I’ve been thinking about this a bit and come up with something I like, so I’m gonna answer my own question here in case other people run into this.
So, there’s 2 things to know. This is demand driven and I’m trying to utilise backpressure to avoid a build up of data when the next step in the flow is busy. By it’s very nature, if C is busy and E has run out of stuff to process, E can’t demand more data. Because there’s no way of knowing if B will produce data for C or E, B asking for demand could just add a build up of data on C even though C is trying to apply backpressure to avoid that in the first place. Put simpler: you can’t ask an API (B) for errors (E) 
So, C and E may as well be the same step so to speak. There’s no branch, at least ask far as Flow is concerned.
Then how does the right data get to C and E? This is the second thing. I make B output “tagged tuples” like {:ok, data} and {:error, data}. This way C and E can pattern match to receive the data relevant for them. As long as they include a “pass through” default pattern match, the data for later steps flows down. So if you arrange the steps A → B → C → E with C passing through (or “ignoring”) {:error, data}s, then the data makes it to E. Put D → F on the end and ensure they also passthrough and you’re done.
It looks something like this (removing partitioning and simplifying to all maps):
Flow.from_enumerable(A.data)
|> Flow.map(&B.do_stuff/1)
|> Flow.map(&C.do_stuff/1)
|> Flow.map(&E.do_stuff/1)
|> Flow.map(&D.do_stuff/1)
|> Flow.map(&F.do_stuff/1)
|> Flow.run
# Example passthrough work module
defmodule C do
def do_stuff({:ok, data}) do
# actually do stuff
end
# passthrough stuff that isn't for C (ie, errors)
def do_stuff(other_data_passthrough), do: other_data_passthrough
end
There’s also different ways of arranging the overlapping steps. Like, if it’s really important that errors propagate to your services quickly, you’d arrange it A → B → E → F → C → D so that C and D can’t hold up E and F. And vice versa.
I’m starting to wonder if there’s a way of abstracting this structure up into Flow. Like extending map and it’s friends to take an atom, and Flow handles the pattern matching and passthrough for you. Hmm.
