Dawdle weaponizes Amazon SQS for use in your Elixir applications. Use it when you want to handle something later, or, better yet, when you want someone else to handle it.
Put simply, you can use Dawdle to signal an event. That event can then be
handled by another process within the current Erlang node, or another node
altogether. You can even signal events with a delay, similar to
Process.send_after/4
, but in a distributed and node-failure-tolerant manner.
The events are encoded and enqueued into Amazon AWS's Simple Queue Service (SQS). This means that if, for example, you're running your BEAM system in a Kubernetes cluster, and one or more of your pods die or are restarted or terminated or what have you, your events will still fire and be handled by whatever pods are available. If no pods are available for some reason, the events will still be preserved and can be handled when a pod is available again to service them.
Add dawdle
to your list of dependencies in mix.exs
:
def deps do
[
{:dawdle, "~> 0.7.0"}
]
end
Dawdle provides two backends: A local one for development and testing which does not require access to SQS (the default), and an SQS one for use in an AWS environment.
To enable the SQS backend, set the following in your application's config.exs
:
config :dawdle,
backend: Dawdle.Backend.SQS
To disable the Dawdle queue listener on a node, use the following:
config :dawdle,
start_listener: false
You can also set the environment variables DAWDLE_BACKEND
or
DAWDLE_START_LISTENER
.
To configure your SQS queue, set the queue URL in config.exs
:
config :dawdle, Dawdle.Backend.SQS,
region: "us-east-1",
queue_url: "https://sqs.us-east-1.amazonaws.com/XXXXXXXXXXXX/my-dawdle-message-queue"
These values can also be set by using the environment variables
DAWDLE_SQS_REGION
and DAWDLE_SQS_QUEUE_URL
.
The configuration should be managed either via config.exs
or by setting the
environment variables. Trying to mix the two will result in some changes being
overwritten in surprising ways. Caveat emptor.
Obviously the configured SQS queue needs to exist and be accessible by your
application. AWS authentication is handled by
ex_aws. If you're already using ex_aws
for
something else, your configuration should already be good. If not, follow the
configuration instructions on that page to set up your AWS key.
Dawdle uses two queues: one for normal messages and one for delayed events. The
message queue must be a Standard Queue. They can be configured with default
values, except that Receive Message Wait Time
should be set to 20 seconds
to enable long polling. It is also a good idea to set Default Visibility Timeout
to a short value, like 2 seconds.
The queue can be created using the aws
CLI or from the AWS Control Panel.
Here are example commands for creating the queue from the CLI:
$ aws sqs create-queue --queue-name my-dawdle-message-queue --attributes ReceiveMessageWaitTimeSeconds=20
{
"QueueUrl": "https://xx-xxxx-x.queue.amazonaws.com/XXXXXXXXXXXX/my-dawdle-message-queue"
}
Full docs can be found at https://hexdocs.pm/dawdle.
Event handlers are where Dawdle really begins to shine. An event is essentially just an Elixir struct. Define an event and event handler, then when you signal that event using Dawdle, the event handler will be called to process the event.
- Define the event
defmodule MyApp.TestEvent do
defstruct :foo, :bar
end
- Create an event handler
defmodule MyApp.TestEventHandler do
use Dawdle.Handler, only: [MyApp.TestEvent]
alias MyApp.TestEvent
def handle_event(%TestEvent{} = event) do
IO.puts("Handling event #{inspect(event)}")
:ok
end
end
- Signal an event
t = %MyApp.TestEvent{foo: 1, bar: 2}
Dawdle.signal(t)
It is possible to signal an event that will bypass the queue and be delivered
directly to the appropriate handlers. The handlers will execute immediately on
the current node, but in a separate process. Simply pass direct: true
to
Dawdle.signal/2
:
t = %MyApp.TestEvent{foo: 1, bar: 2}
Dawdle.signal(t, direct: true)
The event handler will execute on a node running the Dawdle application with pollers enabled.
Note that if you are handling events on nodes different from where they are signaled, then you need to ensure that the event definintions are available in both places.
There is a basic, experimental API which involves passing an anonymous function
to Dawdle.call/1
. The function will execute on a node running the Dawdle
application with pollers enabled.
iex> Dawdle.call(fn -> IO.puts("Hello World!") end)
:ok
Hello World!
Passing a function to Dawdle.call_after/2
will result in that function being
called after the specified delay.
iex> Dawdle.call_after(2000, fn -> IO.puts("Hello Future!") end)
:ok
# 2 seconds later
Hello Future!
This API is included for feedback and may be discontinued or extracted into a separate library for Dawdle 1.0.
Because the events are managed outside of your BEAM VM(s), they will be preserved and handled by an available node even if the node that originally signaled them no longer exists.
Dawdle does not support FIFO queues. The cost paid for strict ordering is reduced throughput and the possibility of the queue becoming clogged if an event handler takes a long time. Dawdle resolves that tradeoff in favor of using standard queues to maximize throughput and avoid queue clogs.
SQS standard queues are not millisecond-precision timing devices. Their maximum delay precision is 1 second, so any timeouts given in fractions of a second will be rounded down.
SQS standard queues guarantee at least once delivery. In practice it's almost always exactly once, but your code needs to handle the possibility that a given handler will execute multiple times.
SQS does not guarantee ordering on its standard queues, and Dawdle leverages concurrency when dispatching events to handlers in order to keep latency low. So if you signal two events in quick succession, it's not guaranteed that the handlers will execute in the same order that the events were signaled.
SQS has an upper message size limit of 256KB, and the terms sent via it are Base64 encoded, so avoid sending large structures in your message. If you need a large bit of data as part of your message, stash it in a persistent store first and send the key through Dawdle.
SQS delays are limited to 15 minutes. We handle longer delays by using multiple chained messages, so factor this into any capacity calculations you're doing.
We are still doing some experimentation to get a feel for what works best. The core of Dawdle is working well, and we are mostly refining and focusing internal message handling. Some of the ideas on this list are speculative and may not make it into the 1.0 release. We are happy to listen to feedback on any of these features, and pull requests are always welcome.
- More backends (RabbitMQ?, Redis?, GCP pub/sub?)
- More robust local delivery of events
- Having multiple backend queues configured at once (i.e., SQS and RabbitMQ)
- Having multiple instances of a single backend (i.e., multiple SQS queues)
- Allowing more discretion on when and where events are handled
That last item would allow, for example, having an instance of the application that generates events, and another instance that only processes certain events. So, some events would be handled in the main application, while other, more expensive events, could be handled on nodes that aren't loaded with other work.