separated the init process into phases.

before applying this change we need an estimate of the implications.

And on openrc by default kubelet is not restarted once it fails with config. My proposition is to change order of configuration and launching:

is there a way to tell an openrc service to loop-restart (crash loop) the same way as systemd is doing?

/kind design
/area ecosystem

[kvaster]

is there a way to tell an openrc service to loop-restart (crash loop) the same way as systemd is doing?

There is no 'default' way, but there are several different ways. Using supervised for example.
I think it might make sense to add one more phase which will check and restart kubelet if it's not running.

[neolit123]

In any case the change must not be disruptive to systemd users as those compose the majority.

[kvaster]

I will try to propose some non-disruptive proposition.

[neolit123]

Ok, we can also discuss this in a meeting after the holidays. Let's not send the pr yet. Thanks.

[neolit123]

xref #1295