Highly Available Queues

If your RabbitMQ broker consists of a single node, then a failure of that node will cause downtime, temporary unavailability of service, and potentially loss of messages (especially non-persistent messages held by non-durable queues). You could publish all messages persistent, to durable queues, but even then, due to buffering there is an amount of time between the message being sent and the message being written to disk and fsync'd. Using publisher confirms is one means to ensure the client understands which messages have been written to disk, but even so, you may not wish to suffer the downtime and inconvenience of the unavailability of service caused by a node failure, or the performance degradation of having to write every message to disk.

You could use a cluster of RabbitMQ nodes to construct your RabbitMQ broker. This will be resilient to the loss of individual nodes in terms of the overall availability of service, but some important caveats apply: whilst exchanges and bindings survive the loss of individual nodes, queues and their messages do not. This is because a queue and its contents reside on exactly one node, thus the loss of a node will render its queues unavailable.

You could use an active/passive pair of nodes such that should one node fail, the passive node will be able to come up and take over from the failed node. This can even be combined with clustering. Whilst this approach ensures that failures are quickly detected and recovered from, there can be reasons why the passive node can take a long time to start up, or potentially even fail to start. This can cause at best, temporary unavailability of queues which were located on the failed node.

To solve these various problems, we have developed active/active high availability for queues. This works by allowing queues to be mirrored on other nodes within a RabbitMQ cluster. The result is that should one node of a cluster fail, the queue can automatically switch to one of the mirrors and continue to operate, with no unavailability of service. This solution still requires a RabbitMQ cluster, which means that it will not cope seamlessly with network partitions within the cluster and, for that reason, is not recommended for use across a WAN (though of course, clients can still connect from as near and as far as needed).

Mirrored Queue Behaviour

In normal operation, for each mirrored-queue, there is one master and several slaves, each on a different node. The slaves apply the operations that occur to the master in exactly the same order as the master and thus maintain the same state. All actions other than publishes go only to the master, and the master then broadcasts the effect of the actions to the slaves. Thus clients consuming from a mirrored queue are in fact consuming from the master.

Should a slave fail, there is little to be done other than some bookkeeping: the master remains the master and no client need take any action or be informed of the failure.

If the master fails, then one of the slaves must be promoted. At this point, the following happens:

  1. A slave is promoted to become the new master. The slave chosen for promotion is the eldest slave. As such, it has the best chance of being synchronised with the master. However, note that should there be no slave that is synchronised with the master, messages that only the master held will be lost.
  2. The slave considers all previous consumers to have been abruptly disconnected. As such, it requeues (at the rear of the queue) all messages that have been delivered to clients but are pending acknowledgement. This can include messages for which a client has issued acknowledgements: either the acknowledgement was lost on the wire before reaching the master, or it was lost during broadcast from the master to the slaves. In either case, the new master has no choice but to requeue all messages it thinks have not been acknowledged.
  3. Clients that were consuming from the mirrored-queue and support our Consumer Cancellation Notifications extension will receive a notification that their subscription to the mirrored-queue has been abruptly cancelled. At this point they should re-consume from the queue, which will pick up the new master. The reason for sending this notification is that informing clients of the loss of the master is essential: otherwise the client may continue to issue acknowledgements for messages they were sent by the old, failed master, and not expect that they might be about to see the same messages again, this time sent by the new master. Of course, clients that were connected to the failed node will find their connections failed, and will need to reconnect to a surviving node of the cluster.
  4. As a result of the requeuing, clients that re-consume from the queue must be aware that they are likely to subsequently receive messages that they have seen previously.

As the chosen slave becomes the master, no messages that are published to the mirrored-queue during this time will be lost: messages published to a mirrored-queue are always published directly to the master and all slaves. Thus should the master fail, the messages continue to be sent to the slaves and will be added to the queue once the promotion of a slave to the master completes.

Similarly, messages published by clients using publisher confirms will still be confirmed correctly even if the master (or any slaves) fail between the message being published and the message being able to be confirmed to the publisher. Thus from the point of view of the publisher, publishing to a mirrored-queue is no different from publishing to any other sort of queue. It is only consumers that need to be aware of the possibility of needing to re-consume from a mirrored-queue upon receipt of a Consumer Cancellation Notification.

