Craig Weber

I've recently been working on my Raspberry Pi Kubernetes cluster. I also use Tailscale for my home VPN (because it's performant and absurdly easy to setup and configure). I wanted to run Kubernetes services on my VPN using private DNS names (e.g., foo.local) and addresses from the Tailscale address space (e.g., 100.*) as opposed to the host network address space (e.g., 192.168.*).

Since I'm using Klipper LB, the default load balancer controller for k3s, services' external IP addresses are always the IP addresses of a node in the cluster, and since each node is running on the Tailscale network, configuring DNS is pretty straightforward: get the external IP address of the service (in the 192.168.* address space), figure out what host that IP address corresponds to, and then figure out the Tailscale IP address (from the 100.* address space) and create the A record mapping a DNS name to that Tailscale IP address.

However, because these external service IP addresses change from time to time and to satiate my compulsion to automate everything, I'm running a DNS controller--a Kubernetes service that watches the Kubernetes cluster for changes to Ingresses and Services and CRUDs DNS records accordingly. So when a service (or ingress) is added to the cluster with an external IP address, the controller will fetch the DNS name for the service from an attribute and create an A record mapping that DNS name to the service's external IP address.

Since that IP address is in the host network IP space (192.168.*), the DNS won't resolve on hosts that are inside of the VPN but outside of the physical network (e.g., my cell phone is connected to the VPN over its 4G network). So either the controller needs to be Tailscale-aware (determine the host for the 192.168.* IP address and then determine the Tailscale IP address for that host) or Klipper needs to assign services external IP addresses from the 100.* address space instead of the 192.168.* address space.

Solution: Configure K3s to use the Tailscale network

The latter seemed easier and generally keeps my DNS controller decoupled from the details of my VPN and Load Balancer controller, so I went that route. I use k3sup to install k3s on my nodes, so my invocation install and join (for installing k3s on the master and worker nodes, respectively) looks like this:

# Install k3s; merge the kubeconfig with the user's home kubeconfig instead of
# sporadically dropping kubeconfig files all over.
$ k3sup install \
    --ip "$HOST" \
    --user "$USER" \
    --ssh-key "$PRIVKEY" \
    --k3s-extra-args "--flannel-iface tailscale0 \
        --advertise-address $TAILSCALE_ADDR \
        --node-ip $TAILSCALE_ADDR \
        --node-external-ip $TAILSCALE_ADDR" \
    --k3s-channel latest \
    --context rpis \
    --local-path $HOME/.kube/config \
    --merge

And

$ k3sup join \
    --ip "$HOST" \
    --server-ip "$MASTER_NODE" \
    --user "$USER" \
    --ssh-key "$PRIVKEY" \
    --k3s-extra-args "--flannel-iface tailscale0 \
        --node-ip $TAILSCALE_ADDR \
        --node-external-ip $TAILSCALE_ADDR" \
    --k3s-channel latest

For the purposes of this article, the noteworthy bits are the --k3s-extra-args. This string is plumbed through to the k3s server and k3s agent invocations in the resulting systemd units, and they tell k3s to use the Tailscale network interface and IP addresses. Effectively, as far as k3s is concerned, these nodes are all connected via Tailscale rather than the host network, so in theory (i.e., I haven't confirmed this yet) I should be able to add a node in the same way even if it's not on the same host network (e.g., an EC2 instance running in an AWS VPC).

Alternative Solution: Bridge Host and Tailscale Networks

Another solution would have been to bridge the host network with the Tailscale network such that a device running on a different physical network (but still on the VPN) could resolve addresses in the 192.168.* address space (and vice versa)--Tailscale has documentation for that here.

I like that a little less because it requires a different Tailscale configuration for one node (that node must be configured to advertise 100.* addresses on the physical network and 192.168.* addresses on the Tailscale network, as well as configuring subnets and disabling key rotation. Further, traffic that needs to cross this boundary must be routed through a particular node, adding a runtime cost (no real idea if the runtime cost exceeds Tailscale overhead). Mostly, it just adds a fair amount of complexity at the networking level, where I'm least comfortable debugging should something go wrong.