Have you looked at NATS at all? We're using it as a message bus for one app and it's been fantastic. It is, however, an in-memory queue, and the current version cannot replace Rabbit for queues that require durability.
tbh, i never used clustering (because it's one of the shittiest clustering implementations i've ever seen) but we do use two servers (publishers connect to one randomly and consumers connect to both) and it seems to handle millions of messages without any issues.
of all servers i've ever used, rabbitmq is by far the most stable (together with ejabberd).
Right now, the main annoyance is that it's impossible, as far as I understand, to limit its memory usage. You can set a "VM high watermark" and some other things, but beyond that, it will — much like, say, Elasticsearch — use a large amount of mysterious memory that you have no control over. You can't just say "use 1GB and nothing more", which is problematic on Kubernetes where you want to pack things a bit tightly. This happens even if all the queues are marked as durable.
So it's awesome for realtime firehose-type use cases where a websocket client connects, receives messages (every client gets all the messages, although NATS also supports load-balanced fanout) for a while, then eventually disconnects.
NATS is ridiculously fast [1], too.
There's an add-on currently in beta, NATS Streaming [1], which [2] has durability, acking/redelivery and replay, so covers most of what you get from both RabbitMQ and Kafka. It looks very promising.
[1] http://bravenewgeek.com/tag/nats/
[2] https://nats.io/documentation/streaming/nats-streaming-intro...