Quorum Queues

Overview

The RabbitMQ quorum queue is a modern queue type which implements a durable, replicated queue based on the Raft consensus algorithm and should be considered the default choice when needing a replicated, highly available queue.

Quorum queues are designed for excellent data safety as well as reliable and fast leader election properties to ensure high availability even during upgrades or other turbulence.

Quorum queues are optimized for certain use cases where data safety is the top priority. This is covered in the Motivation section.

Quorum queues have differences in behaviour compared to classic queues as well as some limitations that it is important to be aware of when converting an application from using classic to quorum queues.

Some features are specific to quorum queues such as poison message handling , at least once dead-lettering and the modified outcome when using AMQP.

For cases that would benefit from replication and repeatable reads, streams may be a better option than quorum queues.

Quorum queues and streams are the two replicated data structures available. Classic queue mirroring was removed starting with RabbitMQ 4.0.

Topics Covered

Topics covered in this document include:

• What are quorum queues and why they were introduced
• How are they different from classic queues
• Primary use cases of quorum queues and when not to use them
• How to declare a quorum queue
• Replication-related topics: replica management, replica leader rebalancing, optimal number of replicas, etc
• What guarantees quorum queues offer in terms of leader failure handling, data safety and availability
• Continuous Membership Reconciliation
• Performance characteristics of quorum queues and performance tuning relevant to them
• Poison message handling provided by quorum queues
• Configurable settings of quorum queues
• Resource use of quorum queues, most importantly their memory footprint

and more.

General familiarity with RabbitMQ clustering would be helpful here when learning more about quorum queues.

Motivation

Quorum queues adopt a different replication and consensus protocol and give up support for certain "transient" in nature features, which results in some limitations. These limitations are covered later in this information.

Quorum queues pass a refactored and more demanding version of the original Jepsen test. This ensures they behave as expected under network partitions and failure scenarios. The new test runs continuously to spot possible regressions and is enhanced regularly to test new features (e.g. dead lettering).

What is a Quorum?

If intentionally simplified, quorum in a distributed system can be defined as an agreement between the majority of nodes ((N/2)+1 where N is the total number of system participants).

When applied to queue mirroring in RabbitMQ clusters this means that the majority of replicas (including the currently elected queue leader) agree on the state of the queue and its contents.

Use Cases

Quorum queues intended use is in topologies where queues exist for a long time and are critical to certain aspects of an application's architecture.

Examples of good use cases would be incoming orders in a sales system or votes cast in an electoral system where potentially losing messages would have a significant impact on overall system correctness and function.

Quorum queues are not designed to be used for every problem. Stock tickers, instant messaging systems and RPC reply queues benefit less or not at all from use of quorum queues.

Publishers should use publisher confirms as this is how clients can interact with the quorum queue consensus system. Publisher confirms will only be issued once a published message has been successfully replicated to a quorum of replicas and is considered "safe" within the context of the queue.

Consumers should use manual acknowledgements to ensure messages that aren't successfully processed are returned to the queue so that another consumer can re-attempt processing.

When Not to Use Quorum Queues

In some cases quorum queues should not be used. They typically involve:

• Temporary queues: transient or exclusive queues, high queue churn (declaration and deletion rates)
• Lowest possible latency: the underlying consensus algorithm has an inherently higher latency due to its data safety features
• When data safety is not a priority (e.g. applications do not use manual acknowledgements and publisher confirms are not used)
• Very long queue backlogs (5M+ messages) (streams are likely to be a better fit)
• Large fanouts: (streams are likely to be a better fit)

Features

Quorum queues share most of the fundamentals with other queue types.

