What is needed to create a solution that has no single point of failure?

The switch fabric that connects the cluster nodes with each other and the outside network should be as redundant as possible. A modular switch chassis is a good way to do this. We support using multiple NICs on Swarm nodes. This enables a dual-attached fabric for this exact purpose.  

Swarm nodes cannot be connected to multiple different subnets, though.