Skip to main content

MetalLB on Hetzner Dedicated with vSwitch

· 4 min read

When running Kubernetes on Hetzner Dedicated, there is no cloud load balancer. But you can provide public LoadBalancer IPs by attaching a routed IP range to a vSwitch and letting MetalLB announce addresses over L2.

Our setup:

  • Calico (VXLAN + WireGuard)
  • kube-proxy IPVS with strictARP
  • ingress-nginx for ingress traffic

The diagram below illustrates the traffic flow: MetalLB advertises a public VIP from one node at a time, ingress-nginx receives it, and traffic is forwarded to the application pod running anywhere in the cluster.

MetalLB + Hetzner vSwitch topology

1. Assign a public subnet to your vSwitch​

Example routed block Hetzner provides:

Subnet:     123.45.67.32/29
Gateway:    123.45.67.33
Usable:     123.45.67.34–38
Broadcast:  123.45.67.39

Attach your dedicated servers to the vSwitch (VLAN ID e.g. 4000).

2. Configure vSwitch VLAN on each node​

Each node gets a /32 from the subnet – Hetzner routes the whole /29 to your server.

Create /etc/netplan/10-vlan-4000.yaml:

network:
  version: 2
  renderer: networkd
  vlans:
    vlan4000:
      id: 4000
      link: eno1
      mtu: 1400
      addresses:
        - 123.45.67.38/32 # node-specific
      routes:
        - to: 0.0.0.0/0
          via: 123.45.67.33
          on-link: true
          table: 200
        - to: 123.45.67.32/29
          scope: link
          table: 200
      routing-policy:
        - from: 123.45.67.32/29
          table: 200
          priority: 10
        - to: 123.45.67.32/29
          table: 200
          priority: 10
        - from: 123.45.67.32/29
          to: 10.233.0.0/18
          table: 254
          priority: 0
        - from: 123.45.67.32/29
          to: 10.233.64.0/18
          table: 254
          priority: 0

Apply:

netplan apply

3. Required sysctl settings​

Create /etc/sysctl.d/999-metallb.conf:

net.ipv4.conf.all.arp_ignore=1
net.ipv4.conf.all.arp_announce=2
net.ipv4.conf.all.rp_filter=0
net.ipv4.conf.default.rp_filter=0

Why:

SettingPurpose
arp_ignore=1Only reply to ARP queries for an IP on the correct interface – prevents conflicting replies from Calico/VXLAN.
arp_announce=2Send ARP only from the interface that owns the VIP, required when MetalLB moves VIPs between nodes.
rp_filter=0Disable strict reverse-path filtering – otherwise nodes drop return traffic sourced from VIPs or remote pods.

4. kube-proxy + Calico adjustments​

Enable strictARP in kube-proxy (IPVS mode):

apiVersion: kubeproxy.config.k8s.io/v1alpha1
kind: KubeProxyConfiguration
ipvs:
  strictARP: true

MTU must account for VXLAN + vSwitch + WireGuard overhead:

Calico MTU: 1280 (consistent across nodes)

5. Deploy MetalLB​

apiVersion: metallb.io/v1beta1
kind: IPAddressPool
metadata:
  name: vswitch
  namespace: metallb-system
spec:
  addresses:
    - 123.45.67.34-123.45.67.36 # free VIPs
---
apiVersion: metallb.io/v1beta1
kind: L2Advertisement
metadata:
  name: l2
  namespace: metallb-system
spec:
  ipAddressPools: ["vswitch"]
  interfaces: ["vlan4000"]

Restart MetalLB speakers to pick up interface binding.

6. Ingress service configuration​

For ingress-nginx:

spec:
  externalTrafficPolicy: Local

Pro:

  • Preserves client IP
  • Prevents traffic hairpin across nodes

Tradeoff:

  • Only one node handles a given connection (acceptable for ingress)

7. Verification​

Confirm that your ingress-nginx Service received a public VIP:

kubectl get svc -n ingress-nginx ingress-nginx-controller

Expected example:

NAME                       TYPE           CLUSTER-IP      EXTERNAL-IP    PORT(S)                      AGE
ingress-nginx-controller   LoadBalancer   10.233.53.156   123.45.67.35   80:30440/TCP,443:30477/TCP   17d

Inspect the Service events to see which node currently advertises the VIP:

kubectl describe svc -n ingress-nginx ingress-nginx-controller

Look for:

Events:
  Normal  nodeAssigned  ...  metallb-speaker  announcing from node "control-plane-1" with protocol "layer2"

Then verify reachability from outside:

curl -I http://123.45.67.35

Failover test​

  1. Identify the active announcer from the above events
  2. Shut that node down abruptly:
sudo poweroff
  1. Re-run:
curl -I http://123.45.67.35

Expected: traffic continues within ~1–2 seconds as another node picks up the VIP.

➡️ Note: VIPs do not appear in ip addr on nodes; they are held in IPVS and advertised via ARP. That is normal.

Wrapping up​

This gives:

  • Public LoadBalancer IPs
  • Fast failover (~1-2s)
  • Clean separation: pod networking via VXLAN/WireGuard, external via vSwitch

Alternative approaches:

  • Hetzner Cloud Load Balancer (simpler, works with Dedicated too)
  • Cilium with L2 announcements

We run hosting infrastructure, so controlling ingress networking ourselves matters (mostly to prove the point really). Hetzner still runs the vSwitch underneath, but it's more independent than relying on the cloud LB.

And if you'd rather not handle any of this yourself – Hostim.dev is now live. You can deploy your Docker or Compose apps with built-in databases, volumes, HTTPS, and logs – all in one place, ready in minutes.