This plugin adds a consistent-hash exchange type to RabbitMQ. This exchange type uses consistent hashing (intro blog posts: one, two, three) to distribute messages between the bound queues. It is recommended to get a basic understanding of the concept before evaluating this plugin and its alternatives.

rabbitmq-sharding is another plugin that provides a way to partition a stream of messages among a set of consumers while trading off total stream ordering for processing parallelism.

In various scenarios it may be desired to ensure that messages sent to an exchange are reasonably uniformly distributed across a number of queues based on the routing key of the message, a nominated header, or a message property. Technically this can be accomplished using a direct or topic exchange, binding queues to that exchange and then publishing messages to that exchange that match the various binding keys.

However, arranging things this way can be problematic:

1. It is difficult to ensure that all queues bound to the exchange will receive a (roughly) equal number of messages (distribution uniformity) without baking in to the publishers quite a lot of knowledge about the number of queues and their bindings.

2. When the number of queues changes, it is not easy to ensure that the new topology still distributes messages between the different queues roughly evenly.

Consistent Hashing is a hashing technique whereby each bucket appears at multiple points throughout the hash space, and the bucket selected is the nearest higher (or lower, it doesn't matter, provided it's consistent) bucket to the computed hash (and the hash space wraps around). The effect of this is that when a new bucket is added or an existing bucket removed, only a very few hashes change which bucket they are routed to.

The purpose of this exchange type is to help developers achieve a reasonably even message flow distribution between a number of queues.

This plugin ships with RabbitMQ.

This plugin ships with RabbitMQ. Like all other RabbitMQ plugins, it has to be enabled before it can be used:

rabbitmq-plugins enable rabbitmq_consistent_hash_exchange

The exchange type is "x-consistent-hash".

In the case of Consistent Hashing as an exchange type, the hash is calculated from a message property (most commonly the routing key).

When a queue is bound to this exchange, it is assigned one or more partitions on the consistent hashing ring depending on its binding weight (covered below).

For every property hash (e.g. routing key), a hash position computed and a corresponding hash ring partition is picked. That partition corresponds to a bound queue, and the message is routed to that queue.

Assuming a reasonably even routing key distribution of inbound messages, routed messages should be reasonably evenly distributed across all ring partitions, and thus queues according to their binding weights.

This exchange type assumes a single binding between a queue and an exchange. This will be enforced in the code: when multiple bindings are created, only the first one will actually update the ring.

This limitation makes most semantic sense: the purpose is to achieve a reasonably even message flow distribution between queues.

When a queue is bound to a Consistent Hash exchange, the binding key is a number-as-a-string which indicates the binding weight: the number of buckets (sections of the range) that will be associated with the target queue.

In most environments, using one bucket per binding (and thus queue) is highly recommended as it is the simplest way to achieve reasonably even balancing.

The hashing distributes routing keys among queues, not message payloads among queues; all messages with the same routing key will go the same queue. So, if you wish for queue A to receive twice as many routing keys routed to it than are routed to queue B, then you bind the queue A with a binding key of twice the number (as a string -- binding keys are always strings) of the binding key of the binding to queue B. Note this is only the case if your routing keys are evenly distributed in the hash space. If, for example, only two distinct routing keys are used on all the messages, there's a chance both keys will route (consistently!) to the same queue, even though other queues have higher values in their binding key. With a larger set of routing keys used, the statistical distribution of routing keys approaches the ratios of the binding keys.

Each message gets delivered to at most one queue. On average, a message gets delivered to exactly one queue. Concurrent binding changes and queue primary replica failures can temporarily affect this but over the long term, assuming equal weights of every binding, the distribution should be roughly even.

Consistent hashing ring is stored in memory and will be re-populated from exchange bindings when the node boots. Relative positioning of queues on the ring is not guaranteed to be the same between restarts. In practice this means that after a restart, all queues will still receive roughly the same number of messages routed to them (assuming routing key distribution does not change) but a given routing key now may route to a different queue.

In other words, this exchange type provides consistent message distribution between queues but cannot guarantee stable queue routing locality for messages with a given routing key.