If you are consuming from a mirrored-queue with noAck=true (i.e. the client is not sending message acknowledgements) then messages can be lost. This is no different from the norm of course: the broker considers a message acknowledged as soon as it has been sent to a noAck=true consumer, and should the client disconnect abruptly, the message may never be received. In the case of a mirrored-queue, should the master die, messages that are in-flight on their way to noAck=true consumers may never be received by those clients, and will not be requeued by the new master. Because of the possibility the the consuming client is connected to a node that survives, the Consumer Cancellation Notification is useful in identifying when such events may have occurred. Of course, in practise, if you care about not losing messages then you are advised to consume with noAck=false.

Publisher Confirms and Transactions

Mirrored queues support both Publisher Confirms and Transactions. The semantics chosen are that in the case of both confirms and transactions, the action spans all mirrors of the queue. So in the case of a transaction, a tx.commit-ok will only be returned to a client when the transaction has been applied across all mirrors of the queue. Equally, in the case of publisher confirms, a message will only be confirmed to the publisher when it has been accepted by all of the mirrors. It is correct to think of the semantics as being the same as a message being routed to multiple normal queues, and of a transaction with publications within that similarly are routed to multiple queues.

Unsynchronised Slaves

A node may join a cluster at any time. Depending on the configuration of a queue, when a node joins a cluster, queues may add a slave on the new node. At this point, the new slave will be empty: it will not contain any existing contents of the queue, and currently, there is no synchronisation protocol. Such a slave will receive new messages published to the queue, and thus over time will accurately represent the tail of the mirrored-queue. As messages are drained from the mirrored-queue, the size of the head of the queue for which the new slave is missing messages, will shrink until eventually the slave's contents precisely match the master's contents. At this point, the slave can be considered fully synchronised, but it is important to note that this has occured because of actions of clients in terms of draining the pre-existing head of the queue.

Thus a newly added slave provides no additional form of redundancy or availability of the queue's contents until the contents of the queue that existed before the slave was added have been removed. As a result of this, it is preferable to bring up all nodes on which slaves will exist prior to creating mirrored queues, or even better to ensure that your use of messaging generally results in very short or empty queues that rapidly drain.

You can determine which slaves have synchronised with the following rabbitmqctl invocation, or through the management plugin:

rabbitmqctl list_queues name slave_pids synchronised_slave_pids

Starting and Stopping Nodes

If you stop a RabbitMQ node which contains the master of a mirrored-queue, some slave on some other node will be promoted to the master (assuming there is one). If you continue to stop nodes then you will reach a point where a mirrored-queue has no more slaves: it exists only on one node, which is now its master. If the mirrored-queue was declared durable then, if its last remaining node is shutdown, durable messages in the queue will survive the restart of that node. In general, as you restart other nodes, if they were previously part of a mirrored-queue then they will rejoin the mirrored queue.

However, there is currently no way for a slave to know whether or not its queue contents have diverged from the master to which it is rejoining (this could happen during a network partition, for example). As such, when a slave rejoins a mirrored-queue, it throws away any durable local contents it already has and starts empty. Its behaviour is at this point the same as if it were a new node joining the cluster.

Creating a Mirrored Queue

A mirrored-queue must be created as a mirrored-queue. It is not possible to convert a non-mirrored-queue to a mirrored-queue at some later point. It is perfectly acceptable to create a mirrored-queue with no slaves initially, though be aware of the behaviour of adding nodes to a cluster.

To create a mirrored-queue, you provide an x-ha-policy entry in the argument table presented to queue.declare. The value of this entry is a long string which gives the name of the policy you wish to use for this queue. There are currently two policies available:

Some examples:

Map<String, Object> args = new HashMap<String, Object>();
args.put("x-ha-policy", "all");
channel.queueDeclare("myqueue", false, false, false, args);
This declares a queue myqueue which has a mirror on every node that is in the cluster or joins the cluster.
Map<String, Object> args = new HashMap<String, Object>();
args.put("x-ha-policy", "nodes");
args.put("x-ha-policy-params", Arrays.asList("node1@rabbit", "node2@rabbit", "node4@rabbit"));
channel.queueDeclare("myqueue", false, false, false, args);
This declares a queue myqueue which has a mirror on each of the nodes node1@rabbit, node2@rabbit and node4@rabbit.