Comparison with Classic Queues

The following operations work the same way for quorum queues as they do for classic queues:

• Consumption, consumer registration
• Consumer acknowledgements (except for global QoS and prefetch)
• Consumer cancellation
• Purging
• Deletion

With some queue operations there are minor differences:

• Declaration
• Setting prefetch for consumers

Feature Matrix

Feature	Classic queues	Quorum queues
Non-durable queues	yes	no
Message replication	no	yes
Exclusivity	yes	no
Per message persistence	per message	always
Membership changes	no	semi-automatic
Message TTL (Time-To-Live)	yes	yes
Queue TTL	yes	partially (lease is not renewed on queue re-declaration)
Queue length limits	yes	yes (except x-overflow: reject-publish-dlx)
Keeps messages in memory	see Classic Queues	never (see Resource Use)
Message priority	yes	yes
Single Active Consumer	yes	yes
Consumer exclusivity	yes	no (use Single Active Consumer)
Consumer priority	yes	yes
Dead letter exchanges	yes	yes
Adheres to policies	yes	yes (see Policy support)
Poison message handling	no	yes
Global QoS Prefetch	yes	no
Server-named queues	yes	no

Modern quorum queues also offer higher throughput and less latency variability for many workloads.

Queue and Per-Message TTL

Quorum queues support both Queue TTL and message TTL (since RabbitMQ 3.10) (including Per-Queue Message TTL in Queues and Per-Message TTL in Publishers). When using any form of message TTL, the memory overhead increases by 2 bytes per message.

Length Limit

Quorum queues has support for queue length limits.

The drop-head and reject-publish overflow behaviours are supported but they do not support reject-publish-dlx configurations as Quorum queues take a different implementation approach than classic queues.

The current implementation of reject-publish overflow behaviour does not strictly enforce the limit and allows a quorum queue to overshoot its limit by at least one message, therefore it should be taken with care in scenarios where a precise limit is required.

When a quorum queue reaches the max-length limit and reject-publish is configured it notifies each publishing channel who from thereon will reject all messages back to the client. This means that quorum queues may overshoot their limit by some small number of messages as there may be messages in flight whilst the channels are notified. The number of additional messages that are accepted by the queue will vary depending on how many messages are in flight at the time.

Dead Lettering

Quorum queues support dead letter exchanges (DLXs).

Traditionally, using DLXs in a clustered environment has not been safe.

Since RabbitMQ 3.10 quorum queues support a safer form of dead-lettering that uses at-least-once guarantees for the message transfer between queues (with the limitations and caveats outlined below).

This is done by implementing a special, internal dead-letter consumer process that works similarly to a normal queue consumer with manual acknowledgements apart from it only consumes messages that have been dead-lettered.

This means that the source quorum queue will retain the dead-lettered messages until they have been acknowledged. The internal consumer will consume dead-lettered messages and publish them to the target queue(s) using publisher confirms. It will only acknowledge once publisher confirms have been received, hence providing at-least-once guarantees.

at-most-once remains the default dead-letter-strategy for quorum queues and is useful for scenarios where the dead lettered messages are more of an informational nature and where it does not matter so much if they are lost in transit between queues or when the overflow configuration restriction outlined below is not suitable.

Activating at-least-once dead-lettering

To activate or turn on at-least-once dead-lettering for a source quorum queue, apply all of the following policies (or the equivalent queue arguments starting with x-):

• Set dead-letter-strategy to at-least-once (default is at-most-once).
• Set overflow to reject-publish (default is drop-head).
• Configure a dead-letter-exchange.
• Turn on feature flag stream_queue (turned on by default for RabbitMQ clusters created in 3.9 or later).

It is recommended to additionally configure max-length or max-length-bytes to prevent excessive message buildup in the source quorum queue (see caveats below).

Optionally, configure a dead-letter-routing-key.

Limitations

at-least-once dead lettering does not work with the default drop-head overflow strategy even if a queue length limit is not set. Hence if drop-head is configured the dead-lettering will fall back to at-most-once. Use the overflow strategy reject-publish instead.

Caveats

at-least-once dead-lettering will require more system resources such as memory and CPU. Therefore, turn on at-least-once only if dead lettered messages should not be lost.

at-least-once guarantees opens up some specific failure cases that needs handling. As dead-lettered messages are now retained by the source quorum queue until they have been safely accepted by the dead-letter target queue(s) this means they have to contribute to the queue resource limits, such as max length limits so that the queue can refuse to accept more messages until some have been removed. Theoretically it is then possible for a queue to only contain dead-lettered messages, in the case where, say a target dead-letter queue isn't available to accept messages for a long time and normal queue consumers consume most of the messages.