In the below example the queues q0 and q1 get bound each with the weight of 1 in the hash space to the exchange e which means they'll each get roughly the same number of routing keys. The queues q2 and q3 however, get 2 buckets each (their weight is 2) which means they'll each get roughly the same number of routing keys too, but that will be approximately twice as many as q0 and q1.

Note the routing_keys in the bindings are numbers-as-strings. This is because AMQP 0-9-1 specifies the routing_key field must be a string.

The example uses low weight values intentionally.

Higher values will reduce throughput of the exchange, primarily for workloads that experience a high binding churn (queues are bound to and unbound from a consistent hash exchange frequently).

Equal weights of 1 for all bindings are recommended (and sufficient for most use cases).

The example then publishes 100,000 messages to our exchange with random routing keys, the queues will get their share of messages roughly equal to the binding keys ratios. After this has completed, message distribution between queues can be inspected using RabbitMQ's management UI and rabbitmqctl list_queues.

It is important to ensure that the messages being published to the exchange have varying routing keys: if a very small set of routing keys are being used then there's a possibility of messages not being evenly distributed between the bound queues. With a large number of bound queues some queues may get no messages routed to them at all.

If pseudo-random or unique values such as client/session/request identifiers are used for routing keys (or another property used for hashing) then reasonably uniform distribution should be observed.

Executable versions of some of the code examples can be found under ./examples.

This version of the example uses Pika, the most widely used Python client for RabbitMQ:

#!/usr/bin/env python

import pika
import time

conn = pika.BlockingConnection(pika.ConnectionParameters(host='localhost'))
ch   = conn.channel()

ch.exchange_declare(exchange="e", exchange_type="x-consistent-hash", durable=True)

for q in ["q1", "q2", "q3", "q4"]:
    ch.queue_declare(queue=q, durable=True)
    ch.queue_purge(queue=q)

for q in ["q1", "q2"]:
    ch.queue_bind(exchange="e", queue=q, routing_key="1")

for q in ["q3", "q4"]:
    ch.queue_bind(exchange="e", queue=q, routing_key="2")

n = 100000

for rk in list(map(lambda s: str(s), range(0, n))):
    ch.basic_publish(exchange="e", routing_key=rk, body="")
print("Done publishing.")

print("Waiting for routing to finish...")
# in order to keep this example simpler and focused,
# wait for a few seconds instead of using publisher confirms and waiting for those
time.sleep(5)

print("Done.")
conn.close()

Below is a version of the example that uses the official RabbitMQ Java client:

package com.rabbitmq.examples;

import com.rabbitmq.client.*;

import java.io.IOException;
import java.util.Arrays;
import java.util.concurrent.TimeoutException;

public class ConsistentHashExchangeExample1 {
  private static String CONSISTENT_HASH_EXCHANGE_TYPE = "x-consistent-hash";

  public static void main(String[] argv) throws IOException, TimeoutException, InterruptedException {
    ConnectionFactory cf = new ConnectionFactory();
    Connection conn = cf.newConnection();
    Channel ch = conn.createChannel();

    for (String q : Arrays.asList("q1", "q2", "q3", "q4")) {
      ch.queueDeclare(q, true, false, false, null);
      ch.queuePurge(q);
    }

    ch.exchangeDeclare("e1", CONSISTENT_HASH_EXCHANGE_TYPE, true, false, null);

    for (String q : Arrays.asList("q1", "q2")) {
      ch.queueBind(q, "e1", "1");
    }

    for (String q : Arrays.asList("q3", "q4")) {
      ch.queueBind(q, "e1", "2");
    }

    ch.confirmSelect();

    AMQP.BasicProperties.Builder bldr = new AMQP.BasicProperties.Builder();
    for (int i = 0; i < 100000; i++) {
      ch.basicPublish("e1", String.valueOf(i), bldr.build(), "".getBytes("UTF-8"));
    }

    ch.waitForConfirmsOrDie(10000);

    System.out.println("Done publishing!");
    System.out.println("Evaluating results...");
    // wait for one stats emission interval so that queue counters
    // are up-to-date in the management UI
    Thread.sleep(5);

