Skip to main content

Clustering Guide

Overview

This guide covers fundamental topics related to RabbitMQ clustering:

and more. Cluster Formation and Peer Discovery is a closely related guide that focuses on peer discovery and cluster formation automation-related topics. For queue contents (message) replication, see the Quorum Queues guide.

VMware RabbitMQ provides an Intra-cluster Compression feature.

A RabbitMQ cluster is a logical grouping of one or several nodes, each sharing users, virtual hosts, queues, exchanges, bindings, runtime parameters and other distributed state.

Cluster Formation

Ways of Forming a Cluster

A RabbitMQ cluster can be formed in a number of ways:

Please refer to the Cluster Formation guide for details.

The composition of a cluster can be altered dynamically. All RabbitMQ brokers start out as running on a single node. These nodes can be joined into clusters, and subsequently turned back into individual brokers again.

Node Names (Identifiers)

RabbitMQ nodes are identified by node names. A node name consists of two parts, a prefix (usually rabbit) and hostname. For example, rabbit@node1.messaging.svc.local is a node name with the prefix of rabbit and hostname of node1.messaging.svc.local.

Node names in a cluster must be unique. If more than one node is running on a given host (this is usually the case in development and QA environments), they must use different prefixes, e.g. rabbit1@hostname and rabbit2@hostname.

In a cluster, nodes identify and contact each other using node names. This means that the hostname part of every node name must resolve. CLI tools also identify and address nodes using node names.

When a node starts up, it checks whether it has been assigned a node name. This is done via the RABBITMQ_NODENAME environment variable. If no value was explicitly configured, the node resolves its hostname and prepends rabbit to it to compute its node name.

If a system uses fully qualified domain names (FQDNs) for hostnames, RabbitMQ nodes and CLI tools must be configured to use so called long node names. For server nodes this is done by setting the RABBITMQ_USE_LONGNAME environment variable to true.

For CLI tools, either RABBITMQ_USE_LONGNAME must be set or the --longnames option must be specified.

Cluster Formation Requirements

Hostname Resolution

RabbitMQ nodes address each other using a node name, a combination of a prefix and domain name, either short or fully-qualified (FQDNs).

Therefore every cluster member must be able to resolve hostnames of every other cluster member, its own hostname, as well as machines on which command line tools such as rabbitmqctl might be used.

Nodes will perform hostname resolution early on node boot. In container-based environments it is important that hostname resolution is ready before the container is started. For Kubernetes users, this means the DNS cache interval for CoreDNS to a value in the 5-10 second range.

Hostname resolution can use any of the standard OS-provided methods:

  • DNS records
  • Local host files (e.g. /etc/hosts)

In more restrictive environments, where DNS record or hosts file modification is restricted, impossible or undesired, Erlang VM can be configured to use alternative hostname resolution methods, such as an alternative DNS server, a local file, a non-standard hosts file location, or a mix of methods. Those methods can work in concert with the standard OS hostname resolution methods.

To use FQDNs, see RABBITMQ_USE_LONGNAME in the Configuration guide. See Node Names above.

Port Access

RabbitMQ nodes bind to ports (open server TCP sockets) in order to accept client and CLI tool connections. Other processes and tools such as SELinux may prevent RabbitMQ from binding to a port. When that happens, the node will fail to start.

CLI tools, client libraries and RabbitMQ nodes also open connections (client TCP sockets). Firewalls can prevent nodes and CLI tools from communicating with each other. The following ports are most relevant to inter-node communication in a cluster:

  • 4369: epmd, a helper discovery daemon used by RabbitMQ nodes and CLI tools
  • 6000 through 6500: used by RabbitMQ Stream replication
  • 25672: used for inter-node and CLI tools communication (Erlang distribution server port) and is allocated from a dynamic range (limited to a single port by default, computed as AMQP port + 20000). Unless external connections on these ports are really necessary (e.g. the cluster uses federation or CLI tools are used on machines outside the subnet), these ports should not be publicly exposed. See networking guide for details.
  • 35672-35682: used by CLI tools (Erlang distribution client ports) for communication with nodes and is allocated from a dynamic range (computed as server distribution port + 10000 through server distribution port + 10010).

It is possible to configure RabbitMQ to use different ports and specific network interfaces. See RabbitMQ Networking guide to learn more.

Nodes in a Cluster

What is Replicated?

All data/state required for the operation of a RabbitMQ broker is replicated across all nodes. An exception to this are message queues, which by default reside on one node, though they are visible and reachable from all nodes. To replicate queues across nodes in a cluster, use a queue type that supports replication. This topic is covered in the Quorum Queues guide.

Nodes are Equal Peers

Some distributed systems have leader and follower nodes. This is generally not true for RabbitMQ. All nodes in a RabbitMQ cluster are equal peers: there are no special nodes in RabbitMQ core. This topic becomes more nuanced when quorum queues and plugins are taken into consideration but for most intents and purposes, all cluster nodes should be considered equal.

Many CLI tool operations can be executed against any node. An HTTP API client can target any cluster node.

Individual plugins can designate (elect) certain nodes to be "special" for a period of time. For example, federation links are colocated on a particular cluster node. Should that node fail, the links will be restarted on a different node.

