Set provider ID for AWS nodes

[vixus0]

Sets the provider ID for controller and worker nodes on AWS to aws:///<az>/<instance id>.

• Use afterburn.service on Fedora and coreos-metadata.service on Flatcar to write instance metadata to file.
• Read instance metadata file in kubelet.service.
• Set --provider-id in kubelet arguments.

Testing

• Provisioned two clusters on AWS, one each of Fedora and Flatcar.
• Checked kubectl describe node had a correct value for ProviderID in the format aws:///<az>/<instance id> for both controllers and workers.

Rationale

Autoscaling can be pretty essential for many cluster operators.
Without a provider ID, it is impossible (as far as I know) to use tools such as cluster-autoscaler.

I understand this change isn't ideal as it makes use of the instance metadata service only available on AWS EC2 instances.
However, given there is an aws subdirectory, we already have some explicit notion of which cloud provider we're running against.

Alternative options

• Autoscale some other way, but nothing comes readily to mind.
• Set hostname in kubelet to the value of hostname --fqdn.
  This would maybe provide the option for running with an external cloud-controller-manager as it will match the node name to the EC2 instance and set providerID accordingly.
  But then we might need a way to set --cloud-provider=external as well, not sure.

[dghubble]

I think the relevant link from the cluster-autoscaler project is here with their needs. I'm open to setting provider-id for other unrelated reasons (on controllers too).

However, this should not need an additional systemd unit. Fedora CoreOS has afterburn and Flatcar Linux has a fork as well. At present, the only platforms that need afterburn enabled are DigitalOcan so you can see them for examples. There is some potential reliability concern since metadata can be a new flake point in bringup.

[vixus0]

Yes, the only bits missing from Typhoon that I could see were the following:

If you're managing your own kubelets, they need to be started with the --provider-id flag. The provider id has the format aws:////, e.g. aws:///us-east-1a/i-01234abcdef.

There's also the requirement for adding extra tags to the auto scaling group but that can be done with a local-exec provisioner.

I'll look into using afterburn and then I'll update the PR, thanks for the pointer.

[vixus0]

After spending some time faffing around with Fedora's recent SSH key issues I've managed to successfully test the changes using afterburn and coreos-metadata.

Testing:

• Provisioned two clusters on AWS, one each of Fedora and Flatcar.
• Checked kubectl describe node had a correct value for ProviderID in the format aws:///<az>/<instance id> for both controllers and workers.

[dghubble]

Rebased and merged as 507c646, thanks Anshul