Dead-lettered messages are considered "live" until they have been confirmed by the dead-letter target queue(s).

There are few cases for which dead lettered messages will not be removed from the source queue in a timely manner:

• The configured dead-letter exchange does not exist.
• The messages cannot be routed to any queue (equivalent to the mandatory message property).
• One (of possibly many) routed target queues does not confirm receipt of the message. This can happen when a target queue is not available or when a target queue rejects a message (e.g. due to exceeded queue length limit).

The dead-letter consumer process will retry periodically if either of the scenarios above occur which means there is a possibility of duplicates appearing at the DLX target queue(s).

For each quorum queue with at-least-once dead-lettering turned on, there will be one internal dead-letter consumer process. The internal dead-letter consumer process is co-located on the quorum queue leader node. It keeps all dead-lettered message bodies in memory. It uses a prefetch size of 32 messages to limit the amount of message bodies kept in memory if no confirms are received from the target queues.

That prefetch size can be increased by the dead_letter_worker_consumer_prefetch setting in the rabbit app section of the advanced config file if high dead-lettering throughput (thousands of messages per second) is required.

For a source quorum queue, it is possible to switch dead-letter strategy dynamically from at-most-once to at-least-once and vice versa. If the dead-letter strategy is changed either directly from at-least-once to at-most-once or indirectly, for example by changing overflow from reject-publish to drop-head, any dead-lettered messages that have not yet been confirmed by all target queues will be deleted.

Messages published to the source quorum queue are persisted on disk regardless of the message delivery mode (transient or persistent). However, messages that are dead lettered by the source quorum queue will keep the original message delivery mode. This means if dead lettered messages in the target queue should survive a broker restart, the target queue must be durable and the message delivery mode must be set to persistent when publishing messages to the source quorum queue.

Priorities

important

Quorum queue priority support is available as of RabbitMQ 4.0. However, there are differences in how quorum queues and classic queues implement priorities.

Quorum queues support consumer priorities and starting with 4.0, they also support a type of message prioritisation that is quite different from classic queue message priorities.

Quorum queue message priorities are always active and do not require a policy to work. As soon as a quorum queue receives a message with a priority set it will enable prioritization.

Quorum queues internally only support two priorities: high and normal. Messages without a priority set will be mapped to normal as will priorities 0 - 4. Messages with a priority higher than 4 will be mapped to high.

High priority messages will be favoured over normal priority messages at a ratio of 2:1, i.e. for every 2 high priority message the queue will deliver 1 normal priority message (if available). Hence, quorum queues implement a kind of non-strict, "fair share" priority processing. This ensures progress is always made on normal priority messages but high priorities are favoured at a ratio of 2:1.

important

If a high priority message was published before a normal priority one, the high priority message will always be delivered first even if it is the normal priority's turn.

More Advanced Scenarios

For more advanced message priority scenarios, separate queues should be used for different message types, one for each type (priority). Queues used for more important message types should generally have more (overprovisioned) consumers.

Poison Message Handling

Quorum queues support handling of poison messages, that is, messages that cause a consumer to repeatedly requeue a delivery (possibly due to a consumer failure) such that the message is never consumed completely and positively acknowledged so that it can be marked for deletion by RabbitMQ.

Quorum queues keep track of the number of unsuccessful delivery attempts and expose it in the x-delivery-count header/message-annotation that is included with any redelivered message. The x-delivery-count header/message-annotation includes all returns to the queue including failed and deliberate ones. For AMQP the header section delivery-count field will only include failures (client connection lost, returns using the modified outcome with delivery-failed=true) and is thus a separate counter.

info

Starting with RabbitMQ 4.0, the delivery limit for quorum queues defaults to 20

This can be increased or decreased using a policy argument, delivery-limit. There is no way to set it to the old, unlimited default.

