phoenixframework/flame

Imagine if we could auto scale simply by wrapping any existing app code in a function and have that block of code run in a temporary copy of the app.

Enter the FLAME pattern.

FLAME - Fleeting Lambda Application for Modular Execution

With FLAME, you treat your entire application as a lambda, where modular parts can be executed on short-lived infrastructure.

Check the screencast to see it in action:

Setup

First add :flame as a dependency:

defp deps do
  [
    # For Erlang/OTP 26 and earlier, you also need Jason
    # {:jason, ">= 0.0.0"},
    {:flame, "~> 0.5"}
  ]
end

Then start a FLAME pool in your supervision tree, typically on application.ex. The example below uses Fly.io's backend:

children = [
  {FLAME.Pool,
   name: MyApp.SamplePool,
   backend: FLAME.FlyBackend,
   min: 0,
   max: 10,
   max_concurrency: 5,
   idle_shutdown_after: 30_000,
   log: :debug}
]

Now you can wrap any block of code in a FLAME.call and it will find or boot a copy of the app, execute the work there, and return the results:

def generate_thumbnails(%Video{} = vid, interval) do
  FLAME.call(MyApp.FFMpegRunner, fn ->
    # I'm runner on a short-lived, temporary server
    tmp_dir = Path.join(System.tmp_dir!(), Ecto.UUID.generate())
    File.mkdir!(tmp_dir)
    System.cmd("ffmpeg", ~w(-i #{vid.url} -vf fps=1/#{interval} #{tmp_dir}/%02d.png))
    urls = VideoStore.put_thumbnails(vid, Path.wildcard(tmp_dir <> "/*.png"))
    Repo.insert_all(Thumbnail, Enum.map(urls, &%{video_id: vid.id, url: &1}))
  end)
end

Here we wrapped up our CPU expensive ffmpeg operation in a FLAME.call/2. FLAME accepts a function and any variables that the function closes over. In this example, the %Video{} struct and interval are passed along automatically. The work happens in a temporary copy of the app. We can do any work inside the FLAME call because we are running the entire application, database connection(s) and all.

FLAME provides the following interfaces for elastically scaled operations:

• FLAME.call/3 - used for synchronous calls
• FLAME.cast/3 - used for async casts where you don't need to wait on the results
• FLAME.place_child/3 – used for placing a child spec somewhere to run, in place of DynamicSupervisor.start_child, Task.Supervisor.start_child, etc

The FLAME.Pool handles elastically scaling runners up and down, as well as remote monitoring of resources. Check the moduledoc for example usage.