Distributed portals with Elixir


Distributed portals with Elixir


This notebook is a fast-paced introduction to the Elixir programming language. We will explore both basic and advanced concepts to implement our own version of the Portal game to transfer data across notebooks using Elixir’s distribution capabilities.

For a more structured introduction to the language, see Elixir’s Getting Started guide and the many learning resources available.

The plan ahead

The Portal game consists of a series of puzzles that must be solved by teleporting the player’s character and simple objects from one place to another.

In order to teleport, the player uses the Portal gun to shoot doors onto flat planes, like a floor or a wall. Entering one of those doors teleports you to the other:

Our version of the Portal game will use Elixir to shoot doors of different colors and transfer data between them! We will even learn how we can distribute doors across different machines in our network:

Here is what we will learn:

  • Elixir’s basic data structures
  • Pattern matching
  • Using agents for state
  • Using structs for custom data structures
  • Extending the language with protocols
  • Supervision trees and applications
  • Distributed Elixir nodes

At the end of this notebook, we will make the following code work:

# Shoot two doors: one orange, another blue

# Start transferring the list [1, 2, 3, 4] from orange to blue
portal = Portal.transfer(:orange, :blue, [1, 2, 3, 4])

# This will output:
#   #Portal<
#          :orange <=> :blue
#     [1, 2, 3, 4] <=> []
#   >

# Now every time we call push_right, data goes to blue

# This will output:
#   #Portal<
#          :orange <=> :blue
#        [1, 2, 3] <=> [4]
#   >

Intrigued? Let’s get started!

Basic data structures

Elixir has numbers, strings, and variables. Code comments start with #:

# Numbers
IO.inspect(40 + 2)

# Strings
variable = "hello" <> " world"

Executing the cell above prints the number 42 and the string "hello world". To do so, we called the function inspect in the IO module, using the IO.inspect(...) notation. This function prints the given data structure to your terminal - in this case, our notebook - and returns the value given to it.

Elixir also has three special values, true, false, and nil. Everything in Elixir is considered to be a truthy value, except for false and nil:

# && is the logical and operator
IO.inspect(true &amp;&amp; true)
IO.inspect(13 &amp;&amp; 42)

# || is the logical or operator
IO.inspect(true || false)
IO.inspect(nil || 42)

For working with collections of data, Elixir has three data types:

# Lists (typically hold a dynamic amount of items)
IO.inspect([1, 2, "three"])

# Tuples (typically hold a fixed amount of items)
IO.inspect({:ok, "value"})

# Maps (key-value data structures)
IO.inspect(%{"key" => "value"})

In the snippet above, we also used a new data structure represented as :ok. All values starting with a leading : in Elixir are called atoms. Atoms are used as identifiers across the language. Common atoms are :ok and :error. Which brings us to the next topic: pattern matching.

Pattern matching

The = operator in Elixir is a bit different from the ones we see in other languages:

x = 1

So far so good, but what happens if we invert the operands?

1 = x

It worked! That’s because Elixir tries to match the right side against the left side. Since both are set to 1, it works. Let’s try something else:

2 = x

Now the sides did not match, so we got an error. We use pattern matching in Elixir to match on collection too. For example, we can use [head | tail] to extract the head (the first element) and tail (the remaining ones) from a list:

[head | tail] = [1, 2, 3]

Matching an empty list against [head | tail] causes a match error:

[head | tail] = []

Finally, we can also use the [head | tail] expression to add elements to the head of a list:

list = [1, 2, 3]
[0 | list]

We can also pattern match on tuples. This is often used to match on the return types of function calls. For example, take the function Date.from_iso8601(string), which returns {:ok, date} if the string represents a valid date, in the format YYYY-MM-DD, otherwise it returns {:error, reason}:

# A valid date
# An invalid date

Now, what happens if we want our code to behave differently depending if the date is valid or not? We can use case to pattern match on the different tuples:

# Give an invalid date as input
input = "2020-02-30"

# And then match on the return value
case Date.from_iso8601(input) do
  {:ok, date} ->
    "We got a valid date: #{inspect(date)}"

  {:error, reason} ->
    "Oh no, the date is invalid. Reason: #{inspect(reason)}"

In this example, we are using case to pattern match on the different outcomes of the Date.from_iso8601. We say the case above has two clauses, one matching on {:ok, date} and another on {:error, reason}. Now try changing the input variable above and reevaluate the cell accordingly. What happens when you give it a valid date?