When a message has been returned more times than the limit the message will be dropped or dead-lettered (if a DLX is configured).

It is recommended that all quorum queues have a dead letter configuration of some sort to ensure messages aren't dropped and lost unintentionally. Using a single stream for a low priority dead letter policy is a good, low resource way to ensure dropped messages are retained for some time after.

Policy Support

Quorum queues can be configured via RabbitMQ policies. The below table summarises the policy keys they adhere to.

Definition Key	Type
max-length	Number
max-length-bytes	Number
overflow	"drop-head" or "reject-publish"
expires	Number (milliseconds)
dead-letter-exchange	String
dead-letter-routing-key	String
delivery-limit	Number

Features that are not Supported

Transient (non-Durable) Queues

Classic queues can be non-durable. Quorum queues are always durable per their assumed use cases.

Exclusivity

Exclusive queues are tied to the lifecycle of their declaring connection. Quorum queues by design are replicated and durable, therefore the exclusive property makes no sense in their context. Therefore quorum queues cannot be exclusive.

Quorum queues are not meant to be used as temporary queues.

Global QoS

Quorum queues do not support global QoS prefetch where a channel sets a single prefetch limit for all consumers using that channel. If an attempt is made to consume from a quorum queue from a channel with global QoS activated a channel error will be returned.

Use per-consumer QoS prefetch, which is the default in several popular clients.

Usage

Quorum queues share most of the fundamentals with other queue types. Any AMQP 0.9.1 client library that can specify optional queue arguments when declaring will be able to use quorum queues.

First we will cover how to declare a quorum queue using AMQP 0.9.1.

Declaring

To declare a quorum queue set the x-queue-type queue argument to quorum (the default is classic). This argument must be provided by a client at queue declaration time; it cannot be set or changed using a policy. This is because policy definition or applicable policy can be changed dynamically but queue type cannot. It must be specified at the time of declaration.

Declaring a queue with an x-queue-type argument set to quorum will declare a quorum queue with up to three replicas (default replication factor), one per each cluster node.

For example, a cluster of three nodes will have three replicas, one on each node. In a cluster of five nodes, three nodes will have one replica each but two nodes won't host any replicas.

After declaration a quorum queue can be bound to any exchange just as any other RabbitMQ queue.

If declaring using management UI, queue type must be specified using the queue type drop down menu.

Client Operations

The following operations work the same way for quorum queues as they do for classic queues:

• Consumption (subscription)
• Consumer acknowledgements (keep QoS Prefetch Limitations in mind)
• Cancellation of consumers
• Purging of queue messages
• Queue deletion

With some queue operations there are minor differences:

• Declaration (covered above)
• Setting QoS prefetch for consumers

Replication Factor and Membership Management

When a quorum queue is declared, an initial number of replicas for it must be started in the cluster. By default the number of replicas to be started is up to three, one per RabbitMQ node in the cluster.

Three nodes is the practical minimum of replicas for a quorum queue. In RabbitMQ clusters with a larger number of nodes, adding more replicas than a quorum (majority) will not provide any improvements in terms of quorum queue availability but it will consume more cluster resources.

Therefore the recommended number of replicas for a quorum queue is the quorum of cluster nodes (but no fewer than three). This assumes a fully formed cluster of at least three nodes.

Controlling the Initial Replication Factor

For example, a cluster of three nodes will have three replicas, one on each node. In a cluster of seven nodes, three nodes will have one replica each but four more nodes won't host any replicas of the newly declared queue.

The replication factor (number of replicas a queue has) can be configured for quorum queues.

The minimum factor value that makes practical sense is three. It is highly recommended for the factor to be an odd number. This way a clear quorum (majority) of nodes can be computed. For example, there is no "majority" of nodes in a two node cluster. This is covered with more examples below in the Fault Tolerance and Minimum Number of Replicas Online section.

This may not be desirable for larger clusters or for cluster with an even number of nodes. To control the number of quorum queue members set the x-quorum-initial-group-size queue argument when declaring the queue. The group size argument provided should be an integer that is greater than zero and smaller or equal to the current RabbitMQ cluster size. The quorum queue will be launched to run on a random subset of RabbitMQ nodes present in the cluster at declaration time.

