Is this how Phoenix.Channel and PubSub should interact?

I came across some interactions between Channels and PubSub yesterday which surprised me. This is using phoenix 1.7.1 and phoenix_pubsub 2.1.1.

If I have a Phoenix.Channel on topic "some_topic and a client connected to it:

  • Phoenix.PubSub.broadcast(MyApp.PubSub, "some_topic", %{some: "message"} will go to that channel process’s handle_info/2 even if it did not call Phoenix.PubSub.subscribe/3, and will go N+1 times if it did explicitly subscribe N times.
  • MyApp.Endpoint.broadcast("some_topic", "some_event", %{some: "message"}) will go directly to the connected client of the channel without the channel process ever seeing it unless the channel process intercepts it.

Is this how it’s intended to work? My mental model was “only processes which explicitly PubSub.subscribe/3 will get a PubSub event, and it will go to their handle_info/2, where they can do what they will.” I see that handle_out/3 exists for interception, so clearly that’s by design…

I’ve been using Phoenix for years and didn’t realize that the topics were shared between Channels and PubSub. But I may not be the sharpest knife in the drawer. :wink:

1 Like

Yes this is working exactly it is supposed to :slight_smile:

Channel broadcasts (broacast(socket, ...) and Endpoint.broadcast )) are “fast-laned” directly to the transport process unless you intercept. This allows us to encode the the message a single time vs N times for every subscriber. Phoenix.PubSub.broadcast is lower level and broadcasts the raw elixir term as is, vs channel broadcast where it is wrapped internally in a Phoenix.Socket.Broadcast struct. As far as your dup subscriptions go, it’s the subscribers job to avoid duplicate subscriptions as there is a performance cost in checking on the subscribe side. Hope that helps!


OK, got it, thanks. For our purposes we want to have the Channel process get the published message and do some standard formatting before it goes out, so we’ll just need to make sure our topics are different for Channels vs PubSub.

the transport process

Is it accurate to say that each Channel process is paired with one transport process?

That’s essentially what intercept would do as well.

Yes, with the difference that if I forget to intercept a particular message, it will go to the client, but if I forget to handle_info one, it will blow up and I’ll notice in development.