    System.out.println("Done.");
    conn.close();
  }
}

Below is a version that uses Bunny, the most widely used Ruby client for RabbitMQ:

#!/usr/bin/env ruby

require 'bunny'

conn = Bunny.new
conn.start

ch = conn.create_channel
ch.confirm_select

q1 = ch.queue("q1", durable: true)
q2 = ch.queue("q2", durable: true)
q3 = ch.queue("q3", durable: true)
q4 = ch.queue("q4", durable: true)

[q1, q2, q3, q4]. each(&:purge)

x  = ch.exchange("chx", type: "x-consistent-hash", durable: true)

[q1, q2].each { |q| q.bind(x, routing_key: "1") }
[q3, q4].each { |q| q.bind(x, routing_key: "2") }

n = 100_000
n.times do |i|
  x.publish(i.to_s, routing_key: i.to_s)
end

ch.wait_for_confirms
puts "Done publishing!"

# wait for queue stats to be emitted so that management UI numbers
# are up-to-date
sleep 5
conn.close
puts "Done"

Below is a version of the example that uses the RabbitMQ Erlang client:

-include_lib("amqp_client/include/amqp_client.hrl").

test() ->
    {ok, Conn} = amqp_connection:start(#amqp_params_network{}),
    {ok, Chan} = amqp_connection:open_channel(Conn),
    Queues = [<<"q0">>, <<"q1">>, <<"q2">>, <<"q3">>],
    amqp_channel:call(Chan,
                  #'exchange.declare'{
                    exchange = <<"e">>, type = <<"x-consistent-hash">>
                  }),
    [amqp_channel:call(Chan, #'queue.declare'{queue = Q}) || Q <- Queues],
    [amqp_channel:call(Chan, #'queue.bind'{queue = Q,
                                           exchange = <<"e">>,
                                           routing_key = <<"1">>})
        || Q <- [<<"q0">>, <<"q1">>]],
    [amqp_channel:call(Chan, #'queue.bind' {queue = Q,
                                            exchange = <<"e">>,
                                            routing_key = <<"2">>})
        || Q <- [<<"q2">>, <<"q3">>]],
    RK = list_to_binary(integer_to_list(random:uniform(1000000))),
    Msg = #amqp_msg{props = #'P_basic'{}, payload = <<>>},
    [amqp_channel:call(Chan,
                   #'basic.publish'{
                     exchange = <<"e">>,
                     routing_key = RK
                   }, Msg) || _ <- lists:seq(1, 100000)],
amqp_connection:close(Conn),
ok.

Under most circumstances the routing key is a good choice for something to hash. However, in some cases it is necessary to use the routing key for some other purpose (for example with more complex routing involving exchange to exchange bindings). In this case it is possible to configure the consistent hash exchange to route based on a named header instead. To do this, declare the exchange with a string argument called "hash-header" naming the header to be used.

When a "hash-header" is specified, the chosen header should be provided. If published messages do not contain the header, they will all get routed to the same arbitrarily chosen queue.

#!/usr/bin/env python

import pika
import time

conn = pika.BlockingConnection(pika.ConnectionParameters(host='localhost'))
ch   = conn.channel()

args = {u'hash-header': u'hash-on'}
ch.exchange_declare(exchange='e2',
                    exchange_type='x-consistent-hash',
                    arguments=args,
                    durable=True)

for q in ['q1', 'q2', 'q3', 'q4']:
    ch.queue_declare(queue=q, durable=True)
    ch.queue_purge(queue=q)

for q in ['q1', 'q2']:
    ch.queue_bind(exchange='e2', queue=q, routing_key='1')

for q in ['q3', 'q4']:
    ch.queue_bind(exchange='e2', queue=q, routing_key='2')

n = 100000

for rk in list(map(lambda s: str(s), range(0, n))):
    hdrs = {u'hash-on': rk}
    ch.basic_publish(exchange='e2',
                     routing_key='',
                     body='',
                     properties=pika.BasicProperties(content_type='text/plain',
                                                     delivery_mode=2,
                                                     headers=hdrs))
print('Done publishing.')

