We're planting a tree for every job application! Click here to learn more

Elixir: Practical Concurrency Cookbook

Javier Garcia

22 Jul 2021

7 min read

Elixir: Practical Concurrency Cookbook
  • Erlang

While the Erlang runtime is known for being a highly concurrent platform to which Elixir compiles, most of us end up solving the same problems in our daily jobs. We use Phoenix to bootstrap our web applications, write JSON APIs and sprinkle our pages with some javascript. And although it is true that simply by using Phoenix we're already getting for free the whole concurrency model, the majority of the features we develop for our products don't leverage all that often distributed Erlang, complex supervision trees or fleets of GenServers.

In this post, my intention is to highlight some common scenarios where we can leverage Erlang's concurrency model as well as Elixir's abstractions to build better, faster and more secure software. Think of it like... A cookbook.

Fire and forget

The easiest problem we may have to solve is how to do a "fire and forget" computation. In other words, how can we tell our system to execute some code asynchronously and not care about when it finishes, nor the result. To do this Elixir provides with the Task abstraction.

Tasks are processes meant to run a single action within their life-cycle. It can be a long-lived operation, like batch processing of records, or a short one, like sending a Slack notification.

Usually, to run a one-off task the easiest thing is to use Task.start/1. However, here is the first tip: don't do that. The best way to run a one-off action in Elixir is to spawn the tasks under a supervision tree, with Task.Supervisor.start_child/2.

The main reason why it's preferable to run tasks under their own supervision tree is to allow for a proper clean-up of processes. When a Supervisor is taken down, so are all its children, which allows you to take down the application cleanly.

This doesn't mean that you want your tasks to be restarted. In fact, the default strategy for the Task.Supervisor is :temporary, which means they're never restarted. It's just a means to avoid dangling processes if things go weird at some point. And they always do.

Here is an example:

defmodule FireAndForgetExample.Application do
  use Application

  def start(_type, _args) do
    children = [
      # Start the supervision tree under the OTP Application.
      {Task.Supervisor, name: FireAndForgetExample.TaskSupervisor}

    Supervisor.start_link(children, strategy: :one_for_one)

defmodule FireAndForgetExample.OtherModule do
  def process_event(event) do
    # Start the task under the supervision tree.
    Task.Supervisor.start_child(FireAndForgetExample.TaskSupervisor, fn ->
      send_slack_notification("Hey! We got an event!")

    |> do_something()
    |> do_something_else()


However, what if we do care about the results? Sometimes it's useful to run a certain operation and forget about it, but most times we actually do want to do something with its result.

If the problem we have at hand is one which consists of multiple operations which you can run asynchronously because they don't depend on each other, like for example, uploading a bunch of documents to S3, or sending a batch of emails to different people, the easiest solution is to implement a fan-in/fan-out strategy.

This can be done by using tasks too. We can do it naively without using a Supervisor or we can spin them up under a Supervisor as we've mentioned before. I always recommend using a Supervisor in code that is going to be shipped to production, however, for the sake of simplicity, let's see an example without it:

defmodule FanInFanOutExample do
  def send_notifications(notifications) do
    # Spin a task per element
    |> Enum.map(&Task.async(fn -> send_single_notification(&1) end)
    # Await all of them
    |> Enum.map(&Task.await/1)

  def send_single_notification(notification) do
    # ...

The good thing about this approach is that it will only take as long as the the longest task, and by the time the function has finished we will have a list with all the results. In Elixir 1.11 there's also Task.await_many/2, which under the hood, it does a little bit more than a simple iteration and an await, but in the end, does get us to the same place.

Scheduling work

A different kind of problem we might come across is "_How can we run some work periodically every N minutes/seconds/etc.". This is fairly simple leveraging a different abstraction available in Elixir but ultimately in OTP: the GenServer.

GenServer is short for a "Generic Server". It's basically a process that can receive messages and allows us to specify callbacks so it does different things with those messages. I won't go very much into detail about GenServers because I already wrote about them a while back.

So, how do we do the scheduling with a GenServer? Simple: by leveraging the handle_info/2 callback and Process.send_after/3. In other words, we will add a callback to the GenServer which does the work and then schedule the message with Process.send_after/3. Lastly, to make sure it runs again after some time, we make sure to call Process.send_after/3 again before the callback returns. Let's see how it looks:

defmodule SchedulingExample do
  use GenServer

  @default_minutes 3

  def start_link(args \\ []) do
    GenServer.start_link(__MODULE__, to_map(args))

  defp to_map(args) do
      minutes: Keyword.get(args, :minutes, @default_minutes),
      forever: Keyword.get(args, :forever, true),

  def init(%{minutes: minutes} = state) do
    {:ok, state}

  def handle_info(:work, %{minutes: minutes, forever: forever} = state) do
    # Do my work here ...

    if forever do

    {:noreply, state}

  defp schedule_work(minutes) do
    milliseconds = to_milliseconds(minutes)
    Process.send_after(self(), :work, milliseconds)

  defp to_milliseconds(minutes) do
    |> :timer.minutes()
    |> Kernel.trunc()

Also, when leveraging GenServers and other abstractions over processes, it's usually a good call to keep the server module with as little business logic as possible. In this particular case, if the GenServer can simply call a function from a module, we're golden. This way we can decouple completely process management from our business, allowing for easier testing... And an easier time.

