Message availability depends on both reachable brokers and surviving queue replicas. In a RabbitMQ active-active deployment, PCS can keep a broker service running on each cluster node while RabbitMQ handles message replication and queue-leader elections.
The systemd:rabbitmq-server resource controls the installed service, and a Pacemaker clone gives each node its own instance. It adopts an already formed RabbitMQ cluster with matching Erlang cookies and separate local data directories; cloning three unrelated brokers does not join them or copy their messages. Its service monitor also does not establish that a queue can accept or deliver messages.
A three-member quorum queue can remain available when one broker stops because the other two replicas retain a majority. Client applications still need multiple broker addresses or a health-checked load balancer, reconnect handling, publisher confirms, and consumer acknowledgements. A controlled node-standby test checks service withdrawal and message survival; power-loss fencing and application reconnection need their own tests.
Steps to set up RabbitMQ active-active with PCS:
- Confirm that all three Pacemaker nodes are online with quorum.
$ sudo pcs status Cluster name: clustername Cluster Summary: * Stack: corosync (Pacemaker is running) * Current DC: node-01 (version 3.0.1-3.0.1) - partition with quorum * Last updated: Mon Nov 9 15:10:11 2026 on node-01 ##### snipped ##### Node List: * Online: [ node-01 node-02 node-03 ] Full List of Resources: * fence-ipmi-node-01 (stonith:fence_ipmilan): Started node-02 * fence-ipmi-node-02 (stonith:fence_ipmilan): Started node-03 * fence-ipmi-node-03 (stonith:fence_ipmilan): Started node-01 ##### snipped #####
The nodes here are node-01, node-02, and node-03, with RabbitMQ already installed as rabbitmq-server.service. Cluster commands run on node-01 unless another node is named. A listed fence resource is not proof that power control works; tested fencing is a prerequisite.
Related: How to check Pacemaker cluster status
Related: How to create a STONITH resource in Pacemaker - Confirm that RabbitMQ reports the same three running cluster members.
$ sudo rabbitmqctl cluster_status Cluster status of node rabbit@node-01 ... ##### snipped ##### Disk Nodes rabbit@node-01 rabbit@node-02 rabbit@node-03 Running Nodes rabbit@node-01 rabbit@node-02 rabbit@node-03 ##### snipped #####
The remaining commands use PCS 0.12 clone syntax and a RabbitMQ 4.x quorum queue. The messaging test needs the management plugin reachable locally at 127.0.0.1:15672 on each broker, curl, Python 3, and a test account named ha-check with the management tag. Its permissions belong only to a dedicated ha-check virtual host, with configure/read access to pcs-check and write access to the default exchange. Existing application queues and credentials stay separate.
The handover changes who starts RabbitMQ after boot. A maintenance window allows recovery if cluster management interrupts broker connections. If the handover must be reversed, removing rabbitmq_service-clone through PCS stops its brokers; sudo systemctl enable --now rabbitmq-server.service on each node then restores systemd startup. RabbitMQ data directories must remain intact.
Related: How to delete a Pacemaker resource
Related: How to manage Pacemaker cluster services with systemctl in Linux - Disable independent RabbitMQ boot startup on every broker node.
$ sudo systemctl disable rabbitmq-server.service
The same command applies on all three nodes. Without --now it leaves the running brokers available for Pacemaker to discover. The service remains unmasked so Pacemaker can start it.
- Create the RabbitMQ service and its clone in one PCS operation.
$ sudo pcs resource create rabbitmq_service systemd:rabbitmq-server \ op start timeout=180s op stop timeout=120s \ op monitor interval=30s clone --future
The default clone placement allows one instance per cluster node. Creating the clone together with its service avoids an intermediate singleton resource that could stop the other brokers. The --future flag selects the PCS 0.12 syntax.
Related: How to create a Pacemaker resource
Related: How to create a cloned resource in Pacemaker
Related: How to set Pacemaker clone resource options - Check that the RabbitMQ clone is started on all three nodes.
