Skip to main content
Version: 3.13

Queues

What is a Queue?​

A queue in RabbitMQ is an ordered collection of messages. Messages are enqueued and dequeued (delivered to consumers) in a (FIFO ("first in, first out") manner.

To define a queue in generic terms, it is a sequential data structure with two primary operations: an item can be enqueued (added) at the tail and dequeued (consumed) from the head.

Queues play a major role in the messaging technology space. Many messaging protocols and tools assume that publishers and consumers communicate using a queue-like storage mechanism.

Many features in a messaging system are related to queues. Some RabbitMQ queue features such as priorities and requeueing by consumers can affect the ordering as observed by consumers.

The information in this topic includes an overview of queues in RabbitMQ and also links out to other topics so you can learn more about using queues in RabbitMQ.

This information primarily covers queues in the context of the AMQP 0-9-1 protocol, however, much of the content is applicable to other supported protocols.

Some protocols (for example: STOMP and MQTT) are based around the idea of topics. For these protocols, queues act as a data accumulation buffer for consumers. However, it is still important to understand the role queues play because many features still operate at the queue level, even for those protocols.

Streams is an alternative messaging data structure available in RabbitMQ. Streams provide different features from queues.

The information about RabbitMQ queues covered in this topic includes:

For topics related to consumers, see the Consumers guide. Classic queues, quorum queues and streams also have dedicated guides.

Queue Names​

Queues have names so that applications can reference them.

Applications may pick queue names or ask the broker to generate a name for them. Queue names may be up to 255 bytes of UTF-8 characters.

Queue names starting with "amq." are reserved for internal use by the broker. Attempts to declare a queue with a name that violates this rule will result in a channel-level exception with reply code 403 (ACCESS_REFUSED).

Server-named Queues​

In AMQP 0-9-1, the broker can generate a unique queue name on behalf of an app. To use this feature, pass an empty string as the queue name argument: the same generated name may be obtained by subsequent methods in the same channel by using the empty string where a queue name is expected. This works because the channel remembers the last server-generated queue name.

Server-named queues are meant to be used for state that is transient in nature and specific to a particular consumer (application instance). Applications can share such names in message metadata to let other applications respond to them (as demonstrated in tutorial six). Otherwise, the names of server-named queues should be known and used only by the declaring application instance. The instance should also set up appropriate bindings (routing) for the queue, so that publishers can use well-known exchanges instead of the server-generated queue name directly.

Queue Properties​

Queues have properties that define how they behave. There is a set of mandatory properties and a map of optional ones:

  • Name
  • Durable (the queue will survive a broker restart)
  • Exclusive (used by only one connection and the queue will be deleted when that connection closes)
  • Auto-delete (queue that has had at least one consumer is deleted when last consumer unsubscribes)
  • Arguments (optional; used by plugins and broker-specific features such as message TTL, queue length limit, etc)

Note that not all property combination make sense in practice. For example, auto-delete and exclusive queues should be server-named. Such queues are supposed to be used for client-specific or connection (session)-specific data.

When auto-delete or exclusive queues use well-known (static) names, in case of client disconnection and immediate reconnection there will be a natural race condition between RabbitMQ nodes that will delete such queues and recovering clients that will try to re-declare them. This can result in client-side connection recovery failure or exceptions, and create unnecessary confusion or affect application availability.

Declaration and Property Equivalence​

tip

Specifically for the queue type property, the property equivalence check can be relaxed. Alternatively, a default queue type (DQT) can be configured.

Before a queue can be used it has to be declared. Declaring a queue will cause it to be created if it does not already exist. The declaration will have no effect if the queue does already exist and its attributes are the same as those in the declaration. When the existing queue attributes are not the same as those in the declaration a channel-level exception with code 406 (PRECONDITION_FAILED) will be raised.

Specifically for the queue type property, the property equivalence checks can be relaxed or configured to use a default.

See the Virtual Hosts guide to learn more.

Optional Arguments​

Optional queue arguments, also known as "x-arguments" because of their field name in the AMQP 0-9-1 protocol, is a map (dictionary) of arbitrary key/value pairs that can be provided by clients when a queue is declared.

