RabbitMQ tutorial - Work Queues

Work Queues

(using the amqp Elixir library)

info

Prerequisites

This tutorial assumes RabbitMQ is installed and running on localhost on the standard port (5672). In case you use a different host, port or credentials, connections settings would require adjusting.

Where to get help

If you're having trouble going through this tutorial you can contact us through the mailing list or RabbitMQ community Slack.

In the first tutorial we wrote programs to send and receive messages from a named queue. In this one we'll create a Work Queue that will be used to distribute time-consuming tasks among multiple workers.

The main idea behind Work Queues (aka: Task Queues) is to avoid doing a resource-intensive task immediately and having to wait for it to complete. Instead we schedule the task to be done later. We encapsulate a task as a message and send it to the queue. A worker process running in the background will pop the tasks and eventually execute the job. When you run many workers the tasks will be shared between them.

This concept is especially useful in web applications where it's impossible to handle a complex task during a short HTTP request window.

Preparation

In the previous part of this tutorial we sent a message containing "Hello World!". Now we'll be sending strings that stand for complex tasks. We don't have a real-world task, like images to be resized or pdf files to be rendered, so let's fake it by just pretending we're busy - by using the :timer.sleep/1 function. We'll take the number of dots in the string as its complexity; every dot will account for one second of "work". For example, a fake task described by Hello... will take three seconds.

We will slightly modify the send.exs code from our previous example, to allow arbitrary messages to be sent from the command line. This program will schedule tasks to our work queue, so let's name it new_task.exs:

message =
  case System.argv do
    []    -> "Hello World!"
    words -> Enum.join(words, " ")
  end

AMQP.Basic.publish(channel, "", "task_queue", message, persistent: true)

IO.puts " [x] Send '#{message}'"

Our old receive.exs script also requires some changes: it needs to fake a second of work for every dot in the message body. It will pop messages from the queue and perform the task, so let's call it worker.exs:

defmodule Worker do
  def wait_for_messages(channel) do
    receive do
      {:basic_deliver, payload, meta} ->
        IO.puts " [x] Received #{payload}"
        payload
        |> to_char_list
        |> Enum.count(fn x -> x == ?. end)
        |> Kernel.*(1000)
        |> :timer.sleep
        IO.puts " [x] Done."

        wait_for_messages(channel)
    end
  end
end

Round-robin dispatching

One of the advantages of using a Task Queue is the ability to easily parallelise work. If we are building up a backlog of work, we can just add more workers and that way, scale easily.

First, let's try to run two worker.exs scripts at the same time. They will both get messages from the queue, but how exactly? Let's see.

You need three consoles open. Two will run the worker.exs script. These consoles will be our two consumers - C1 and C2.

mix run worker.exs
# => [*] Waiting for messages. To exit press CTRL+C, CTRL+C

mix run worker.exs
# => [*] Waiting for messages. To exit press CTRL+C, CTRL+C

In the third one we'll publish new tasks. Once you've started the consumers you can publish a few messages:

# shell 3
mix run new_task.exs First message.
mix run new_task.exs Second message..
mix run new_task.exs Third message...
mix run new_task.exs Fourth message....
mix run new_task.exs Fifth message.....

Let's see what is delivered to our workers:

# shell 1
mix run worker.exs
# => [*] Waiting for messages. To exit press CTRL+C, CTRL+C
# => [x] Received 'First message.'
# => [x] Received 'Third message...'
# => [x] Received 'Fifth message.....'

# shell 2
mix run worker.exs
# => [*] Waiting for messages. To exit press CTRL+C, CTRL+C
# => [x] Received 'Second message..'
# => [x] Received 'Fourth message....'

By default, RabbitMQ will send each message to the next consumer, in sequence. On average every consumer will get the same number of messages. This way of distributing messages is called round-robin. Try this out with three or more workers.

Message acknowledgment

Doing a task can take a few seconds so you may wonder what happens if a consumer starts a long task and it terminates when it is not fully complete. With our current code, once RabbitMQ delivers a message to the consumer, it immediately marks it for deletion. In this case, if you terminate a worker, the message it was just processing is lost. All the messages that were dispatched to this particular worker but were not yet handled are also lost.

But we don't want to lose any tasks. If a worker dies, we'd like the task to be delivered to another worker.

In order to make sure a message is never lost, RabbitMQ supports message acknowledgments. An ack(nowledgement) is sent back by the consumer to tell RabbitMQ that a particular message had been received, processed and that RabbitMQ is free to delete it.

If a consumer dies (its channel is closed, connection is closed, or TCP connection is lost) without sending an ack, RabbitMQ will understand that a message wasn't processed fully and will re-queue it. If there are other consumers online at the same time, it will then quickly redeliver it to another consumer. That way you can be sure that no message is lost, even if the workers occasionally die.

A timeout (30 minutes by default) is enforced on consumer delivery acknowledgement. This helps detect buggy (stuck) consumers that never acknowledge deliveries. You can increase this timeout as described in Delivery Acknowledgement Timeout.

Manual message acknowledgments are turned on by default. In previous examples we explicitly turned them off via the no_ack: true flag. It's time to remove this flag and send a proper acknowledgment from the worker, once we're done with a task.

defmodule Worker do
  def wait_for_messages(channel) do
    receive do
      {:basic_deliver, payload, meta} ->
        IO.puts " [x] Received #{payload}"
        payload
        |> to_char_list
        |> Enum.count(fn x -> x == ?. end)
        |> Kernel.*(1000)
        |> :timer.sleep
        IO.puts " [x] Done."
        AMQP.Basic.ack(channel, meta.delivery_tag)

        wait_for_messages(channel)
    end
  end
end

AMQP.Basic.consume(channel, "hello")

Using this code, you can ensure that even if you terminate a worker node using CTRL+C while it was processing a message, nothing is lost. Soon after the worker is terminated, all unacknowledged messages are redelivered.