In case a quorum queue is declared before all cluster nodes have joined the cluster, and the initial replica count is greater than the total number of cluster members, the effective value used will be equal to the total number of cluster nodes. When more nodes join the cluster, the replica count will not be automatically increased but it can be increased by the operator.

Managing Replicas

Replicas of a quorum queue are explicitly managed by the operator. When a new node is added to the cluster, it will host no quorum queue replicas unless the operator explicitly adds it to a member (replica) list of a quorum queue or a set of quorum queues.

When a node has to be decommissioned (permanently removed from the cluster), the forget_cluster_node command will automatically attempt to remove all quorum queue members on the decommissioned node. Alternatively the shrink command below can be used ahead of node removal to move any replicas to a new node.

Also see Continuous Membership Reconciliation for a more automated way to grow quorum queues.

Several CLI commands are provided to perform the above operations:

rabbitmq-queues add_member [-p <vhost>] <queue-name> <node>

rabbitmq-queues delete_member [-p <vhost>] <queue-name> <node>

rabbitmq-queues grow <node> <all | even> [--vhost-pattern <pattern>] [--queue-pattern <pattern>]

rabbitmq-queues shrink <node> [--errors-only]

To successfully add and remove members a quorum of replicas in the queue must be available because membership changes are treated as queue state changes.

Care needs to be taken not to accidentally make a queue unavailable by losing the quorum whilst performing maintenance operations that involve membership changes.

When replacing a cluster node, it is safer to first add a new node, grow any quorum queues that need a member on the new node and then decommission the node it replaces.

Queue Leader Location

Every quorum queue has a primary replica. That replica is called queue leader. All queue operations go through the leader first and then are replicated to followers. This is necessary to guarantee FIFO ordering of messages.

To avoid some nodes in a cluster hosting the majority of queue leader replicas and thus handling most of the load, queue leaders should be reasonably evenly distributed across cluster nodes.

When a new quorum queue is declared, the set of nodes that will host its replicas is randomly picked, but will always include the node the client that declares the queue is connected to.

Which replica becomes the initial leader can controlled using three options:

1. Setting the queue-leader-locator policy key (recommended)
2. By defining the queue_leader_locator key in the configuration file (recommended)
3. Using the x-queue-leader-locator optional queue argument

Supported queue leader locator values are

• client-local: Pick the node the client that declares the queue is connected to. This is the default value.
• balanced: If there are overall less than 1000 queues (classic queues, quorum queues, and streams), pick the node hosting the minimum number of quorum queue leaders. If there are overall more than 1000 queues, pick a random node.

Rebalancing Replicas

Once declared, the RabbitMQ quorum queue leaders may be unevenly distributed across the RabbitMQ cluster. To re-balance use the rabbitmq-queues rebalance command. It is important to know that this does not change the nodes which the quorum queues span. To modify the membership instead see managing replicas.

# rebalances all quorum queues
rabbitmq-queues rebalance quorum

It is possible to rebalance a subset of queues selected by name:

# rebalances a subset of quorum queues
rabbitmq-queues rebalance quorum --queue-pattern "orders.*"

or quorum queues in a particular set of virtual hosts:

# rebalances a subset of quorum queues
rabbitmq-queues rebalance quorum --vhost-pattern "production.*"

Continuous Membership Reconciliation (CMR)

important

The continuous membership reconciliation (CMR) feature exists in addition to, and not as a replacement for, explicit replica management. In certain cases where nodes are permanently removed from the cluster, explicitly removing quorum queue replicas may still be necessary.

In addition to controlling quorum queue replica membership by using the initial target size and explicit replica management, nodes can be configured to automatically try to grow the quorum queue replica membership to a configured target group size by enabling the continuous membership reconciliation feature.

When activated, every quorum queue leader replica will periodically check its current membership group size (the number of replicas online), and compare it with the target value.

