Comment 7 for bug 1561894

According to 'list_policies' output HA is enabled only for ceilometer queues. And it is the way it should be. For RPC loss of rabbitmq broker will result in loss of all queue contents located on that broker. Queue is created on the node where some client issued 'queue.declare' for the first time - so its evident that one the the brokers you've killed contained those 10000 message.

Actually, even in pause_minority mode of clustering and with HA enabled there are no guarantees of no data loss (see https://aphyr.com/posts/315-jepsen-rabbitmq)