Here’s what I came up with. It all seems to work as expected but there’s a lot I’m unsure about:

I added an analysis registry and supervisor:

      {Registry, keys: :unique, name: MyApp.Analysis.Registry},
      {Task.Supervisor, name: MyApp.Analysis.Supervisor}

I wondered briefly whether to namespace them since it seems fairly easy to use them in different contexts, or at least Task.Supervisor, but it seemed like most examples I could find used namespaces by default.

Next I added a new function to my analysis module to handle the “launch or kill” spec

  def request_analysis(%Sources.Source{id: source_id} = source) do
    case Registry.register(Analysis.Registry, "source-#{source_id}-analyzer", :value) do
      {:ok, _} ->
        analyze(source_id)

      {:error, {:already_registered, pid}} ->
        Process.exit(pid, :kill)
        request_analysis(source)
    end

    :ok
  end

I’m not sure I understand the point of the “value” argument in Registry.register/3 and I’m just using a placeholder there. I even considered just using nil. My naive expectation would be that the “value” of this function would naturally be the pid itself. The Registry docs were a bit confusing because they assumed you wanted to register a new process by naming it instead of using register.

I then updated my LiveView to call this function instead of the analsys:

  defp request_reanalysis(socket) do
    Task.Supervisor.async_nolink(
      Analysis.Supervisor,
      Analysis.Analyzer,
      :request_analysis,
      [socket.assigns.source],
      shutdown: :brutal_kill
    )

    socket
    |> assign(analyzer_data: nil)
  end

The analyze function I left mostly intact, aside from adding the PubSub call. It uses Task.async and Task.await to concurrently build the data and compose it when each sub-part is complete. I suppose now that I have introduced Task.Supervisor I should use that given the advice above? But then I’m not sure what the purpose of Task.await would be. Poking around a bit it just seems like people mostly just always use one or the other. The documentation doesn’t seem to focus on the trade offs involved, of which surely there must be some?