Acknowledgement must be sent on the same channel that received the delivery. Attempts to acknowledge using a different channel will result in a channel-level protocol exception. See the doc guide on confirmations to learn more.

Forgotten acknowledgment

It's a common mistake to miss the AMQP.Basic.ack. It's an easy error, but the consequences are serious. Messages will be redelivered when your client quits (which may look like random redelivery), but RabbitMQ will eat more and more memory as it won't be able to release any unacked messages.

In order to debug this kind of mistake you can use rabbitmqctl to print the messages_unacknowledged field:
sudo rabbitmqctl list_queues name messages_ready messages_unacknowledged
On Windows, drop the sudo:
rabbitmqctl.bat list_queues name messages_ready messages_unacknowledged

Message durability

We have learned how to make sure that even if the consumer dies, the task isn't lost. But our tasks will still be lost if RabbitMQ server stops.

When RabbitMQ quits or crashes it will forget the queues and messages unless you tell it not to. Two things are required to make sure that messages aren't lost: we need to mark both the queue and messages as durable.

First, we need to make sure that the queue will survive a RabbitMQ node restart. In order to do so, we need to declare it as durable:

AMQP.Queue.declare(channel, "hello", durable: true)

Although this command is correct by itself, it won't work in our setup. That's because we've already defined a queue called hello which is not durable. RabbitMQ doesn't allow you to redefine an existing queue with different parameters and will return an error to any program that tries to do that. But there is a quick workaround - let's declare a queue with different name, for example task_queue:

AMQP.Queue.declare(channel, "task_queue", durable: true)

This AMQP.Queue.declare change needs to be applied to both the producer and consumer code.

At that point we're sure that the task_queue queue won't be lost even if RabbitMQ restarts. Now we need to mark our messages as persistent

by supplying a persistent: true property.

AMQP.Basic.publish(channel, "", "task_queue", message, persistent: true)

Note on message persistence

Marking messages as persistent doesn't fully guarantee that a message won't be lost. Although it tells RabbitMQ to save the message to disk, there is still a short time window when RabbitMQ has accepted a message and hasn't saved it yet. Also, RabbitMQ doesn't do fsync(2) for every message -- it may be just saved to cache and not really written to the disk. The persistence guarantees aren't strong, but it's more than enough for our simple task queue. If you need a stronger guarantee then you can use publisher confirms.

Fair dispatch

You might have noticed that the dispatching still doesn't work exactly as we want. For example in a situation with two workers, when all odd messages are heavy and even messages are light, one worker will be constantly busy and the other one will do hardly any work. Well, RabbitMQ doesn't know anything about that and will still dispatch messages evenly.

This happens because RabbitMQ just dispatches a message when the message enters the queue. It doesn't look at the number of unacknowledged messages for a consumer. It just blindly dispatches every n-th message to the n-th consumer.

In order to defeat that we can use the AMQP.Basic.qos function with the prefetch_count: 1 setting. This tells RabbitMQ not to give more than one message to a worker at a time. Or, in other words, don't dispatch a new message to a worker until it has processed and acknowledged the previous one. Instead, it will dispatch it to the next worker that is not still busy.

AMQP.Basic.qos(channel, prefetch_count: 1)

Note about queue size

If all the workers are busy, your queue can fill up. You will want to keep an eye on that, and maybe add more workers, or use message TTL.

Putting it all together

Final code of our new_task.exs script:

{:ok, connection} = AMQP.Connection.open
{:ok, channel} = AMQP.Channel.open(connection)

AMQP.Queue.declare(channel, "task_queue", durable: true)

message =
  case System.argv do
    []    -> "Hello World!"
    words -> Enum.join(words, " ")
  end

AMQP.Basic.publish(channel, "", "task_queue", message, persistent: true)
IO.puts " [x] Sent '#{message}'"

AMQP.Connection.close(connection)

(new_task.exs source)

And our worker:

defmodule Worker do
  def wait_for_messages(channel) do
    receive do
      {:basic_deliver, payload, meta} ->
        IO.puts " [x] Received #{payload}"
        payload
        |> to_char_list
        |> Enum.count(fn x -> x == ?. end)
        |> Kernel.*(1000)
        |> :timer.sleep
        IO.puts " [x] Done."
        AMQP.Basic.ack(channel, meta.delivery_tag)

        wait_for_messages(channel)
    end
  end
end

{:ok, connection} = AMQP.Connection.open
{:ok, channel} = AMQP.Channel.open(connection)

AMQP.Queue.declare(channel, "task_queue", durable: true)
AMQP.Basic.qos(channel, prefetch_count: 1)

AMQP.Basic.consume(channel, "task_queue")
IO.puts " [*] Waiting for messages. To exit press CTRL+C, CTRL+C"

Worker.wait_for_messages(channel)

(worker.exs source)

Using message acknowledgments and prefetch_count you can set up a work queue. The durability options let the tasks survive even if RabbitMQ is restarted.

Now we can move on to tutorial 3 and learn how to deliver the same message to many consumers.

RabbitMQ tutorial - Work Queues

Work Queues​

(using the amqp Elixir library)​

Prerequisites​

Where to get help​

Preparation​

Round-robin dispatching​

Message acknowledgment​

Forgotten acknowledgment​

Message durability​

Note on message persistence​

Fair dispatch​

Note about queue size​

Putting it all together​