In older (long maintained) versions, RabbitMQ management plugin used a dedicated node for stats collection and aggregation.

RabbitMQ nodes and CLI tools (e.g. rabbitmqctl) use a cookie to determine whether they are allowed to communicate with each other. For two nodes to be able to communicate they must have the same shared secret called the Erlang cookie. The cookie is just a string of alphanumeric characters up to 255 characters in size. It is usually stored in a local file. The file must be only accessible to the owner (e.g. have UNIX permissions of 600 or similar). Every cluster node must have the same cookie.

If the file does not exist, Erlang VM will try to create one with a randomly generated value when the RabbitMQ server starts up. Using such generated cookie files are appropriate in development environments only. Since each node will generate its own value independently, this strategy is not really viable in a clustered environment.

Erlang cookie generation should be done at cluster deployment stage, ideally using automation and orchestration tools.

In distributed deployment

Linux, MacOS, *BSD

On UNIX systems, the cookie will be typically located in /var/lib/rabbitmq/.erlang.cookie (used by the server) and $HOME/.erlang.cookie (used by CLI tools). Note that since the value of $HOME varies from user to user, it's necessary to place a copy of the cookie file for each user that will be using the CLI tools. This applies to both non-privileged users and root.

RabbitMQ nodes will log its effective user's home directory location early on boot.

Community Docker Image and Kubernetes

Docker community RabbitMQ image uses RABBITMQ_ERLANG_COOKIE environment variable value to populate the cookie file.

Configuration management and container orchestration tools that use this image must make sure that every RabbitMQ node container in a cluster uses the same value.

In the context of Kubernetes, the value must be specified in the pod template specification of the stateful set. For instance, this can be seen in the RabbitMQ on Kubernetes examples repository.

Windows

On Windows, the cookie location depends on a few factors:

  • Whether the HOMEDRIVE and HOMEPATH environment variables are both set
  • Erlang version: prior to 20.2 (these are no longer supported by any maintained release series of RabbitMQ) or 20.2 and later
Erlang 20.2 or later

With Erlang versions starting with 20.2, the cookie file locations are:

  • %HOMEDRIVE%%HOMEPATH%\.erlang.cookie (usually C:\Users\%USERNAME%\.erlang.cookie for user %USERNAME%) if both the HOMEDRIVE and HOMEPATH environment variables are set
  • %USERPROFILE%\.erlang.cookie (usually C:\Users\%USERNAME%\.erlang.cookie) if HOMEDRIVE and HOMEPATH are not both set
  • For the RabbitMQ Windows service - %USERPROFILE%\.erlang.cookie (usually C:\WINDOWS\system32\config\systemprofile)

If the Windows service is used, the cookie should be copied from C:\Windows\system32\config\systemprofile\.erlang.cookie to the expected location for users running commands like rabbitmqctl.bat.

Overriding Using CLI and Runtime Command Line Arguments

As an alternative, the option "-setcookie <value>" can be added to RABBITMQ_SERVER_ADDITIONAL_ERL_ARGS environment variable value to override the cookie value used by a RabbitMQ node:

RABBITMQ_SERVER_ADDITIONAL_ERL_ARGS="-setcookie cookie-value"

CLI tools can take a cookie value using a command line flag:

rabbitmq-diagnostics status --erlang-cookie "cookie-value"

Both are the least secure options and generally not recommended.

When a node starts, it will log the home directory location of its effective user:

node           : rabbit@cdbf4de5f22d
home dir : /var/lib/rabbitmq

Unless any server directories were overridden, that's the directory where the cookie file will be looked for, and created by the node on first boot if it does not already exist.

In the example above, the cookie file location will be /var/lib/rabbitmq/.erlang.cookie.

Authentication Failures

When the cookie is misconfigured (for example, not identical), RabbitMQ nodes will log errors such as "Connection attempt from disallowed node", "", "Could not auto-cluster".

For example, when a CLI tool connects and tries to authenticate using a mismatching secret value:

2020-06-15 13:03:33 [error] <0.1187.0> ** Connection attempt from node 'rabbitmqcli-99391-rabbit@warp10' rejected. Invalid challenge reply. **

When a CLI tool such as rabbitmqctl fails to authenticate with RabbitMQ, the message usually says

* epmd reports node 'rabbit' running on port 25672
* TCP connection succeeded but Erlang distribution failed
* suggestion: hostname mismatch?
* suggestion: is the cookie set correctly?
* suggestion: is the Erlang distribution using TLS?

An incorrectly placed cookie file or cookie value mismatch are most common scenarios for such failures.

When a recent Erlang/OTP version is used, authentication failures contain more information and cookie mismatches can be identified better:

* connected to epmd (port 4369) on warp10
* epmd reports node 'rabbit' running on port 25672
* TCP connection succeeded but Erlang distribution failed

* Authentication failed (rejected by the remote node), please check the Erlang cookie

See the CLI Tools guide for more information.