The map is used by various features and plugins such as

and so on.

The same idea is also used with other protocol operations, for example, when registering a consumer:

Some optional arguments are set at queue declaration time and remain immutable over the entire lifetime of the queue. Others can be dynamically changed after queue declaration via policies.

tip

For keys that can be set via policies, always first consider using a policy instead of setting these values in application code

Optional queue arguments can be set differently:

The former option is more flexible, non-intrusive, does not require application modifications and redeployments. Therefore it is highly recommended for most users. Note that some optional arguments such as queue type or max number of priorities can only be provided by clients because they cannot be dynamically changed and must be known at declaration time.

The way optional arguments are provided by clients varies from client library to client library but is usually an argument next to the durable, auto_delete and other arguments of the function (method) that declares queues.

Optional Arguments and Policy-Defined Key Precedence​

When the same key is provided by both client-provided x-arguments and by a policy, the former take precedence.

However, if an operator policy is also used, that will take precedence over the client-provided arguments, too. Operator policies are a protection mechanism and override client-provided values and user policy values.

For numerical values such as maximum queue length or TTL, the lower value of the two will be used. If an application needs or chooses to use a lower value, that will be allowed by an operator policy. A value higher than that defined in the operator policy, however, cannot be used.

Use operator policies to introduce guardrails for application-controlled parameters related to resource use (e.g. peak disk space usage).

Message Ordering in RabbitMQ​

Queues in RabbitMQ are ordered collections of messages. Messages are enqueued and dequeued (delivered to consumers) in the FIFO manner.

FIFO ordering is not guaranteed for priority and sharded queues.

Ordering also can be affected by the presence of multiple competing consumers, consumer priorities, message redeliveries. This applies to redeliveries of any kind: automatic after channel closure and negative consumer acknowledgements.

Applications can assume messages published on a single channel will be enqueued in publishing order in all the queues they get routed to. When publishing happens on multiple connections or channels, their sequences of messages will be routed concurrently and interleaved.

Consuming applications can assume that initial deliveries (those where the redelivered property is set to false) to a single consumer are performed in the same FIFO order as they were enqueued. For repeated deliveries (the redelivered property is set to true), original ordering can be affected by the timing of consumer acknowledgements and redeliveries, and thus not guaranteed.

In case of multiple consumers, messages will be dequeued for delivery in the FIFO order but actual delivery will happen to multiple consumers. If all of the consumers have equal priorities, they will be picked on a round-robin basis. Only consumers on channels that have not exceeded their prefetch value (the number of outstanding unacknowledged deliveries) will be considered.

Durability​

Queues can be durable or transient. Metadata of a durable queue is stored on disk, while metadata of a transient queue is stored in memory when possible. The same distinction is made for messages at publishing time in some protocols, e.g. AMQP 0-9-1 and MQTT.

In environments and use cases where durability is important, applications must use durable queues and make sure that publishers mark published messages as persisted.

Transient queues will be deleted on node boot. They therefore will not survive a node restart, by design. Messages in transient queues will also be discarded.

Durable queues will be recovered on node boot, including messages in them published as persistent. Messages published as transient will be discarded during recovery, even if they were stored in durable queues.

How to Choose​

In most other cases, durable queues are the recommended option. For replicated queues, the only reasonable option is to use durable queues.

Throughput and latency of a queue is not affected by whether a queue is durable or not in most cases. Only environments with very high queue or binding churn — that is, where queues are deleted and re-declared hundreds or more times a second — will see latency improvements for some operations, namely on bindings. The choice between durable and transient queues therefore comes down to the semantics of the use case.

Temporary queues can be a reasonable choice for workloads with transient clients, for example, temporary WebSocket connections in user interfaces, mobile applications and devices that are expected to go offline or use switch identities. Such clients usually have inherently transient state that should be replaced when the client reconnects.

Some queue types do not support transient queues. Quorum queues must be durable due to the assumptions and requirements of the underlying replication protocol, for example.

Temporary Queues​

With some workloads queues are supposed to be short lived. While clients can delete the queues they declare before disconnection, this is not always convenient. On top of that, client connections can fail, potentially leaving unused resources (queues) behind.

tip

Consider using server-generated names for temporary queues. Since such queues are not meant to be shared between N consumers, using unique names makes sense.

