When we are scheduling a gateway on the available chassis, it can happen that only one node is available at that time (because the rest of then network nodes may be down for whatever reason). Therefore, the router will be scheduled on it becoming the active chassis for its gateway.
When the rest of the nodes come up back again in the cluster, no rescheduling happens and the previous router is not effectively in HA. If the active node goes down, North-South traffic to that router will be disrupted as it won't be failed over to any of the network nodes available.
The solution would be to schedule all routers which have been scheduled in less than MAX_GW_CHASSIS [0] chassis in the new nodes (perhaps with a lower priority to avoid failovers although it may cause some unbalance).
[0] http://git.openstack.org/cgit/openstack/networking-ovn/tree/networking_ovn/l3/l3_ovn_scheduler.py?id=d40470a51314fc0c60353c9882e0d2d44c9d2aa5#n31
@dalvarez did you already handle this, or was it something else?