Hostname Resolution

Since hostname resolution is a prerequisite for successful inter-node communication, starting with RabbitMQ 3.8.6, CLI tools provide two commands that help verify that hostname resolution on a node works as expected. The commands are not meant to replace dig and other specialised DNS tools but rather provide a way to perform most basic checks while taking Erlang runtime hostname resolver features into account.

The commands are covered in the Networking guide.

CLI Tools

Starting with version 3.8.6, rabbitmq-diagnostics includes a command that provides relevant information on the Erlang cookie file used by CLI tools:

rabbitmq-diagnostics erlang_cookie_sources

The command will report on the effective user, user home directory and the expected location of the cookie file:

Cookie File

Effective user: antares
Effective home directory: /home/cli-user
Cookie file path: /home/cli-user/.erlang.cookie
Cookie file exists? true
Cookie file type: regular
Cookie file access: read
Cookie file size: 20

Cookie CLI Switch

--erlang-cookie value set? false
--erlang-cookie value length: 0

Env variable (Deprecated)

RABBITMQ_ERLANG_COOKIE value set? false
RABBITMQ_ERLANG_COOKIE value length: 0

Node Counts and Quorum

Because several features (e.g. quorum queues, client tracking in MQTT) require a consensus between cluster members, odd numbers of cluster nodes are highly recommended: 1, 3, 5, 7 and so on.

Two node clusters are highly recommended against since it's impossible for cluster nodes to identify a majority and form a consensus in case of connectivity loss. For example, when the two nodes lose connectivity MQTT client connections won't be accepted, quorum queues would lose their availability, and so on.

From the consensus point of view, four or six node clusters would have the same availability characteristics as three and five node clusters.

The Quorum Queues guide covers this topic in more detail.

Clustering and Clients

Messaging Protocols

Assuming all cluster members are available, a messaging (AMQP 0-9-1, AMQP 1.0, MQTT, STOMP) client can connect to any node and perform any operation. Nodes will route operations to the quorum queue leader or queue leader replica transparently to clients.

With all supported messaging protocols a client is only connected to one node at a time.

In case of a node failure, clients should be able to reconnect to a different node, recover their topology and continue operation. For this reason, most client libraries accept a list of endpoints (hostnames or IP addresses) as a connection option. The list of hosts will be used during initial connection as well as connection recovery, if the client supports it. See documentation guides for individual clients to learn more.

With quorum queues and streams, clients will only be able to perform operations on queues that have a quorum of replicas online.

With classic mirrored queues, there are scenarios where it may not be possible for a client to transparently continue operations after connecting to a different node. They usually involve non-mirrored queues hosted on a failed node.

Stream Clients

RabbitMQ Stream protocol clients behave differently from messaging protocols clients: they are more cluster topology-aware. For publishing, they can connect to any node, and that node will forward all relevant operations to the node that hosts the leader replica of the stream.

However, stream consumers should connect to one of the nodes hosting the replicas of the target stream. The protocol includes a topology discovery operation, so well-behaved client libraries will select one of the suitable nodes. This won't be the case when a load balancer is used, however.

See Connecting to Streams to learn more.

Queue and Stream Leader Replica Placement

Every queue and srream in RabbitMQ has a primary replica. That replica is called the leader. All publishing operations on queues and streams go through the leader replica first and then are replicated to followers (secondary replicas). This is necessary to guarantee FIFO ordering of messages.

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

Queue leaders can be distributed between nodes using several strategies. Which strategy is used is controlled in three ways, namely, using the x-queue-master-locator optional queue argument, setting the queue-master-locator policy key or by defining the queue_master_locator key in the configuration file.

There are two options available:

  • balanced, the default strategy, uses the data on how many replicas peer nodes host, when there are relatively few (say, no more than 1000) queues in the cluster; it falls back to a more efficient strategy of picking a random node when there are many queues
  • client-local will always pick the node the client is connected to

The following example sets the queue_leader_locator setting in rabbitmq.conf to its default value:

queue_leader_locator = balanced

The client-provided queue argument takes presedence when both are used.

Note that all Raft-based features, namely quorum queues and streams, use this value as a suggestion. Raft leader election algorithm involves a degree of randomness, therefore the selected recommended node will have a replica placed on it but it will not always be the leader replica.

Clustering and Observability

Client connections, channels and queues will be distributed across cluster nodes. Operators need to be able to inspect and monitor such resources across all cluster nodes.

RabbitMQ CLI tools such as rabbitmq-diagnostics and rabbitmqctl provide commands that inspect resources and cluster-wide state. Some commands focus on the state of a single node (e.g. rabbitmq-diagnostics environment and rabbitmq-diagnostics status), others inspect cluster-wide state. Some examples of the latter include rabbitmqctl list_connections, rabbitmqctl list_mqtt_connections, rabbitmqctl list_stomp_connections, rabbitmqctl list_users, rabbitmqctl list_vhosts and so on.

Such "cluster-wide" commands will often contact one node first, discover cluster members and contact them all to retrieve and combine their respective state. For example, rabbitmqctl list_connections will contact all nodes, retrieve their AMQP 0-9-1 and AMQP 1.0 connections, and display them all to the user. The user doesn't have to manually contact all nodes. Assuming a non-changing state of the cluster (e.g. no connections are closed or opened), two CLI commands executed against two different nodes one after another will produce identical or semantically identical results. "Node-local" commands, however, will not produce identical results since two nodes rarely have identical state: at the very least their node names will be different!