Shared temporary queues can lead to a natural race condition between RabbitMQ node actions and recovering clients.

There are three ways to make queue deleted automatically:

  • Exclusive queues (covered below)
  • TTLs (also covered below)
  • Auto-delete queues

An auto-delete queue will be deleted when its last consumer is cancelled (e.g. using the basic.cancel in AMQP 0-9-1) or gone (closed channel or connection, or lost TCP connection with the server).

If a queue never had any consumers, for instance, when all consumption happens using the basic.get method (the "pull" API), it won't be automatically deleted. For such cases, use exclusive queues or queue TTL.

Exclusive Queues​

An exclusive queue can only be used (consumed from, purged, deleted, etc) by its declaring connection.

tip

Consider using server-generated names for exclusive queues. Since such queues cannot be shared between N consumers, using server-generated names makes most sense.

An attempt to use an exclusive queue from a different connection will result in a channel-level exception RESOURCE_LOCKED with an error message that says cannot obtain exclusive access to locked queue.

Exclusive queues are deleted when their declaring connection is closed or gone (e.g. due to underlying TCP connection loss). They therefore are only suitable for client-specific transient state.

It is common to make exclusive queues server-named.

Exclusive queues are declared on the "client-local" node (the node that the client declaring the queue is connected to), regardless of the queue_leader_locator value.

Replicated and Distributed Queues​

Quorum queues is replicated, data safety and consistency-oriented queue type. Classic queues historically supported replication but this feature was removed for RabbitMQ 4.x.

Any client connection can use any queue, whether it is replicated or not, regardless of the node the queue replica is hosted on or the node the client is connected to. RabbitMQ will route the operations to the appropriate node transparently for clients.

For example, in a cluster with nodes A, B and C, a client connected to node A can consume from a queue Q hosted on B, while a client connected to node C can publish in a way that routes messages to queue Q.

Client libraries or applications may choose to connect to the node that hosts the current leader replica of a specific queue for improved data locality.

This general rule applies to all messaging data types supported by RabbitMQ except for one. Streams are an exception to this rule, and require clients, regardless of the protocol they use, to connect to a node that hosts a replica (a leader of rollower) of the target stream. Consequently, RabbitMQ Stream protocol clients will connect to multiple nodes in parallel.

Queues can also be federated across loosely coupled nodes or clusters.

Note that intra-cluster replication and federation are orthogonal features and should not be considered direct alternatives.

Streams is another replicated data structure supported by RabbitMQ, with a different set of supported operations and features.

Non-Replicated Queues and Client Operations​

Any client connection can use any queue, including non-replicated (single replica) queues, regardless of the node the queue replica is hosted on or the node the client is connected to. RabbitMQ will route the operations to the appropriate node transparently for clients.

For example, in a cluster with nodes A, B and C, a client connected to node A can consume from a queue Q hosted on B, while a client connected to node C can publish in a way that routes messages to queue Q.

Client libraries or applications may choose to connect to the node that hosts the current leader replica of a specific queue for improved data locality.

This general rule applies to all messaging data types supported by RabbitMQ except for one. Streams are an exception to this rule, and require clients, regardless of the protocol they use, to connect to a node that hosts a replica (a leader of rollower) of the target stream. Consequently, RabbitMQ Stream protocol clients will connect to multiple nodes in parallel.

Time-to-Live and Length Limit​

Queues can have their length limited. Queues and messages can have a TTL.

Both features can be used for data expiration and as a way of limiting how many resources (RAM, disk space) a queue can use at most, e.g. when consumers go offline or their throughput falls behind publishers.

In Durable and In-Memory Storage​

In modern RabbitMQ versions, quorum queues and classic queues v2 alike actively move data to disk and only keep a relatively small working set in memory.

In some protocols (e.g. AMQP 0-9-1) clients can publish messages as persistent or transient. Transient messages will still be stored on disk but will be discarded during the next node restart.

In AMQP 0-9-1, this is done via a message property (delivery_mode or, in some clients, persistent).