print('Waiting for routing to finish...')
# in order to keep this example simpler and focused,
# wait for a few seconds instead of using publisher confirms and waiting for those
time.sleep(5)

print('Done.')
conn.close()

package com.rabbitmq.examples;

import com.rabbitmq.client.*;

import java.io.IOException;
import java.util.Arrays;
import java.util.HashMap;
import java.util.Map;
import java.util.concurrent.TimeoutException;

public class ConsistentHashExchangeExample2 {
  public static final String EXCHANGE = "e2";
  private static String EXCHANGE_TYPE = "x-consistent-hash";

  public static void main(String[] argv) throws IOException, TimeoutException, InterruptedException {
    ConnectionFactory cf = new ConnectionFactory();
    Connection conn = cf.newConnection();
    Channel ch = conn.createChannel();

    for (String q : Arrays.asList("q1", "q2", "q3", "q4")) {
      ch.queueDeclare(q, true, false, false, null);
      ch.queuePurge(q);
    }

    Map<String, Object> args = new HashMap<>();
    args.put("hash-header", "hash-on");
    ch.exchangeDeclare(EXCHANGE, EXCHANGE_TYPE, true, false, args);

    for (String q : Arrays.asList("q1", "q2")) {
      ch.queueBind(q, EXCHANGE, "1");
    }

    for (String q : Arrays.asList("q3", "q4")) {
      ch.queueBind(q, EXCHANGE, "2");
    }

    ch.confirmSelect();


    for (int i = 0; i < 100000; i++) {
      AMQP.BasicProperties.Builder bldr = new AMQP.BasicProperties.Builder();
      Map<String, Object> hdrs = new HashMap<>();
      hdrs.put("hash-on", String.valueOf(i));
      ch.basicPublish(EXCHANGE, "", bldr.headers(hdrs).build(), "".getBytes("UTF-8"));
    }

    ch.waitForConfirmsOrDie(10000);

    System.out.println("Done publishing!");
    System.out.println("Evaluating results...");
    // wait for one stats emission interval so that queue counters
    // are up-to-date in the management UI
    Thread.sleep(5);

    System.out.println("Done.");
    conn.close();
  }
}

#!/usr/bin/env ruby

require 'bundler'
Bundler.setup(:default, :test)
require 'bunny'

conn = Bunny.new
conn.start

ch = conn.create_channel
ch.confirm_select

q1 = ch.queue("q1", durable: true)
q2 = ch.queue("q2", durable: true)
q3 = ch.queue("q3", durable: true)
q4 = ch.queue("q4", durable: true)

[q1, q2, q3, q4]. each(&:purge)

x  = ch.exchange("x2", type: "x-consistent-hash", durable: true, arguments: {"hash-header" => "hash-on"})

[q1, q2].each { |q| q.bind(x, routing_key: "1") }
[q3, q4].each { |q| q.bind(x, routing_key: "2") }

n = 100_000
(0..n).map(&:to_s).each do |i|
  x.publish(i.to_s, routing_key: rand.to_s, headers: {"hash-on": i})
end

ch.wait_for_confirms
puts "Done publishing!"

# wait for queue stats to be emitted so that management UI numbers
# are up-to-date
sleep 5
conn.close
puts "Done"

With RabbitMQ Erlang client:

-include_lib("amqp_client/include/amqp_client.hrl").

