Processes

A Rails engine runs inside requests; a mountable library runs on the BEAM, so it can keep state, schedule work and talk to other nodes without Redis or a job queue.

Children of a mount

@impl PhoenixMountable
def children(mount) do
  [
    {MyLib.Cache, name: PhoenixMountable.via(mount, :cache)},
    {MyLib.Poller, mount: mount}
  ]
end

Each mount gets a supervisor of its own, so two mounts of the library never share state, and a crash stays inside the library's subtree.

By default the processes start the first time the mount is used: a request or a LiveView. That suits caches and per-user state. A host that needs them from boot (a scheduler, a poller) adds the supervisor to its application, after its endpoint:

children = [
  MyApp.Repo,
  MyAppWeb.Endpoint,
  {PhoenixMountable.Supervisor, router: MyAppWeb.Router, endpoint: MyAppWeb.Endpoint}
]

The router knows its mounts at compile time, so nothing else is listed.

PubSub

PhoenixMountable.PubSub.subscribe(socket.assigns.mount, "items")
PhoenixMountable.PubSub.broadcast(mount, "items", {:item_updated, item})

Topics are scoped to the mount and travel across the cluster. The server is the mount's :pubsub option, or the endpoint's.

Once per cluster

def children(mount) do
  [
    {PhoenixMountable.Singleton,
     name: {:scheduler, PhoenixMountable.Mount.key(mount)},
     child: {MyLib.Scheduler, mount: mount}}
  ]
end

Every node runs a PhoenixMountable.Singleton; the one that registers the name globally starts the child, and the others take over when its node goes down.

Telemetry

PhoenixMountable emits events for mounts, processes, stale assets, migrations and backfills. See PhoenixMountable.Telemetry.