Clustering, mirrored queues and quorum queues

Key points

• Cluster nodes share users, vhosts and definitions, but a classic queue lives on exactly one node.
• Nodes that do not host the queue proxy to the node that does, so queue placement affects latency.
• Classic mirroring used ha-mode policies, could lose confirmed messages on failover, and 4.0 removed it.
• Quorum queues arrived in RabbitMQ 3.8 and use Raft, so a write needs a majority of the replicas.
• Three replicas tolerate one node loss and five tolerate two, so always pick an odd replica count.
• Erlang clustering uses port 25672 and epmd on 4369, which are not the AMQP port.

Example

rabbitmq-diagnostics cluster_status

# quorum is a queue argument, chosen at declare time
rabbitmqadmin declare queue name=orders durable=true \
  arguments='{"x-queue-type":"quorum","x-delivery-limit":10}'

# who is the Raft leader, and are all replicas in sync?
rabbitmq-queues quorum_status orders

# add a replica on a new node, then rebalance leaders
rabbitmq-queues grow rabbit@node3 all
rabbitmq-queues rebalance quorum

Quorum queues for work that must not be lost, streams for replay, classic for cheap and disposable.