Management UI works similarly: a node that has to respond to an HTTP API request will fan out to other cluster members and aggregate their responses. In a cluster with multiple nodes that have management plugin enabled, the operator can use any node to access management UI. The same goes for monitoring tools that use the HTTP API to collect data about the state of the cluster. There is no need to issue a request to every cluster node in turn.

Node Failure Handling

RabbitMQ brokers tolerate the failure of individual nodes. Nodes can be started and stopped at will, as long as they can contact a cluster member node known at the time of shutdown.

Quorum queue allows queue contents to be replicated across multiple cluster nodes with parallel replication and a predictable leader election and data safety behavior as long as a majority of replicas are online.

Non-replicated classic queues can also be used in clusters. Non-mirrored queue behaviour in case of node failure depends on queue durability.

RabbitMQ clustering has several modes of dealing with network partitions, primarily consistency oriented. Clustering is meant to be used across LAN. It is not recommended to run clusters that span WAN. The Shovel or Federation plugins are better solutions for connecting brokers across a WAN. Note that Shovel and Federation are not equivalent to clustering.

Metrics and Statistics

Every node stores and aggregates its own metrics and stats, and provides an API for other nodes to access it. Some stats are cluster-wide, others are specific to individual nodes. Node that responds to an HTTP API request contacts its peers to retrieve their data and then produces an aggregated result.

In older (long unmaintained) versions RabbitMQ management plugin used a dedicated node for stats collection and aggregation.

Clustering Transcript with rabbitmqctl

The following several sections provide a transcript of manually setting up and manipulating a RabbitMQ cluster across three machines: rabbit1, rabbit2, rabbit3. It is recommended that the example is studied before more automation-friendly cluster formation options are used.

We assume that the user is logged into all three machines, that RabbitMQ has been installed on the machines, and that the rabbitmq-server and rabbitmqctl scripts are in the user's PATH.

This transcript can be modified to run on a single host, as explained more details below.

Starting Independent Nodes

Clusters are set up by re-configuring existing RabbitMQ nodes into a cluster configuration. Hence the first step is to start RabbitMQ on all nodes in the normal way:

# on rabbit1
rabbitmq-server -detached
# on rabbit2
rabbitmq-server -detached
# on rabbit3
rabbitmq-server -detached

This creates three independent RabbitMQ brokers, one on each node, as confirmed by the cluster_status command:

# on rabbit1
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit1 ...
# => [{nodes,[{disc,[rabbit@rabbit1]}]},{running_nodes,[rabbit@rabbit1]}]
# => ...done.

# on rabbit2
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit2 ...
# => [{nodes,[{disc,[rabbit@rabbit2]}]},{running_nodes,[rabbit@rabbit2]}]
# => ...done.

# on rabbit3
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit3 ...
# => [{nodes,[{disc,[rabbit@rabbit3]}]},{running_nodes,[rabbit@rabbit3]}]
# => ...done.

The node name of a RabbitMQ broker started from the rabbitmq-server shell script is rabbit@shorthostname, where the short node name is lower-case (as in rabbit@rabbit1, above). On Windows, if rabbitmq-server.bat batch file is used, the short node name is upper-case (as in rabbit@RABBIT1). When you type node names, case matters, and these strings must match exactly.

Creating a Cluster

In order to link up our three nodes in a cluster, we tell two of the nodes, say rabbit@rabbit2 and rabbit@rabbit3, to join the cluster of the third, say rabbit@rabbit1. Prior to that both newly joining members must be reset.

We first join rabbit@rabbit2 in a cluster with rabbit@rabbit1. To do that, on rabbit@rabbit2 we stop the RabbitMQ application and join the rabbit@rabbit1 cluster, then restart the RabbitMQ application. Note that a node must be reset before it can join an existing cluster. Resetting the node removes all resources and data that were previously present on that node. This means that a node cannot be made a member of a cluster and keep its existing data at the same time. When that's desired, using the Blue/Green deployment strategy or backup and restore are the available options.

# on rabbit2
rabbitmqctl stop_app
# => Stopping node rabbit@rabbit2 ...done.

rabbitmqctl reset
# => Resetting node rabbit@rabbit2 ...

rabbitmqctl join_cluster rabbit@rabbit1
# => Clustering node rabbit@rabbit2 with [rabbit@rabbit1] ...done.

rabbitmqctl start_app
# => Starting node rabbit@rabbit2 ...done.

We can see that the two nodes are joined in a cluster by running the cluster_status command on either of the nodes:

# on rabbit1
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit1 ...
# => [{nodes,[{disc,[rabbit@rabbit1,rabbit@rabbit2]}]},
# => {running_nodes,[rabbit@rabbit2,rabbit@rabbit1]}]
# => ...done.

# on rabbit2
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit2 ...
# => [{nodes,[{disc,[rabbit@rabbit1,rabbit@rabbit2]}]},
# => {running_nodes,[rabbit@rabbit1,rabbit@rabbit2]}]
# => ...done.