If a queue is below the target value, RabbitMQ will attempt to grow the queue onto the availible nodes that do not currently host replicas of said queue, if any, up to the target value.

When is Continuous Membership Reconciliation Triggered?

The default reconciliation interval is 60 minutes. In addition, automatic reconciliation is triggered by certain events in the cluster, such as an addition of a new node, or permanent node removal, or a quorum queue-related policy change.

warning

Note that a node or quorum queue replica failure does not trigger automatic membership reconciliation.

If a node is failed in an unrecoverable way and cannot be brought back, it must be explicitly removed from the cluster or the operator must opt-in and enable the quorum_queue.continuous_membership_reconciliation.auto_remove setting.

This also means that upgrades do not trigger automatic membership reconciliation since nodes are expected to come back and only a minority (often just one) node is stopped for upgrading at a time.

CMR Configuration

rabbitmq.conf

Continuous Membership Reconciliation (CMR) Settings

rabbitmq.conf configuration key	Description
quorum_queue.continuous_membership_reconciliation.enabled	Enables or disables continuous membership reconciliation. Data type: boolean Default: true
quorum_queue.continuous_membership_reconciliation.target_group_size	The target replica count (group size) for queue members. Data type: positive integer Default: none
quorum_queue.continuous_membership_reconciliation.auto_remove	Enables or disables automatic removal of member nodes that are no longer part of the cluster, but still a member of the quorum queue. Data type: boolean Default: false
quorum_queue.continuous_membership_reconciliation.interval	The default evaluation interval in milliseconds. Data type: positive integer Default: 3600000 (60 minutes)
quorum_queue.continuous_membership_reconciliation.trigger_interval	The reconciliation delay in milliseconds, used when a trigger event occurs, for example, a node is added or removed from the cluster or an applicable policy changes. This delay will be applied only once, then the regular interval will be used again. Data type: positive integer Default: 10000 (10 seconds)

Policy Keys

Policy-driven CMR Settings

Policy key	Description
target-group-size	Defines the target replica count (group size) for matching queues. This policy can be set by users and operators. Data type: positive integer Default: none

Optional Arguments-driven CMR Settings

Optional arguments key	Description
x-quorum-target-group-size	Defines the target replica count (group size) for matching queues. This key can be overridden by operator policies. Data type: positive integer Default: none

Quorum Queue Behaviour

A quorum queue relies on a consensus protocol called Raft to ensure data consistency and safety.

Every quorum queue has a primary replica (a leader in Raft parlance) and zero or more secondary replicas (called followers).

A leader is elected when the cluster is first formed and later if the leader becomes unavailable.

Leader Election and Failure Handling

A quorum queue requires a quorum of the declared nodes to be available to function. When a RabbitMQ node hosting a quorum queue's leader fails or is stopped another node hosting one of that quorum queue's follower will be elected leader and resume operations.

Failed and rejoining followers will re-synchronise ("catch up") with the leader. With quorum queues, a temporary replica failure does not require a full re-synchronization from the currently elected leader. Only the delta will be transferred if a re-joining replica is behind the leader. This "catching up" process does not affect leader availability.

When a new replica is added, it will synchronise the entire queue state from the leader.

Fault Tolerance and Minimum Number of Replicas Online

Consensus systems can provide certain guarantees with regard to data safety. These guarantees do mean that certain conditions need to be met before they become relevant such as requiring a minimum of three cluster nodes to provide fault tolerance and requiring more than half of members to be available to work at all.

Failure tolerance characteristics of clusters of various size can be described in a table:

Cluster node count	Tolerated number of node failures	Tolerant to a network partition
1	0	not applicable
2	0	no
3	1	yes
4	1	yes if a majority exists on one side
5	2	yes
6	2	yes if a majority exists on one side
7	3	yes
8	3	yes if a majority exists on one side
9	4	yes

As the table above shows RabbitMQ clusters with fewer than three nodes do not benefit fully from the quorum queue guarantees. RabbitMQ clusters with an even number of RabbitMQ nodes do not benefit from having quorum queue members spread over all nodes. For these systems the quorum queue size should be constrained to a smaller uneven number of nodes.

