Defines a Phoenix Channel.

Channels provide a means for bidirectional communication from clients that integrate with the Phoenix.PubSub layer for soft-realtime functionality.

Topics & Callbacks

Every time you join a channel, you need to choose which particular topic you want to listen to. The topic is just an identifier, but by convention it is often made of two parts: "topic:subtopic". Using the "topic:subtopic" approach pairs nicely with the Phoenix.Socket.channel/3 allowing you to match on all topics starting with a given prefix by using a splat (the * character) as the last character in the topic pattern:

channel "room:*", MyApp.RoomChannel

Any topic coming into the router with the "room:" prefix would dispatch to MyApp.RoomChannel in the above example. Topics can also be pattern matched in your channels' join/3 callback to pluck out the scoped pattern:

# handles the special `"lobby"` subtopic
def join("room:lobby", _payload, socket) do
  {:ok, socket}
end

# handles any other subtopic as the room ID, for example `"room:12"`, `"room:34"`
def join("room:" <> room_id, _payload, socket) do
  {:ok, socket}
end

Authorization

Clients must join a channel to send and receive PubSub events on that channel. Your channels must implement a join/3 callback that authorizes the socket for the given topic. For example, you could check if the user is allowed to join that particular room.

To authorize a socket in join/3, return {:ok, socket}. To refuse authorization in join/3, return {:error, reply}.

Incoming Events

After a client has successfully joined a channel, incoming events from the client are routed through the channel's handle_in/3 callbacks. Within these callbacks, you can perform any action. Typically you'll either forward a message to all listeners with broadcast!/3, or push a message directly down the socket with push/3. Incoming callbacks must return the socket to maintain ephemeral state.

Here's an example of receiving an incoming "new_msg" event from one client, and broadcasting the message to all topic subscribers for this socket.

def handle_in("new_msg", %{"uid" => uid, "body" => body}, socket) do
  broadcast!(socket, "new_msg", %{uid: uid, body: body})
  {:noreply, socket}
end

General message payloads are received as maps, and binary data payloads are passed as a {:binary, data} tuple:

def handle_in("file_chunk", {:binary, chunk}, socket) do
  ...
  {:reply, :ok, socket}
end

You can also push a message directly down the socket, in the form of a map, or a tagged {:binary, data} tuple:

# client asks for their current rank, push sent directly as a new event.
def handle_in("current_rank", _, socket) do
  push(socket, "current_rank", %{val: Game.get_rank(socket.assigns[:user])})
  push(socket, "photo", {:binary, File.read!(socket.assigns.photo_path)})
  {:noreply, socket}
end

Replies

In addition to pushing messages out when you receive a handle_in event, you can also reply directly to a client event for request/response style messaging. This is useful when a client must know the result of an operation or to simply ack messages.

For example, imagine creating a resource and replying with the created record:

def handle_in("create:post", attrs, socket) do
  changeset = Post.changeset(%Post{}, attrs)

  if changeset.valid? do
    post = Repo.insert!(changeset)
    response = MyApp.PostView.render("show.json", %{post: post})
    {:reply, {:ok, response}, socket}
  else
    response = MyApp.ChangesetView.render("errors.json", %{changeset: changeset})
    {:reply, {:error, response}, socket}
  end
end

Alternatively, you may just want to ack the status of the operation:

def handle_in("create:post", attrs, socket) do
  changeset = Post.changeset(%Post{}, attrs)

  if changeset.valid? do
    Repo.insert!(changeset)
    {:reply, :ok, socket}
  else
    {:reply, :error, socket}
  end
end

Like binary pushes, binary data is also supported with replies via a {:binary, data} tuple:

{:reply, {:ok, {:binary, bin}}, socket}

Intercepting Outgoing Events

When an event is broadcasted with broadcast/3, each channel subscriber can choose to intercept the event and have their handle_out/3 callback triggered. This allows the event's payload to be customized on a socket by socket basis to append extra information, or conditionally filter the message from being delivered. If the event is not intercepted with Phoenix.Channel.intercept/1, then the message is pushed directly to the client:

intercept ["new_msg", "user_joined"]

# for every socket subscribing to this topic, append an `is_editable`
# value for client metadata.
def handle_out("new_msg", msg, socket) do
  push(socket, "new_msg", Map.merge(msg,
    %{is_editable: User.can_edit_message?(socket.assigns[:user], msg)}
  ))
  {:noreply, socket}
end

# do not send broadcasted `"user_joined"` events if this socket's user
# is ignoring the user who joined.
def handle_out("user_joined", msg, socket) do
  unless User.ignoring?(socket.assigns[:user], msg.user_id) do
    push(socket, "user_joined", msg)
  end
  {:noreply, socket}
end