Finally, we can also pattern match on maps. This is used to extract the values for the given keys:

map = %{:elixir => :functional, :python => :object_oriented}
%{:elixir => type} = map

If the key does not exist on the map, it raises:

%{:c => type} = map

With pattern matching out of the way, we are ready to start our Portal implementation!

Modeling portal doors with Agents

Elixir data structures are immutable. In the examples above, we never mutated the list. We can break a list apart or add new elements to the head, but the original list is never modified.

That said, when we need to keep some sort of state, like the data transferring through a portal, we must use an abstraction that stores this state for us. One such abstraction in Elixir is called an agent. Before we use agents, we need to briefly talk about anonymous functions.

Anonymous functions are a mechanism to represent pieces of code that we can pass around and execute later on:

adder = fn a, b -> a + b end

An anonymous function is delimited by the words fn and end and an arrow -> is used to separate the arguments from the anonymous function body. We can now call the anonymous function above multiple times by providing two numbers as inputs:

adder.(1, 2)
adder.(3, 5)

In Elixir, we also use anonymous functions to initialize, get, and update the agent state:

{:ok, agent} = Agent.start_link(fn -> [] end)

In the example above, we created a new agent, passing a function that returns the initial state of an empty list. The agent returns {:ok, #PID<...>}, where PID stands for a process identifier, which uniquely identifies the agent. Elixir has many abstractions for concurrency, such as agents, tasks, generic servers, but at the end of the day they all boil down to processes. When we say processes in Elixir, we don’t mean Operating System processes, but rather Elixir Processes, which are lightweight and isolated, allowing us to run hundreds of thousands of them on the same machine.

We store the agent’s PID in the agent variable, which allows us to send messages to get the agent’s state:

Agent.get(agent, fn list -> list end)

As well as update it before reading again:

Agent.update(agent, fn list -> [0 | list] end)
Agent.get(agent, fn list -> list end)

We will use agents to implement our portal doors.

Whenever we need to encapsulate logic in Elixir, we create modules, which are essentially a collection of functions. We define modules with defmodule and functions with def. Our functions will encapsulate the logic to interact with the agent, using the API we learned in the cells above:

defmodule Portal.Door do
  use Agent

  def start_link(color) when is_atom(color) do
    Agent.start_link(fn -> [] end, name: color)

  def get(door) do
    Agent.get(door, fn list -> list end)

  def push(door, value) do
    Agent.update(door, fn list -> [value | list] end)

  def pop(door) do
    Agent.get_and_update(door, fn list ->
      case list do
        [h | t] -> {{:ok, h}, t}
        [] -> {:error, []}

  def stop(door) do

We declare a module by giving it a name, in this case, Portal.Door. At the top of the module, we say use Agent, which brings some Agent-related functionality into the module.

The first function is start_link, which we often refer to as start_link/1, where the number 1 is called the “arity” of the function and it indicates the number of arguments it receives. Then we check that the argument is an atom and proceed to call Agent.start_link/2, as we did earlier in this section, except we are now passing name: color as an argument. By giving a name to the Agent, we can refer to it anywhere by its name, instead of using its PID.

The next two functions, get/1 and push/2 perform simple operation to the agent, reading its state and adding a new element respectively. Let’s take a look at them:


Note how we didn’t need to store the PID anywhere and we can use the atom :pink to refer to the door and read its state. If the door already exists, and we try to start another one with the same name, it returns an {:error, reason} tuple instead of {:ok, pid}:


Next, let’s push some events:

Portal.Door.push(:pink, 1)
Portal.Door.push(:pink, 2)

We pushed some events and they show up in our state. Although, note they appear in reverse order. That’s because we are always adding new entries to the top of the list.

The fourth function we defined is called pop/1. If there is any item in the agent, it takes the head of the list and returns it wrapped in a {:ok, value} tuple. However, if the list is empty, it returns :error.



Finally, the last function, stop/1, simply terminates the agent, effectively closing the door. Let’s try it:


Now, if we try to do anything with it, it will raise:


Note the error message points out why the operation did not work, great!

Portal transfers

Our portal doors are ready so it is time to start working on portal transfers! In order to store the portal data, we are going to create a struct named Portal. Let’s first learn what structs are about.