Performance tails off quite a bit for quorum queue node sizes larger than 5. We do not recommend running quorum queues on more than 7 RabbitMQ nodes. The default quorum queue size is 3 and is controllable using the x-quorum-initial-group-size queue argument.

Data Safety

Quorum queues are designed to provide data safety under network partition and failure scenarios. A message that was successfully confirmed back to the publisher using the publisher confirms feature should not be lost as long as at least a majority of RabbitMQ nodes hosting the quorum queue are not permanently made unavailable.

Generally quorum queues favours data consistency over availability.

No guarantees are provided for messages that have not been confirmed using the publisher confirm mechanism. Such messages could be lost "mid-way", in an operating system buffer or otherwise fail to reach the queue leader.

Availability

A quorum queue should be able to tolerate a minority of queue members becoming unavailable with no or little effect on availability.

Note that depending on the partition handling strategy used RabbitMQ may restart itself during recovery and reset the node but as long as that does not happen, this availability guarantee should hold true.

For example, a queue with three replicas can tolerate one node failure without losing availability. A queue with five replicas can tolerate two, and so on.

If a quorum of nodes cannot be recovered (say if 2 out of 3 RabbitMQ nodes are permanently lost) the queue is permanently unavailable and will need to be force deleted and recreated.

Quorum queue follower replicas that are disconnected from the leader or participating in a leader election will ignore queue operations sent to it until they become aware of a newly elected leader. There will be warnings in the log (received unhandled msg and similar) about such events. As soon as the replica discovers a newly elected leader, it will sync the queue operation log entries it does not have from the leader, including the dropped ones. Quorum queue state will therefore remain consistent.

Performance Characteristics

Quorum queues are designed to trade latency for throughput and have been tested in 3, 5 and 7 node configurations with several different message sizes.

In scenarios using both consumer acks and publisher confirms quorum queues have been observed to have superior throughput to classic mirrored queues (deprecated in 2021, removed in 2024 for RabbitMQ 4.0). For example, take a look at these benchmarks with 3.10 and another with 3.12.

As quorum queues persist all data to disks before doing anything it is recommended to use the fastest disks possible and certain Performance Tuning settings.

Quorum queues also benefit from consumers using higher prefetch values to ensure consumers aren't starved whilst acknowledgements are flowing through the system and allowing messages to be delivered in a timely fashion.

Due to the disk I/O-heavy nature of quorum queues, their throughput decreases as message sizes increase.

Quorum queue throughput is also affected by the number of replicas. The more replicas a quorum queue has, the lower its throughput generally will be since more work has to be done to replicate data and achieve consensus.

Configurable Settings

There are a few new configuration parameters that can be tweaked using the advanced config file.

Note that all settings related to resource footprint are documented in a separate section.

The ra application (which is the Raft library that quorum queues use) has its own set of tunable parameters.

The rabbit application has several quorum queue related configuration items available.

advanced.config configuration key	Description	Default value
rabbit.quorum_cluster_size	Sets the default quorum queue cluster size (can be over-ridden by the x-quorum-initial-group-size queue argument at declaration time.	3
rabbit.quorum_commands_soft_limit	This is a flow control related parameter defining the maximum number of unconfirmed messages a channel accepts before entering flow. The current default is configured to provide good performance and stability when there are multiple publishers sending to the same quorum queue. If the applications typically only have a single publisher per queue this limit could be increased to provide somewhat better ingress rates.	32

Example of a Quorum Queue Configuration

The following advanced.config example modifies all values listed above:

[
 %% five replicas by default, only makes sense for nine node clusters
 {rabbit, [{quorum_cluster_size, 5},
           {quorum_commands_soft_limit, 64}]}
].

Resource Use

Quorum queues are optimised for data safety and performance. Each quorum queue process maintains an in-memory index of the messages in the queue, which requires at least 32 bytes of metadata for each message (more, if the message was returned or has a TTL set). A quorum queue process will therefore use at least 1MB for every 30000 messages in the queue (message size is irrelevant). You can perform back-of-the-envelope calculations based on the number of queues and expected or maximum number of messages in them). Keeping the queues short is the best way to maintain low memory usage. Setting the maximum queue length for all queues is a good way to limit the total memory usage if the queues become long for any reason.