Now we join rabbit@rabbit3 to the same cluster. The steps are identical to the ones above, except this time we'll cluster to rabbit2 to demonstrate that the node chosen to cluster to does not matter - it is enough to provide one online node and the node will be clustered to the cluster that the specified node belongs to.

# on rabbit3
rabbitmqctl stop_app
# => Stopping node rabbit@rabbit3 ...done.

# on rabbit3
rabbitmqctl reset
# => Resetting node rabbit@rabbit3 ...

rabbitmqctl join_cluster rabbit@rabbit2
# => Clustering node rabbit@rabbit3 with rabbit@rabbit2 ...done.

rabbitmqctl start_app
# => Starting node rabbit@rabbit3 ...done.

We can see that the three nodes are joined in a cluster by running the cluster_status command on any of the nodes:

# on rabbit1
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit1 ...
# => [{nodes,[{disc,[rabbit@rabbit1,rabbit@rabbit2,rabbit@rabbit3]}]},
# => {running_nodes,[rabbit@rabbit3,rabbit@rabbit2,rabbit@rabbit1]}]
# => ...done.

# on rabbit2
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit2 ...
# => [{nodes,[{disc,[rabbit@rabbit1,rabbit@rabbit2,rabbit@rabbit3]}]},
# => {running_nodes,[rabbit@rabbit3,rabbit@rabbit1,rabbit@rabbit2]}]
# => ...done.

# on rabbit3
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit3 ...
# => [{nodes,[{disc,[rabbit@rabbit3,rabbit@rabbit2,rabbit@rabbit1]}]},
# => {running_nodes,[rabbit@rabbit2,rabbit@rabbit1,rabbit@rabbit3]}]
# => ...done.

By following the above steps we can add new nodes to the cluster at any time, while the cluster is running.

Restarting Cluster Nodes

Nodes that have been joined to a cluster can be stopped at any time. They can also fail or be terminated by the OS.

In general, if the majority of nodes is still online after a node is stopped, this does not affect the rest of the cluster, although client connection distribution, queue replica placement, and load distribution of the cluster will change.

Schema Syncing from Online Peers

A restarted node will sync the schema and other information from its peers on boot. Before this process completes, the node won't be fully started and functional.

It is therefore important to understand the process node go through when they are stopped and restarted.

A stopping node picks an online cluster member (only disc nodes will be considered) to sync with after restart. Upon restart the node will try to contact that peer 10 times by default, with 30 second response timeouts.

In case the peer becomes available in that time interval, the node successfully starts, syncs what it needs from the peer and keeps going.

If the peer does not become available, the restarted node will give up and voluntarily stop. Such condition can be identified by the timeout (timeout_waiting_for_tables) warning messages in the logs that eventually lead to node startup failure:

2020-07-27 21:10:51.361 [warning] <0.269.0> Error while waiting for Mnesia tables: {timeout_waiting_for_tables,[rabbit@node2,rabbit@node1],[rabbit_durable_queue]}
2020-07-27 21:10:51.361 [info] <0.269.0> Waiting for Mnesia tables for 30000 ms, 1 retries left
2020-07-27 21:11:21.362 [warning] <0.269.0> Error while waiting for Mnesia tables: {timeout_waiting_for_tables,[rabbit@node2,rabbit@node1],[rabbit_durable_queue]}
2020-07-27 21:11:21.362 [info] <0.269.0> Waiting for Mnesia tables for 30000 ms, 0 retries left
2020-07-27 21:15:51.380 [info] <0.269.0> Waiting for Mnesia tables for 30000 ms, 1 retries left
2020-07-27 21:16:21.381 [warning] <0.269.0> Error while waiting for Mnesia tables: {timeout_waiting_for_tables,[rabbit@node2,rabbit@node1],[rabbit_user,rabbit_user_permission, …]}
2020-07-27 21:16:21.381 [info] <0.269.0> Waiting for Mnesia tables for 30000 ms, 0 retries left
2020-07-27 21:16:51.393 [info] <0.44.0> Application mnesia exited with reason: stopped
2020-07-27 21:16:51.397 [error] <0.269.0> BOOT FAILED
2020-07-27 21:16:51.397 [error] <0.269.0> ===========
2020-07-27 21:16:51.397 [error] <0.269.0> Timeout contacting cluster nodes: [rabbit@node1].

When a node has no online peers during shutdown, it will start without attempts to sync with any known peers. It does not start as a standalone node, however, and peers will be able to rejoin it.

When the entire cluster is brought down therefore, the last node to go down is the only one that didn't have any running peers at the time of shutdown. That node can start without contacting any peers first. Since nodes will try to contact a known peer for up to 5 minutes (by default), nodes can be restarted in any order in that period of time. In this case they will rejoin each other one by one successfully. This window of time can be adjusted using two configuration settings:

# wait for 60 seconds instead of 30
mnesia_table_loading_retry_timeout = 60000

# retry 15 times instead of 10
mnesia_table_loading_retry_limit = 15

By adjusting these settings and tweaking the time window in which known peer has to come back it is possible to account for cluster-wide redeployment scenarios that can be longer than 5 minutes to complete.