Broadcasting to an external topic

In some cases, you will want to broadcast messages without the context of a socket. This could be for broadcasting from within your channel to an external topic, or broadcasting from elsewhere in your application like a controller or another process. Such can be done via your endpoint:

# within channel
def handle_in("new_msg", %{"uid" => uid, "body" => body}, socket) do
  ...
  broadcast_from!(socket, "new_msg", %{uid: uid, body: body})
  MyApp.Endpoint.broadcast_from!(self(), "room:superadmin",
    "new_msg", %{uid: uid, body: body})
  {:noreply, socket}
end

# within controller
def create(conn, params) do
  ...
  MyApp.Endpoint.broadcast!("room:" <> rid, "new_msg", %{uid: uid, body: body})
  MyApp.Endpoint.broadcast!("room:superadmin", "new_msg", %{uid: uid, body: body})
  redirect(conn, to: "/")
end

Terminate

On termination, the channel callback terminate/2 will be invoked with the error reason and the socket.

If we are terminating because the client left, the reason will be {:shutdown, :left}. Similarly, if we are terminating because the client connection was closed, the reason will be {:shutdown, :closed}.

If any of the callbacks return a :stop tuple, it will also trigger terminate with the reason given in the tuple.

terminate/2, however, won't be invoked in case of errors nor in case of exits. This is the same behaviour as you find in Elixir abstractions like GenServer and others. Typically speaking, if you want to clean something up, it is better to monitor your channel process and do the clean up from another process. Similar to GenServer, it would also be possible :trap_exit to guarantee that terminate/2 is invoked. This practice is not encouraged though.

Exit reasons when stopping a channel

When the channel callbacks return a :stop tuple, such as:

{:stop, :shutdown, socket}
{:stop, {:error, :enoent}, socket}

the second argument is the exit reason, which follows the same behaviour as standard GenServer exits.

You have three options to choose from when shutting down a channel:

• :normal - in such cases, the exit won't be logged and linked processes do not exit

• :shutdown or {:shutdown, term} - in such cases, the exit won't be logged and linked processes exit with the same reason unless they're trapping exits

• any other term - in such cases, the exit will be logged and linked processes exit with the same reason unless they're trapping exits

Subscribing to external topics

Sometimes you may need to programmatically subscribe a socket to external topics in addition to the internal socket.topic. For example, imagine you have a bidding system where a remote client dynamically sets preferences on products they want to receive bidding notifications on. Instead of requiring a unique channel process and topic per preference, a more efficient and simple approach would be to subscribe a single channel to relevant notifications via your endpoint. For example:

defmodule MyApp.Endpoint.NotificationChannel do
  use Phoenix.Channel

  def join("notification:" <> user_id, %{"ids" => ids}, socket) do
    topics = for product_id <- ids, do: "product:#{product_id}"

    {:ok, socket
          |> assign(:topics, [])
          |> put_new_topics(topics)}
  end

  def handle_in("watch", %{"product_id" => id}, socket) do
    {:reply, :ok, put_new_topics(socket, ["product:#{id}"])}
  end

  def handle_in("unwatch", %{"product_id" => id}, socket) do
    {:reply, :ok, MyApp.Endpoint.unsubscribe("product:#{id}")}
  end

  defp put_new_topics(socket, topics) do
    Enum.reduce(topics, socket, fn topic, acc ->
      topics = acc.assigns.topics
      if topic in topics do
        acc
      else
        :ok = MyApp.Endpoint.subscribe(topic)
        assign(acc, :topics, [topic | topics])
      end
    end)
  end
end

Note: the caller must be responsible for preventing duplicate subscriptions. After calling subscribe/1 from your endpoint, the same flow applies to handling regular Elixir messages within your channel. Most often, you'll simply relay the %Phoenix.Socket.Broadcast{} event and payload:

alias Phoenix.Socket.Broadcast
def handle_info(%Broadcast{topic: _, event: event, payload: payload}, socket) do
  push(socket, event, payload)
  {:noreply, socket}
end

Hibernation

From Erlang/OTP 20, channels automatically hibernate to save memory after 15_000 milliseconds of inactivity. This can be customized by passing the :hibernate_after option to use Phoenix.Channel:

use Phoenix.Channel, hibernate_after: 60_000

You can also set it to :infinity to fully disable it.

Shutdown

You can configure the shutdown of each channel used when your application is shutting down by setting the :shutdown value on use:

use Phoenix.Channel, shutdown: 5_000

It defaults to 5_000.

Logging

By default, channel "join" and "handle_in" events are logged, using the level :info and :debug, respectively. Logs can be customized per event type or disabled by setting the :log_join and :log_handle_in options when using Phoenix.Channel. For example, the following configuration logs join events as :info, but disables logging for incoming events:

use Phoenix.Channel, log_join: :info, log_handle_in: false