test() ->
    {ok, Conn} = amqp_connection:start(#amqp_params_network{}),
    {ok, Chan} = amqp_connection:open_channel(Conn),
    Queues = [<<"q0">>, <<"q1">>, <<"q2">>, <<"q3">>],
    amqp_channel:call(
      Chan, #'exchange.declare'{
              exchange  = <<"e">>,
              type      = <<"x-consistent-hash">>,
              arguments = [{<<"hash-header">>, longstr, <<"hash-on">>}]
            }),
    [amqp_channel:call(Chan, #'queue.declare'{queue = Q}) || Q <- Queues],
    [amqp_channel:call(Chan, #'queue.bind' {queue = Q,
                                            exchange = <<"e">>,
                                            routing_key = <<"1">>})
        || Q <- [<<"q0">>, <<"q1">>]],
    [amqp_channel:call(Chan, #'queue.bind' {queue = Q,
                                            exchange = <<"e">>,
                                            routing_key = <<"2">>})
        || Q <- [<<"q2">>, <<"q3">>]],
    RK = list_to_binary(integer_to_list(random:uniform(1000000))),
    Msg = #amqp_msg {props = #'P_basic'{headers = [{<<"hash-on">>, longstr, RK}]}, payload = <<>>},
    [amqp_channel:call(Chan,
                   #'basic.publish'{
                     exchange = <<"e">>,
                     routing_key = <<"">>,
                   }, Msg) || _ <- lists:seq(1, 100000)],
amqp_connection:close(Conn),
ok.

Instead of a value in the header property, you can route on the message_id, correlation_id, or timestamp message properties. To do so, declare the exchange with a string argument called "hash-property" naming the property to be used. The "hash-header" and "hash-property" are mutually exclusive.

When a "hash-property" is specified, the chosen property should be provided. If published messages do not contain the property, they will all get routed to the same arbitrarily chosen queue.

#!/usr/bin/env python

import pika
import time

conn = pika.BlockingConnection(pika.ConnectionParameters(host='localhost'))
ch   = conn.channel()

args = {u'hash-property': u'message_id'}
ch.exchange_declare(exchange='e3',
                    exchange_type='x-consistent-hash',
                    arguments=args,
                    durable=True)

for q in ['q1', 'q2', 'q3', 'q4']:
    ch.queue_declare(queue=q, durable=True)
    ch.queue_purge(queue=q)

for q in ['q1', 'q2']:
    ch.queue_bind(exchange='e3', queue=q, routing_key='1')

for q in ['q3', 'q4']:
    ch.queue_bind(exchange='e3', queue=q, routing_key='2')

n = 100000

for rk in list(map(lambda s: str(s), range(0, n))):
    ch.basic_publish(exchange='e3',
                     routing_key='',
                     body='',
                     properties=pika.BasicProperties(content_type='text/plain',
                                                     delivery_mode=2,
                                                     message_id=rk))
print('Done publishing.')

print('Waiting for routing to finish...')
# in order to keep this example simpler and focused,
# wait for a few seconds instead of using publisher confirms and waiting for those
time.sleep(5)

print('Done.')
conn.close()

package com.rabbitmq.examples;

import com.rabbitmq.client.*;

import java.io.IOException;
import java.util.Arrays;
import java.util.HashMap;
import java.util.Map;
import java.util.concurrent.TimeoutException;

public class ConsistentHashExchangeExample3 {
  public static final String EXCHANGE = "e3";
  private static String EXCHANGE_TYPE = "x-consistent-hash";

  public static void main(String[] argv) throws IOException, TimeoutException, InterruptedException {
    ConnectionFactory cf = new ConnectionFactory();
    Connection conn = cf.newConnection();
    Channel ch = conn.createChannel();

    for (String q : Arrays.asList("q1", "q2", "q3", "q4")) {
      ch.queueDeclare(q, true, false, false, null);
      ch.queuePurge(q);
    }

    Map<String, Object> args = new HashMap<>();
    args.put("hash-property", "message_id");
    ch.exchangeDeclare(EXCHANGE, EXCHANGE_TYPE, true, false, args);

    for (String q : Arrays.asList("q1", "q2")) {
      ch.queueBind(q, EXCHANGE, "1");
    }

    for (String q : Arrays.asList("q3", "q4")) {
      ch.queueBind(q, EXCHANGE, "2");
    }

    ch.confirmSelect();


    for (int i = 0; i < 100000; i++) {
      AMQP.BasicProperties.Builder bldr = new AMQP.BasicProperties.Builder();
      ch.basicPublish(EXCHANGE, "", bldr.messageId(String.valueOf(i)).build(), "".getBytes("UTF-8"));
    }

    ch.waitForConfirmsOrDie(10000);

    System.out.println("Done publishing!");
    System.out.println("Evaluating results...");
    // wait for one stats emission interval so that queue counters
    // are up-to-date in the management UI
    Thread.sleep(5);