Structs define data structures with pre-defined keys. The keys are verified at compilation time, so if you make a typo in the key name, you get an error early on. Structs are defined inside modules, by calling the defstruct with a list of atom keys. Let’s define a User struct with the fields :name and :age:

defmodule User do
  defstruct [:name, :age]

Now, we can create structs using the %User{...} notation:

user = %User{name: "john doe", age: 27}

We can access struct fields using the struct.field syntax:

We can pattern match on structs too:

%User{age: age} = user

Finally, let’s see what happens if we do a typo in a field:

%User{agee: age} = user

Now we are ready to define our Portal struct. It will have two fields, :left and :right, which point respectively to the portal door on the left and the door on the right. Our goal is to transfer data from the left door to the right one. The Portal module, where we define our struct, will also have four other functions:

  • shoot(color) - shoots a door of the given color. This is a wrapper around Portal.Door.start_link/1

  • transfer(left_door, right_door, data) - starts a transfer by loading the given data to left_door and returns a Portal struct

  • push_right(portal) - receives a portal and continues the transfer by pushing data from the left to the right

  • close(portal) - closes the portal by explicitly stopping both doors

Let’s implement them:

defmodule Portal do
  defstruct [:left, :right]

  def shoot(color) do

  def transfer(left_door, right_door, data) do
    # First add all data to the portal on the left
    for item <- data do
      Portal.Door.push(left_door, item)

    # Returns a portal struct with the doors
    %Portal{left: left_door, right: right_door}

  def push_right(portal) do
    # See if we can pop data from left. If so, push the
    # popped data to the right. Otherwise, do nothing.
    case Portal.Door.pop(portal.left) do
      :error   -> :ok
      {:ok, h} -> Portal.Door.push(portal.right, h)

    # Let's return the portal itself

  def close(portal) do

The Portal modules defines a struct followed by a shoot/1 function. The function is just a wrapper around Portal.Door.start_link/1. Then we define the transfer/3 function, which loads the given data into the left door and returns a Portal struct. Finally, push_right/3 gets data from the door on the left and puts it on the right door. Let’s give it a try:

portal = Portal.transfer(:orange, :blue, [1, 2, 3])

The above returns the %Portal{} struct. We can check the data has been loaded into the left door:


Note the list is reversed - and we knew that! - as we always add items on the top. But we will use that to our advantage soon. Let’s start pushing data to the right:


Since the list is reversed, we can see that we pushed the number 3 to the right, which is exactly what we expected. If you reevaluate the cell above, you will see data moving to the right, as our portal doors are stateful.

Our portal transfer seems to work as expected! Now let’s clean up and close the transfer:


We have made some good progress in our implementation, so now let’s work a bit on the presentation. Currently, the Portal is printed as a struct: %Portal{left: :orange, right: :blue}. It would be nice if we actually had a printed representation of the portal transfer, allowing us to see the portal processes as we push data.

Inspecting portals with Protocols

We already know that Elixir data structures data can be printed by Livebook. After all, when we type 1 + 2, we get 3 back. However, can we customize how our own data structures are printed?

Yes, we can! Elixir provides protocols, which allows behaviour to be extended and implemented for any data type, like our Portal struct, at any time.

For example, every time something is printed in Livebook, or in Elixir’s terminal, Elixir uses the Inspect protocol. Since protocols can be extended at any time, by any data type, it means we can implement it for Portal too. We do so by calling defimpl/2, passing the protocol name and the data structure we want to implement the protocol for. Let’s do it:

defimpl Inspect, for: Portal do
  def inspect(%Portal{left: left, right: right}, _) do
    left_door  = inspect(left)
    right_door = inspect(right)

    left_data  = inspect(Enum.reverse(Portal.Door.get(left)))
    right_data = inspect(Portal.Door.get(right))

    max = max(String.length(left_door), String.length(left_data))

      #{String.pad_leading(left_door, max)} <=> #{right_door}
      #{String.pad_leading(left_data, max)} <=> #{right_data}

In the snippet above, we have implemented the Inspect protocol for the Portal struct. The protocol expects one function named inspect to be implemented. The function expects two arguments, the first is the Portal struct itself and the second is a set of options, which we don’t care about for now.

Then we call inspect multiple times, to get a text representation of both left and right doors, as well as to get a representation of the data inside the doors. Finally, we return a string containing the portal presentation properly aligned.