Lastly, we probably want to spawn this kind of worker processes under a Supervisor too, but probably with a different strategy, like a :one_for_one, so they're restarted in case they crash.

Running it at specific times

In some cases, we might want to run our code at specific times. Not necessarily every 3 minutes, but every day at 08.00AM. While this is perfectly achievable leveraging the same tools, I'll be pragmatic and recommend Quantum. It allows you to schedule the execution of functions with a cron syntax and takes away all the complexity of managing processes. It's a seasoned library, widely-adopted within the community, very lightweight... And extremely simple.

Caching data that's accessed very often

Occasionally you will find yourself with an endpoint that that makes queries that take too long or a process which consistently has to crunch a lot of data and provides a bad user experience. In these situations sometimes caching the results may make sense. Sometimes it might make more sense to spend a few hours tweaking the queries themselves or redesigning the solution, but sometimes caching might make sense. Let's talk about when it does.

The simplest approach: Agents

If you're already experienced with Elixir or Erlang, you'll know that its data structures are immutable, however, it has its own way to work with the shared mutable state: processes. In order to save some state, access it and change it, we can do so with a process and in many different ways.

The most simple solution for holding some state is creating an Agent. Agents are the simplest possible abstraction around the state, and sometimes, if what we need is precisely a simple solution without too many batteries, it might actually be the best option. One of the good things about Agents is that it is a single process, which means that many concurrent clients will get their share of the Agent sequentially, which means you don't have to worry about race conditions. On the other hand, that can also be a bad thing if it starts becoming a bottleneck.

An ETS based approach

Other times, if the Agent doesn't cut it for you, you might something faster. In these cases ETS might be a good option. The good thing about ETS is that it will always be faster because it doesn't go through the Erlang Scheduler, furthermore it also supports concurrent reads and writes, which the Agent does not. However, it's a bit more limited when you want to do atomic operations. Overall it's very well suited for a simple shared key/value store, but if it's better suited or not for your specific problem, that's up to you. A naive approach could look something like the below:

defmodule EtsCacheExample do
  def init!(seed, table_name) when is_atom(table_name) do
    case :ets.info(table_name) do
      :undefined ->
        :ets.new(table_name, [:set, :public, :named_table])

      _ ->
        raise "ETS table with name #{table_name(pool_name)} already exists."

    add(seed, pool_name)

  def teardown(table_name) do

  def add(value, table_name) do
    :ets.insert_new(table_name, {value})

  def exists?(value, pool_name) do
    case :ets.lookup(table_name, value) do
      [] -> false
      _ -> true

  def retrieve_all(table_name) do
    |> :ets.match({:"$1"})
    |> List.flatten()

One per user: GenServers

The third option I'm going to mention is GenServers. Most of the times either the Agent or ETS should be enough, however, in some cases it might make sense to provide each API user with its own little cache. A good reason could be because we need to provide certain atomicity for the read/writes and an Agent would be a bottleneck. One of the strengths of GenServers is that they allow us to spin one up per user very easily, so they don't become a bottleneck where the Agent does.

Some last thoughts on caching

However, while Elixir does provide the necessary abstractions to make caching easy, my recommendation on this one is usually to lean on the community's shoulders. Saša Jurić, for example, wrote a while back ConCache which does exactly this, but there are many others out there. The good thing about not implementing one it's own is that there are many edge cases when dealing with concurrency and it's easy to get it wrong the first few times. As they say, the two hardest problems in computer science are naming and caching.

Wrapping up

With all this, I hope to have shed some light in some potential solutions to some of your maybe problems. Like all problems in IT, every solution sometimes makes sense and sometimes it does not, but at the end of the day, I believe that these are all staple techniques within any Elixir or Erlang developer. Elixir makes it so easy to work with concurrent code that most problems can be tackled without the need to bring a third-party library. There are of course more than could be mentioned, like using gen_statem for state machines, or worker pools for throttling work, but I'll leave those for another day.

Oh, and one last thing, don't forget to supervise your processes... It's fault-tolerance for free :)

Sign up to Functional Works for more articles like this!

Did you like this article?

Javier Garcia

I like coding, poetry, strong coffee, Miles Davis and Pokemon :)

See other articles by Javier

Related jobs

See all


The company

  • Remote


The company

  • Remote


The company

  • Remote


The company

  • Remote

Related articles

JavaScript Functional Style Made Simple

JavaScript Functional Style Made Simple

Daniel Boros

12 Sep 2021

JavaScript Functional Style Made Simple

JavaScript Functional Style Made Simple

Daniel Boros

12 Sep 2021


CareersCompaniesSitemapFunctional WorksBlockchain WorksJavaScript WorksAI WorksGolang WorksJava WorksPython WorksRemote Works

Ground Floor, Verse Building, 18 Brunswick Place, London, N1 6DZ

108 E 16th Street, New York, NY 10003

Subscribe to our newsletter

Join over 111,000 others and get access to exclusive content, job opportunities and more!

© 2024 WorksHub

Privacy PolicyDeveloped by WorksHub