During upgrades, sometimes the last node to stop must be the first node to be started after the upgrade. That node will be designated to perform a cluster-wide schema migration that other nodes can sync from and apply when they rejoin.

Restarts and Health Checks (Readiness Probes)

In some environments, node restarts are controlled with a designated health check. The checks verify that one node has started and the deployment process can proceed to the next one. If the check does not pass, the deployment of the node is considered to be incomplete and the deployment process will typically wait and retry for a period of time. One popular example of such environment is Kubernetes where an operator-defined readiness probe can prevent a deployment from proceeding when the OrderedReady pod management policy is used. Deployments that use the Parallel pod management policy will not be affected but must worry about the natural race condition during initial cluster formation.

Given the peer syncing behavior described above, such a health check can prevent a cluster-wide restart from completing in time. Checks that explicitly or implicitly assume a fully booted node that's rejoined its cluster peers will fail and block further node deployments.

Most health check, even relatively basic ones, implicitly assume that the node has finished booting. They are not suitable for nodes that are awaiting schema table sync from a peer.

One very common example of such check is

# will exit with an error for the nodes that are currently waiting for
# a peer to sync schema tables from
rabbitmq-diagnostics check_running

One health check that does not expect a node to be fully booted and have schema tables synced is

# a very basic check that will succeed for the nodes that are currently waiting for
# a peer to sync schema from
rabbitmq-diagnostics ping

This basic check would allow the deployment to proceed and the nodes to eventually rejoin each other, assuming they are compatible.

Hostname Changes Between Restarts

A node rejoining after a node name or host name change can start as a blank node if its data directory path changes as a result. Such nodes will fail to rejoin the cluster. While the node is offline, its peers can be reset or started with a blank data directory. In that case the recovering node will fail to rejoin its peer as well since internal data store cluster identity would no longer match.

Consider the following scenario:

  1. A cluster of 3 nodes, A, B and C is formed
  2. Node A is shut down
  3. Node B is reset
  4. Node A is started
  5. Node A tries to rejoin B but B's cluster identity has changed
  6. Node B doesn't recognise A as a known cluster member because it's been reset

in this case node B will reject the clustering attempt from A with an appropriate error message in the log:

Node 'rabbit@node1.local' thinks it's clustered with node 'rabbit@node2.local', but 'rabbit@node2.local' disagrees

In this case B can be reset again and then will be able to join A, or A can be reset and will successfully join B.

Cluster Node Restart Example

The below example uses CLI tools to shut down the nodes rabbit@rabbit1 and rabbit@rabbit3 and check on the cluster status at each step:

# on rabbit1
rabbitmqctl stop
# => Stopping and halting node rabbit@rabbit1 ...done.

# on rabbit2
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit2 ...
# => [{nodes,[{disc,[rabbit@rabbit1,rabbit@rabbit2,rabbit@rabbit3]}]},
# => {running_nodes,[rabbit@rabbit3,rabbit@rabbit2]}]
# => ...done.

# on rabbit3
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit3 ...
# => [{nodes,[{disc,[rabbit@rabbit1,rabbit@rabbit2,rabbit@rabbit3]}]},
# => {running_nodes,[rabbit@rabbit2,rabbit@rabbit3]}]
# => ...done.

# on rabbit3
rabbitmqctl stop
# => Stopping and halting node rabbit@rabbit3 ...done.

# on rabbit2
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit2 ...
# => [{nodes,[{disc,[rabbit@rabbit1,rabbit@rabbit2,rabbit@rabbit3]}]},
# => {running_nodes,[rabbit@rabbit2]}]
# => ...done.

In the below example, the nodes are started back, checking on the cluster status as we go along:

# on rabbit1
rabbitmq-server -detached
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit1 ...
# => [{nodes,[{disc,[rabbit@rabbit1,rabbit@rabbit2,rabbit@rabbit3]}]},
# => {running_nodes,[rabbit@rabbit2,rabbit@rabbit1]}]
# => ...done.

# on rabbit2
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit2 ...
# => [{nodes,[{disc,[rabbit@rabbit1,rabbit@rabbit2,rabbit@rabbit3]}]},
# => {running_nodes,[rabbit@rabbit1,rabbit@rabbit2]}]
# => ...done.

# on rabbit3
rabbitmq-server -detached

# on rabbit1
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit1 ...
# => [{nodes,[{disc,[rabbit@rabbit1,rabbit@rabbit2,rabbit@rabbit3]}]},
# => {running_nodes,[rabbit@rabbit2,rabbit@rabbit1,rabbit@rabbit3]}]
# => ...done.

# on rabbit2
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit2 ...
# => [{nodes,[{disc,[rabbit@rabbit1,rabbit@rabbit2,rabbit@rabbit3]}]},
# => {running_nodes,[rabbit@rabbit1,rabbit@rabbit2,rabbit@rabbit3]}]
# => ...done.

# on rabbit3
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit3 ...
# => [{nodes,[{disc,[rabbit@rabbit1,rabbit@rabbit2,rabbit@rabbit3]}]},
# => {running_nodes,[rabbit@rabbit2,rabbit@rabbit1,rabbit@rabbit3]}]
# => ...done.