That’s all we need! Let’s start a new transfer and see how it goes:

portal = Portal.transfer(:red, :purple, [1, 2, 3])

Sweet! Look how Livebook automatically picked up the new representation. Now feel free to call push_right and see what happens:


Feel free to reevaluate the cell above a couple times. Once you are done, run the cell below to clean it all up and close the portal:


There is just one topic left…

Distributed transfers

With our portals working, we are ready to give distributed transfers a try. However, before we start, there is one big disclaimer:

> The feature we are going to implement will allow us to share data across > two separate notebooks using the Erlang Distribution. This section is a great > experiment to understand how things work behind the scenes but sharing data > across notebooks as done here is a bad idea in practice. If this topic > interests you, we recommend picking up one of the many available learning > resources available for Elixir, which will provide a more solid ground to > leverage the Erlang VM and its distributed features.

Distribution 101

When Livebook executes the code in a notebook, it starts a separate Elixir runtime to do so. Since Livebook itself is also written in Elixir, it uses the Erlang Distribution to communicate with this Elixir runtime. We can get the name of the Elixir node our notebook is running on like this:


By executing the code above, we can see node names are atoms. By default, Livebook only connects to nodes running on the same machine, but you can also configure it to connect to runtimes across machines.

We can also get a list of all nodes our runtime is connected to by calling Node.list(), let’s give it a try:


If you execute the code above, you get… an empty list!?

The reason why we get an empty list is because, by default, Erlang Distribution is a fully mesh. This means all nodes can see all nodes in the network. However, because we want the notebook runtimes to be isolated from each other, we start each runtime as a hidden node. We can ask Elixir to give us all hidden nodes instead:


Much better! We see one node, which is the Livebook server itself.

Now there is one last piece of the puzzle: in order for nodes to connect to each other, they need to have the same cookie. The cookie serves as a very simple authentication mechanism. We can read the cookie of our current notebook runtime like this:


Now we have everything we need to connect across notebooks.

Notebook connections

In order to connect across notebooks, open up a new empty notebook in a separate tab, copy and paste the code below to this new notebook, and execute it:

IO.inspect node()
IO.inspect Node.get_cookie()

Now paste the result of the other node name and its cookie in the variables below:

other_node = :"name-of-the@other-node"
other_cookie = :"value-of-the-other-cookie"

With both variables defined, let’s connect to the other node with the given cookie:

Node.set_cookie(other_node, other_cookie)

If it returns true, it means it connected as expected and we are ready to start a distributed transfer. The first step is to go to the other notebook and shoot a door. You can try calling the following there:


However, if you try the above, it will fail! This happens because the Portal code has been defined in this notebook but it is not available in the other notebook. If we were working on an actual Elixir project, this issue wouldn’t exist, because we would start multiple nodes on top of the same codebase with the same modules, but we can’t do so here. To work around this, copy the cell that defines the Portal.Door module from this notebook into the other notebook and execute it. Now in the other node you should be able to start a door:


Cross-node references

Now that we have spawned a door on the other notebook, we can directly read its content from this notebook. So far, we have been using atoms to represent doors, such as :blue, but we can also use the {name, node} notation to refer to a process in another node. Let’s give it a try:

blue = {:blue, other_node}

It works! We could successfully read something from the other node. Now, let’s shoot a yellow portal on this node and start a transfer between them:

yellow = {:yellow, node()}
portal = Portal.transfer(yellow, blue, [1, 2, 3, 4])

Our distributed transfer was started. Now let’s push right:


If you go back to the other notebook and run Portal.Door.get(:blue), you should see it has been updated with entries from this notebook! The best part of all is that we enabled distributed transfers without changing a single line of code!

Our distributed portal transfer works because the doors are just processes and accessing/pushing the data through doors is done by sending messages to those processes via the Agent API. We say sending a message in Elixir is location transparent: we can send messages to any PID regardless if it is in the same node as the sender or in different nodes of the same network.

Wrapping up

So we have reached the end of this notebook with a fast paced introduction to Elixir! It was a fun ride and we went from manually starting doors to distributed portal transfers.

To learn more about Elixir, we welcome you to explore our website, read our Getting Started guide, and many of the available learning resources.

You can also learn how to use some of Elixir and Livebook’s unique features together in the Elixir and Livebook notebook.

Finally, huge thanks to Augie De Blieck Jr. for the drawings in this tutorial.

See you around!