The BGP speaker currently uses the Kubernetes node IP as the next-hop for all traffic. This works fine on clusters with single interfaces/IPs, but falls apart in clusters where machines have several NICs and IPs.
A more flexible solution would be to use the BGP session's source IP as the nexthop, and stop getting IPs via the k8s Downward API. That way, the kernel's source IP selection will help us make the correct choice in all cases.
The BGP speaker currently uses the Kubernetes node IP as the next-hop for all traffic. This works fine on clusters with single interfaces/IPs, but falls apart in clusters where machines have several NICs and IPs.
A more flexible solution would be to use the BGP session's source IP as the nexthop, and stop getting IPs via the k8s Downward API. That way, the kernel's source IP selection will help us make the correct choice in all cases.