Additionally, quorum queues on a given node share a write-ahead-log (WAL) for all operations. WAL entries are stored both in memory and written to disk. When the current WAL file reaches a predefined limit, it is flushed to segment files on disk for each quorum queue member on a node and the system will begin to release the memory used by that batch of log entries. The segment files are then compacted over time as consumers acknowledge deliveries. Compaction is the process that reclaims disk space.

The WAL file size limit at which it is flushed to disk can be controlled:

# Flush current WAL file to a segment file on disk once it reaches 64 MiB in size
raft.wal_max_size_bytes = 64000000

The value defaults to 512 MiB. This means that during steady load, the WAL table memory footprint can reach 512 MiB. You can expect your memory usage to look like this:

Because memory deallocation may take some time, we recommend that the RabbitMQ node is allocated at least 3 times the memory of the default WAL file size limit. More will be required in high-throughput systems. 4 times is a good starting point for those.

Repeated Requeues

Internally quorum queues are implemented using a log where all operations including messages are persisted. To avoid this log growing too large it needs to be truncated regularly. To be able to truncate a section of the log all messages in that section needs to be acknowledged. Usage patterns that continuously reject or nack the same message with the requeue flag set to true could cause the log to grow in an unbounded fashion and eventually fill up the disks. Therefore since RabbitMQ 4.0 a default delivery-limit of 20 is always set after which the message will be dropped or dead lettered.

Messages that are rejected or nacked back to a quorum queue will be returned to the back of the queue if no delivery-limit is set. This avoids the above scenario where repeated re-queues causes the Raft log to grow in an unbounded manner. If a delivery-limit is set it will use the original behaviour of returning the message near the head of the queue. (NB: since 4.0 a delivery-limit will only be unset if a queue was upgraded from a prior version, newly declared queues will always have a default of 20).

Increased Atom Use

The internal implementation of quorum queues converts the queue name into an Erlang atom. If queues with arbitrary names are continuously created and deleted it may threaten the long term stability of the RabbitMQ system if the size of the atom table reaches the default limit of 5 million.

While quorum queues were not designed to be used in high churn environments (non-mirrored classic queues are the optimal choice for those), the limit can be increased if really necessary.

See the Runtime guide to learn more.

Performance Tuning

This section aims to cover a couple of tunable parameters that may increase throughput of quorum queues for some workloads. Other workloads may not see any increases, or observe decreases in throughput, with these settings.

Use the values and recommendations here as a starting point and conduct your own benchmark (for example, using PerfTest) to conclude what combination of values works best for a particular workloads.

Tuning: Raft Segment File Entry Count

Workloads with small messages and higher message rates can benefit from the following configuration change that increases the number of Raft log entries (such as enqueued messages) that are allowed in a segment file. Having fewer segment files can be beneficial when consuming from a queue with a long backlog:

# Positive values up to 65535 are allowed, the default is 4096.
raft.segment_max_entries = 32768

Values greater than 65535 are not supported.

Tuning: Linux Readahead

In addition, the aforementioned workloads with a higher rate of small messages can benefit from a higher readahead, a configurable block device parameter of storage devices on Linux.

To inspect the effective readahead value, use blockdev --getra and specify the block device that hosts RabbitMQ node data directory:

# This is JUST AN EXAMPLE.
# The name of the block device in your environment will be different.
#
# Displays effective readahead value device /dev/sda.
sudo blockdev --getra /dev/sda

To configure readahead, use blockdev --setra for the block device that hosts RabbitMQ node data directory:

# This is JUST AN EXAMPLE.
# The name of the block device in your environment will be different.
# Values between 256 and 4096 in steps of 256 are most commonly used.
#
# Sets readahead for device /dev/sda to 4096.
sudo blockdev --setra 4096 /dev/sda