Forcing Node Boot in Case of Unavailable Peers

In some cases the last node to go offline cannot be brought back up. It can be removed from the cluster using the forget_cluster_node rabbitmqctl command.

Alternatively force_boot rabbitmqctl command can be used on a node to make it boot without trying to sync with any peers (as if they were last to shut down). This is usually only necessary if the last node to shut down or a set of nodes will never be brought back online.

Breaking Up a Cluster

Sometimes it is necessary to remove a node from a cluster. The operator has to do this explicitly using a rabbitmqctl command.

Some peer discovery mechanisms support node health checks and forced removal of nodes not known to the discovery backend. That feature is opt-in (deactivated by default).

We first remove rabbit@rabbit3 from the cluster, returning it to independent operation. To do that, on rabbit@rabbit3 we stop the RabbitMQ application, reset the node, and restart the RabbitMQ application.

# on rabbit3
rabbitmqctl stop_app
# => Stopping node rabbit@rabbit3 ...done.

rabbitmqctl reset
# => Resetting node rabbit@rabbit3 ...done.
rabbitmqctl start_app
# => Starting node rabbit@rabbit3 ...done.

Note that it would have been equally valid to list rabbit@rabbit3 as a node.

Running the cluster_status command on the nodes confirms that rabbit@rabbit3 now is no longer part of the cluster and operates independently:

# on rabbit1
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit1 ...
# => [{nodes,[{disc,[rabbit@rabbit1,rabbit@rabbit2]}]},
# => {running_nodes,[rabbit@rabbit2,rabbit@rabbit1]}]
# => ...done.

# on rabbit2
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit2 ...
# => [{nodes,[{disc,[rabbit@rabbit1,rabbit@rabbit2]}]},
# => {running_nodes,[rabbit@rabbit1,rabbit@rabbit2]}]
# => ...done.

# on rabbit3
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit3 ...
# => [{nodes,[{disc,[rabbit@rabbit3]}]},{running_nodes,[rabbit@rabbit3]}]
# => ...done.

We can also remove nodes remotely. This is useful, for example, when having to deal with an unresponsive node. We can for example remove rabbit@rabbit1 from rabbit@rabbit2.

# on rabbit1
rabbitmqctl stop_app
# => Stopping node rabbit@rabbit1 ...done.

# on rabbit2
rabbitmqctl forget_cluster_node rabbit@rabbit1
# => Removing node rabbit@rabbit1 from cluster ...
# => ...done.

Note that rabbit1 still thinks it's clustered with rabbit2, and trying to start it will result in an error. We will need to reset it to be able to start it again.

# on rabbit1
rabbitmqctl start_app
# => Starting node rabbit@rabbit1 ...
# => Error: inconsistent_cluster: Node rabbit@rabbit1 thinks it's clustered with node rabbit@rabbit2, but rabbit@rabbit2 disagrees

rabbitmqctl reset
# => Resetting node rabbit@rabbit1 ...done.

rabbitmqctl start_app
# => Starting node rabbit@rabbit1 ...
# => ...done.

The cluster_status command now shows all three nodes operating as independent RabbitMQ brokers:

# on rabbit1
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit1 ...
# => [{nodes,[{disc,[rabbit@rabbit1]}]},{running_nodes,[rabbit@rabbit1]}]
# => ...done.

# on rabbit2
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit2 ...
# => [{nodes,[{disc,[rabbit@rabbit2]}]},{running_nodes,[rabbit@rabbit2]}]
# => ...done.

# on rabbit3
rabbitmqctl cluster_status
# => Cluster status of node rabbit@rabbit3 ...
# => [{nodes,[{disc,[rabbit@rabbit3]}]},{running_nodes,[rabbit@rabbit3]}]
# => ...done.

Note that rabbit@rabbit2 retains the residual state of the cluster, whereas rabbit@rabbit1 and rabbit@rabbit3 are freshly initialised RabbitMQ brokers. If we want to re-initialise rabbit@rabbit2 we follow the same steps as for the other nodes:

# on rabbit2
rabbitmqctl stop_app
# => Stopping node rabbit@rabbit2 ...done.
rabbitmqctl reset
# => Resetting node rabbit@rabbit2 ...done.
rabbitmqctl start_app
# => Starting node rabbit@rabbit2 ...done.

Besides rabbitmqctl forget_cluster_node and the automatic cleanup of unknown nodes by some peer discovery plugins, there are no scenarios in which a RabbitMQ node will permanently remove its peer node from a cluster.

How to Reset a Node

Sometimes it may be necessary to reset a node (wipe all of its data) and later make it rejoin the cluster. Generally speaking, there are two possible scenarios: when the node is running, and when the node cannot start or won't respond to CLI tool commands e.g. due to an issue such as ERL-430.

Resetting a node will delete all of its data, cluster membership information, configured runtime parameters, users, virtual hosts and any other node data. It will also permanently remove the node from its cluster.

To reset a running and responsive node, first stop RabbitMQ on it using rabbitmqctl stop_app and then reset it using rabbitmqctl reset:

# on rabbit1
rabbitmqctl stop_app
# => Stopping node rabbit@rabbit1 ...done.
rabbitmqctl reset
# => Resetting node rabbit@rabbit1 ...done.