    System.out.println("Done.");
    conn.close();
  }
}

#!/usr/bin/env ruby

require 'bundler'
Bundler.setup(:default, :test)
require 'bunny'

conn = Bunny.new
conn.start

ch = conn.create_channel
ch.confirm_select

q1 = ch.queue("q1", durable: true)
q2 = ch.queue("q2", durable: true)
q3 = ch.queue("q3", durable: true)
q4 = ch.queue("q4", durable: true)

[q1, q2, q3, q4].each(&:purge)

x  = ch.exchange("x3", type: "x-consistent-hash", durable: true, arguments: {"hash-property" => "message_id"})

[q1, q2].each { |q| q.bind(x, routing_key: "1") }
[q3, q4].each { |q| q.bind(x, routing_key: "2") }

n = 100_000
(0..n).map(&:to_s).each do |i|
  x.publish(i.to_s, routing_key: rand.to_s, message_id: i)
end

ch.wait_for_confirms
puts "Done publishing!"

# wait for queue stats to be emitted so that management UI numbers
# are up-to-date
sleep 5
conn.close
puts "Done"

-include_lib("amqp_client/include/amqp_client.hrl").

test() ->
    {ok, Conn} = amqp_connection:start(#amqp_params_network{}),
    {ok, Chan} = amqp_connection:open_channel(Conn),
    Queues = [<<"q0">>, <<"q1">>, <<"q2">>, <<"q3">>],
    amqp_channel:call(Chan,
                  #'exchange.declare'{
                    exchange = <<"e">>, type = <<"x-consistent-hash">>,
                    arguments = {<<"hash-property">>, longstr, <<"message_id">>}
                  }),
    [amqp_channel:call(Chan, #'queue.declare'{queue = Q}) || Q <- Queues],
    [amqp_channel:call(Chan, #'queue.bind'{queue = Q,
                                           exchange = <<"e">>,
                                           routing_key = <<"1">>})
        || Q <- [<<"q0">>, <<"q1">>]],
    [amqp_channel:call(Chan, #'queue.bind' {queue = Q,
                                            exchange = <<"e">>,
                                            routing_key = <<"2">>})
        || Q <- [<<"q2">>, <<"q3">>]],
    RK = list_to_binary(integer_to_list(random:uniform(1000000)),
    Msg = #amqp_msg{props = #'P_basic'{message_id = RK}, payload = <<>>},
    [amqp_channel:call(Chan,
                   #'basic.publish'{
                     exchange = <<"e">>,
                     routing_key = <<"">>,
                     )
                   }, Msg) || _ <- lists:seq(1, 100000)],
amqp_connection:close(Conn),
ok.

If you have questions or need help, feel free to ask on the RabbitMQ mailing list.

The hash function used in this plugin as of RabbitMQ 3.7.8 is A Fast, Minimal Memory, Consistent Hash Algorithm by Lamping and Veach. Erlang's phash2 function is used to convert non-integer values to an integer one that can be used by the jump consistent hash function by Lamping and Veach.

A Chi-squared test was used to evaluate distribution uniformity. Below are the results for 18 bucket counts and how they compare to two commonly used p-value thresholds:

When a queue is bound to a consistent hash exchange, the protocol method, queue.bind, carries a weight in the routing (binding) key. The binding is given a number of buckets on the hash ring (hash space) equal to the weight. When a queue is unbound, the buckets added for the binding are deleted. These two operations use linear algorithms to update the ring.

To perform routing the exchange extract the appropriate value for hashing, hashes it and retrieves a bucket number from the ring, then the bucket and its associated queue.

The implementation assumes there is only one binding between a consistent hash exchange and a queue. Having more than one binding is unnecessary because queue weight can be provided at the time of binding.

The state of the hash space is distributed across all cluster nodes.

(c) 2007-2024 Broadcom. The term “Broadcom” refers to Broadcom Inc. and/or its subsidiaries. All rights reserved.

Released under the Mozilla Public License 2.0, same as RabbitMQ. See LICENSE for details.