I believe this NPE can be fixed easily, but I'm not sure that will address the root cause here. Multiple connections for single user session are currently expected to terminate in the same host, rather than switching among members within the cluster. We are planning to address this limitation at some point (for example, see OF-689), but it is not supported at this time.
As a workaround, we recommend that you configure your load balancer with a persistence rule that will select one of the cluster members based on the client source IP address (or subnet) in the inbound request. In the event that the selected member goes offline, the LB should be able to re-balance the client session using one of the remaining active nodes.