Comment 3 for bug 1538539

This is a regression from https://review.openstack.org/#/c/252852
The main problem though is DVR scheduling mechanism which is a big mess (will be fixed in Mitaka!)

So after rescheduling server tries to notify each agent that router is scheduled to and verifies that agent has received that notification - in case of failure scheduling is considered as failed and no more agents are notified (if there are many for dvr case).
For dvr routers after rescheduling snat from down agent router may still be scheduled to down agent as there are dvr serviceable ports on that node. In this case server still tries to notify dead agent, fails and gives up. The fix will be to not try to notify down agents.