$ sudo pcs status resources * Clone Set: rabbitmq_service-clone [rabbitmq_service]: * Started: [ node-01 node-02 node-03 ]A Starting, Stopped, or failed instance means the handover is incomplete. If startup is still in progress, the status can be checked again after it settles.
- Declare a temporary quorum queue through the local API on node-01.
$ curl --fail --silent --show-error --user ha-check \ -H 'content-type: application/json' -X PUT \ --data '{"durable":true,"arguments":{"x-queue-type":"quorum","x-expires":600000}}' \ http://127.0.0.1:15672/api/queues/ha-check/pcs-check Enter host password for user 'ha-check':Curl prompts for the account password rather than placing it in shell history. The request creates a three-replica queue while all three brokers are running and prints no response body on success. The unused queue expires after ten minutes. HTTP is confined to loopback here; remote management access needs HTTPS.
- Publish a persistent test message through node-01.
$ curl --fail --silent --show-error --user ha-check \ -H 'content-type: application/json' -X POST \ --data '{"properties":{"delivery_mode":2},"routing_key":"pcs-check","payload":"pcs-ha-message","payload_encoding":"string"}' \ http://127.0.0.1:15672/api/exchanges/ha-check/amq.default/publish Enter host password for user 'ha-check': {"routed":true}The routed value must be true. The HTTP API provides a small messaging check; it does not replace an AMQP application test.
Node standby evacuates every Pacemaker resource from node-01 and disconnects clients using that broker. The remaining two RabbitMQ nodes must stay online to preserve queue quorum. If the test fails, sudo pcs node unstandby node-01 restores the node's eligibility to run resources.
- Put node-01 into standby from node-02.
$ sudo pcs node standby node-01
This withdraws one broker under cluster control without disabling fencing or stopping the other brokers.
Related: How to run a Pacemaker failover test with PCS - Check that only node-01's RabbitMQ instance has stopped.
$ sudo pcs status resources * Clone Set: rabbitmq_service-clone [rabbitmq_service]: * Started: [ node-02 node-03 ] * Stopped: [ node-01 ]The messaging check belongs after this state appears; otherwise the first broker might still be serving the queue.
- Retrieve the retained message through node-02.
$ curl --fail --silent --show-error --user ha-check \ -H 'content-type: application/json' -X POST \ --data '{"count":1,"ackmode":"ack_requeue_true","encoding":"auto"}' \ http://127.0.0.1:15672/api/queues/ha-check/pcs-check/get | python3 -m json.tool Enter host password for user 'ha-check': [ { "payload_bytes": 14, "redelivered": false, ##### snipped ##### "payload": "pcs-ha-message", "payload_encoding": "string" } ]The returned pcs-ha-message shows that the queue remains usable after the first broker withdraws. A brief leader election can delay the response; [] or an HTTP error does not prove message survival. The ack_requeue_true mode puts this test message back for the final check.
- Return node-01 to service from node-02.
$ sudo pcs node unstandby node-01
- Confirm that Pacemaker has restarted the RabbitMQ clone on node-01.
$ sudo pcs status resources * Clone Set: rabbitmq_service-clone [rabbitmq_service]: * Started: [ node-01 node-02 node-03 ]The expected state is one started instance on each node, with no failed actions. If node-01 is still starting, it has not yet completed recovery.
- Consume the test message through node-03 after all brokers return.
$ curl --fail --silent --show-error --user ha-check \ -H 'content-type: application/json' -X POST \ --data '{"count":1,"ackmode":"ack_requeue_false","encoding":"auto"}' \ http://127.0.0.1:15672/api/queues/ha-check/pcs-check/get | python3 -m json.tool Enter host password for user 'ha-check': [ { "payload_bytes": 14, "redelivered": true, ##### snipped ##### "payload": "pcs-ha-message", "payload_encoding": "string" } ]The returned pcs-ha-message is readable through the third broker after recovery. This request acknowledges and removes only the test message; the clone remains running on all three nodes.