Other relevant guides on the topic are Quorum Queues, Streams, Reasoning About Memory Usage, Alarms, Memory Alarms, Free Disk Space Alarms, Deployment Guidelines, and Message Store Configuration.

Priorities​

Queues can have 0 or more priorities. This feature is opt-in: only queues that have maximum number of priorities configured via an optional argument (see above) will do prioritisation.

Publishers specify message priority using the priority field in message properties.

If priority queues are desired, we recommend using between 1 and 10. Currently using more priorities will consume more resources (Erlang processes).

CPU Utilisation and Parallelism Considerations​

Currently a single queue replica (whether leader or follower) is limited to a single CPU core on its hot code path. This design therefore assumes that most systems use multiple queues in practice. A single queue is generally considered to be an anti-pattern (and not just for resource utilisation reasons).

In case when it is desirable to trade off message ordering for parallelism (better CPU core utilisation), rabbitmq-sharding provides an opinionated way of doing so transparently to the clients.

Metrics and Monitoring​

RabbitMQ collects multiple metrics about queues. Most of them are available via RabbitMQ HTTP API and management UI, which is designed for monitoring. This includes queue length, ingress and egress rates, number of consumers, number of messages in various states (e.g. ready for delivery or unacknowledged), number of messages in RAM vs. on disk, and so on.

rabbitmqctl can list queues and some basic metrics.

Runtime metrics such as VM scheduler usage, queue (Erlang) process GC activity, amount of RAM used by the queue process, queue process mailbox length can be accessed using the rabbitmq-top plugin and individual queue pages in the management UI.

Consumers and Acknowledgements​

Messages can be consumed by registering a consumer (subscription), which means RabbitMQ will push messages to the client, or fetched individually for protocols that support this (e.g. the basic.get AMQP 0-9-1 method), similarly to HTTP GET.

Delivered messages can be acknowledged by consumer explicitly or automatically as soon as a delivery is written to connection socket.

Automatic acknowledgement mode generally will provide higher throughput rate and uses less network bandwidth. However, it offers the least number of guarantees when it comes to failures. As a rule of thumb, consider using manual acknowledgement mode first.

Prefetch and Consumer Overload​

Automatic acknowledgement mode can also overwhelm consumers which cannot process messages as quickly as they are delivered. This can result in permanently growing memory usage and/or OS swapping for the consumer process.

Manual acknowledgement mode provides a way to set a limit on the number of outstanding (unconfirmed) deliveries: channel QoS (prefetch).

Consumers using higher (several thousands or more) prefetch levels can experience the same overload problem as consumers using automatic acknowledgements.

High number of unacknowledged messages will lead to higher memory usage by the broker.

Message States​

Enqueued messages therefore can be in one of two states:

Message breakdown by state can be found in the management UI.

Determining Queue Length​

It is possible to determine queue length in a number of ways:

  • With AMQP 0-9-1, using a property on the queue.declare method response (queue.declare-ok). The field name is message_count. How it is accessed varies from client library to client library.
  • Using RabbitMQ HTTP API.
  • Using the rabbitmqctl list_queues command.

Queue length is defined as the number of messages ready for delivery.

Avoid Temporary Queues with Well-Known Names​

A temporary queue that is not exclusive can be client named and shared between multiple consumers. This, however, is not recommended and can lead to a race condition between RabbitMQ node operations and client recovery.

Consider the following scenario:

  • A consumer uses an auto-delete queue with a well-known names
  • Client's connection fails
  • Client detects it and initiates connection recovery

As the failed connection which had the only consumer on an auto-delete queue, the queue must be deleted by RabbitMQ. This operation will take some time, during which the consumer may recover.

Then depending on the timing of operations, the queue can be

  1. Declared by the recovering client and then deleted
  2. Deleted and then re-declared

In the first case, the client will try to re-register its consumer on a queue that's been concurrently deleted, which will lead to a channel exception.

There are two solutions to this fundamental race condition:

  1. Introduce a connection recovery delay. For example, several RabbitMQ client libraries use a connection recovery delay of 5 seconds by default
  2. Use server-named queues, which side steps the problem entirely since the new client connection will use a different queue name from its predecessor