metallb: metallb sending wrong arp when using multiple ethernet interfaces

When having multiple (2) ethernet interfaces, metallb Version 0.9.3 sometimes uses the wrong interface/MAC-address for the arp-tables.

Example:

Client outside the kubernetes cluster:

Every 2.0s: arp -a | grep '94:XX:28:XX:26:70'
? (10.139.0.149) at 94:XX:28:XX:26:70 [ether] on ens160
? (192.168.73.92) at 94:XX:28:XX:26:70 [ether] on ens192
? (10.139.0.130) at 94:XX:28:XX:26:70 [ether] on ens160

Only the second line is the correct one here, while 192.168.xxx is the address of the kubernetes-node. The first and the third line are metallb-interfaces, which belong to the wrong MAC-address here (you see, that this is another interface).

Deleting arp-caches, pods and re-deploy the services was not a solution.

As a workaround, I downgraded metallb to version v0.8.1. After clearing the arp-Cache on the client again, my problem was solved:

# arp -a | grep 94:XX:28:XX:26:70
? (192.168.73.92) at 94:XX:28:XX:26:70 [ether] on ens192

# arp -a | grep 149
? (10.139.0.149) at f4:XX:43:4e:XX:64 [ether] on ens160

• MetalLB 0.9.3
• Kubernetes Version 1.17.1
• NetworkAddon flannel with edited config:

  net-conf.json: |
    {
      "Network": "10.244.0.0/16",
      "Backend": {
        "Type": "host-gw"
      }
    }

• Whether you’ve configured kube-proxy for iptables or ipvs mode

I think this is kube-proxy - the cluster was initiated by kubeadm, so I did not do anything like that manually.