In case of a non-responsive node, it must be stopped first using any means necessary. For nodes that fail to start this is already the case. Then override the node's data directory location or [re]move the existing data store. This will make the node start as a blank one. It will have to be instructed to rejoin its original cluster, if any.

A node that's been reset and rejoined its original cluster will sync all virtual hosts, users, permissions and topology (queues, exchanges, bindings), runtime parameters and policies. Quorum queue contents will be replicated if the node will be selected to host a replica. Non-replicated queue contents on a reset node will be lost.

Upgrading clusters

You can find instructions for upgrading a cluster in the upgrade guide.

A Cluster on a Single Machine

Under some circumstances it can be useful to run a cluster of RabbitMQ nodes on a single machine. This would typically be useful for experimenting with clustering on a desktop or laptop without the overhead of starting several virtual machines for the cluster.

In order to run multiple RabbitMQ nodes on a single machine, it is necessary to make sure the nodes have distinct node names, data store locations, log file locations, and bind to different ports, including those used by plugins. See RABBITMQ_NODENAME, RABBITMQ_NODE_PORT, and RABBITMQ_DIST_PORT in the Configuration guide, as well as RABBITMQ_MNESIA_DIR, RABBITMQ_CONFIG_FILE, and RABBITMQ_LOG_BASE in the File and Directory Locations guide.

You can start multiple nodes on the same host manually by repeated invocation of rabbitmq-server ( rabbitmq-server.bat on Windows). For example:

RABBITMQ_NODE_PORT=5672 RABBITMQ_NODENAME=rabbit rabbitmq-server -detached
RABBITMQ_NODE_PORT=5673 RABBITMQ_NODENAME=hare rabbitmq-server -detached
rabbitmqctl -n hare stop_app
rabbitmqctl -n hare join_cluster rabbit@`hostname -s`
rabbitmqctl -n hare start_app

will set up a two node cluster, both nodes as disc nodes. Note that if the node listens on any ports other than AMQP 0-9-1 and AMQP 1.0 ones, those must be configured to avoid a collision as well. This can be done via command line:

RABBITMQ_NODE_PORT=5672 RABBITMQ_SERVER_START_ARGS="-rabbitmq_management listener [{port,15672}]" RABBITMQ_NODENAME=rabbit rabbitmq-server -detached
RABBITMQ_NODE_PORT=5673 RABBITMQ_SERVER_START_ARGS="-rabbitmq_management listener [{port,15673}]" RABBITMQ_NODENAME=hare rabbitmq-server -detached

will start two nodes (which can then be clustered) when the management plugin is installed.

Hostname Changes

RabbitMQ nodes use hostnames to communicate with each other. Therefore, all node names must be able to resolve names of all cluster peers. This is also true for tools such as rabbitmqctl.

In addition to that, by default RabbitMQ names the database directory using the current hostname of the system. If the hostname changes, a new empty database is created. To avoid data loss it's crucial to set up a fixed and resolvable hostname.

Whenever the hostname changes RabbitMQ node must be restarted.

A similar effect can be achieved by using rabbit@localhost as the broker nodename. The impact of this solution is that clustering will not work because the chosen hostname does not resolve to a routable address from the remote hosts. The rabbitmqctl command fails when invoked from a remote host. A better solution is to use DNS, for example, Amazon Route 53 if running on EC2. If you want to use the full hostname for your nodename (RabbitMQ defaults to the short name), and that full hostname is resolvable using DNS, you may want to investigate setting the environment variable RABBITMQ_USE_LONGNAME=true.

See the section on hostname resolution for more information.

Firewalled Nodes

Nodes can have a firewall enabled on them. In such case, traffic on certain ports must be allowed by the firewall in both directions, or nodes won't be able to join each other and perform all the operations they expect to be available on cluster peers.

Learn more in the section on ports above and dedicated RabbitMQ Networking guide.

Erlang Versions Across the Cluster

All nodes in a cluster are highly recommended to run the same major version of Erlang: 26.2.0 and 26.1.2 can be mixed but 25.3.2.8 and 26.2.0 can potentially introduce breaking changes in inter-node communication protocols. While such breaking changes are rare, they are possible.

Incompatibilities between patch releases of Erlang/OTP versions are very rare.

Connecting to Clusters from Clients

A client can connect as normal to any node within a cluster. If that node should fail, and the rest of the cluster survives, then the client should notice the closed connection, and should be able to reconnect to some surviving member of the cluster.

Many clients support lists of hostnames that will be tried in order at connection time.

Generally it is not recommended to hardcode IP addresses into client applications: this introduces inflexibility and will require client applications to be edited, recompiled and redeployed should the configuration of the cluster change or the number of nodes in the cluster change.

Instead, consider a more abstracted approach: this could be a dynamic DNS service which has a very short TTL configuration, or a plain TCP load balancer, or a combination of them.

In general, this aspect of managing the connection to nodes within a cluster is beyond the scope of this guide, and we recommend the use of other technologies designed specifically to address these problems.