Clustering Your Application

Elixir and the BEAM have the incredible ability to be clustered together and processes can pass messages seamlessly to each other between nodes. Fly makes clustering easy! This extra (and totally optional) portion of the guide walks you through clustering your Elixir application.

There are 2 parts to getting clustering quickly setup on Fly.

Installing and using libcluster
Scaling our application to multiple VMs

Adding `libcluster`

The widely adopted library libcluster helps here.

Libcluster supports multiple strategies for finding and connecting with other nodes. The strategy we'll use is DNSPoll which was added in version 3.2.2 of libcluster, so make sure you're using that version or newer.

After installing libcluster, add it to the application like this:

defmodule HelloElixir.Application do
  use Application

  def start(_type, _args) do
    topologies = Application.get_env(:libcluster, :topologies) || []

    children = [
      # ...
      # setup for clustering
      {Cluster.Supervisor, [topologies, [name: HelloElixir.ClusterSupervisor]]}
    ]

    # ...
  end

  # ...
end

Our next step is to add the topologies configuration to the file config/runtime.exs.

  app_name =
    System.get_env("FLY_APP_NAME") ||
      raise "FLY_APP_NAME not available"

  config :libcluster,
    debug: true,
    topologies: [
      fly6pn: [
        strategy: Cluster.Strategy.DNSPoll,
        config: [
          polling_interval: 5_000,
          query: "#{app_name}.internal",
          node_basename: app_name
        ]
      ]
    ]

REMEMBER: Deploy your updated app so the clustering code is available, with fly deploy.

This configures libcluster to use the DNSPoll strategy and look for other deployed apps using the same $FLY_APP_NAME on the .internal private network.

This assumes that your rel/env.sh.eex file is configured to name your Elixir node using the $FLY_APP_NAME. We did this earlier in the "Naming Your Elixir Node" section.

Before this app can be clustered, we need more than one VM. We'll do that next!

Running Multiple VMs

There are two ways to run multiple VMs.

Scale our application to have multiple VMs in one region.
Add a VM to another region (multiple regions).

Both approaches are valid and our Elixir application doesn't change at all for the approach you choose!

Let's first start with a baseline of our single deployment.

fly status

...
VMs
ID       VERSION REGION DESIRED STATUS  HEALTH CHECKS      RESTARTS CREATED
f9014bf7 26      sea    run     running 1 total, 1 passing 0        1h8m ago

Scaling in a Single Region

Let's scale up to 2 VMs in our current region.

fly scale count 2

Count changed to 2

Checking on the status we can see what happened.

fly status

...
VMs
ID       VERSION REGION DESIRED STATUS  HEALTH CHECKS      RESTARTS CREATED
eb4119d3 27      sea    run     running 1 total, 1 passing 0        39s ago
f9014bf7 27      sea    run     running 1 total, 1 passing 0        1h13m ago

We now have two VMs in the same region! That was easy.

Let's make sure they are clustered together. We can check the logs:

fly logs

...
app[eb4119d3] sea [info] 21:50:21.924 [info] [libcluster:fly6pn] connected to :"fly-elixir@fdaa:0:1da8:a7b:ac2:f901:4bf7:2"
...

But that's not as rewarding as seeing it from inside a node. From an IEx shell, we can ask the node we're connected to, what other nodes it can see.

fly ssh console
/app/bin/hello_elixir remote

iex(fly-elixir@fdaa:0:1da8:a7b:ac2:f901:4bf7:2)1> Node.list
[:"fly-elixir@fdaa:0:1da8:a7b:ac4:eb41:19d3:2"]

I included the IEx prompt because it shows the IP address of the node I'm connected to. Then getting the Node.list returns the other node. Our two VMs are connected and clustered!

Scaling to Multiple Regions

Fly makes it super easy to run VMs of your applications physically closer to your users. Through the magic of DNS, users are directed to the nearest region where your application is located. You can read about regions here and see the list of regions to choose from.

Starting back from our baseline of a single VM running in sea which is Seattle, Washington (US), I'll add the region ewr which is Parsippany, NJ (US). This puts a VM on both coasts of the US.

fly regions add ewr

Region Pool:
ewr
sea
Backup Region:
iad
lax
sjc
vin

Looking at the status right now shows that we're only in 1 region because our count is set to 1.

fly status

...
VMs
ID       VERSION REGION DESIRED STATUS  HEALTH CHECKS      RESTARTS CREATED
cdf6c422 29      sea    run     running 1 total, 1 passing 0        58s ago

Let's add a 2nd VM and see it deploy to ewr.

fly scale count 2

Count changed to 2

Now our status shows we have two VMs spread across 2 regions!

fly status

...
VMs
ID       VERSION REGION DESIRED STATUS  HEALTH CHECKS      RESTARTS CREATED
0a8e6666 30      ewr    run     running 1 total, 1 passing 0        16s ago
cdf6c422 30      sea    run     running 1 total, 1 passing 0        6m47s ago

Let's ensure they are clustered together.

fly ssh console
/app/bin/hello_elixir remote

iex(fly-elixir@fdaa:0:1da8:a7b:ac2:cdf6:c422:2)1> Node.list
[:"fly-elixir@fdaa:0:1da8:a7b:ab2:a8e:6666:2"]

We have two VMs of our application deployed to the West and East coasts of the North American continent and they are clustered together! Our users will automatically be directed to the server nearest them. That is so cool!

Before two Elixir nodes can cluster together, they must share a secret cookie. The cookie itself isn't meant to be a super secret encryption key or anything like that, it's designed to let us create multiple sets of small clusters on the same network that don't all just connect together. Different cookies means different clusters. For instance, only the nodes that all use the cookie "abc" will connect together.

For us, this means that in order for my_remote node to connect to the cluster on Fly, we need to share the same cookie value used in production.

When we build a mix release, it generates a long random string for the cookie value. When we re-run the mix release command, it keeps the same cookie value. That is, when we don't run it in Docker. The Dockerfile we're using is building a fresh release every time we run it. That's kind of the point of a Docker container. So our cookie value is being randomly generated every time we deploy. This means after every deploy, we would have to figure out what the new cookie value is so our local node can use it.

The easiest solution here is to specify the value to use for our cookie. One that we will know outside of the build and that won't keep changing on us.

If we read the Mix.Tasks.Release docs, in the :cookie section we learn that if we provide an ENV named RELEASE_COOKIE, it will be used. If that ENV is not found, it falls back to the randomly generated one.

To do this, we can create the cookie we want and store it in our fly.toml file like this:

[env]
  RELEASE_COOKIE = "my-app-cookie"

Also from the docs, we can generate the cookie string to use with this Elixir command:

Base.url_encode64(:crypto.strong_rand_bytes(40))

After deploying the application, we can verify that the cookie is being used by getting an IEx shell into our running server and issuing the following command:

Node.get_cookie()

This shows the cookie being used at runtime.

With a known and unchanging cookie deployed in our application, we are ready for the next step!

Edit this page on GitHub

Clustering Your Application

Adding libcluster

Running Multiple VMs

Scaling in a Single Region

Scaling to Multiple Regions

The Cookie Situation

The Cookie Problem

The Cookie Solution

Making the Cookie Changes

Adding `libcluster`