fanchat as unit test

[glyph]

this isolates the problem away from the full integration stack,
because we are not touching Twisted at all here.

[glyph]

The problem here might actually be more fundamental than a bug.

fanchat has an explicit flow-control cycle in it: the participant's router feeds into all the chat rooms that they're participating in; in turn, each chat room broadcasts back into the participant's router.

The router is itself a fan.Out underneath, and conceptually as well, because the router could produce output to all of its downstreams in response to one input from its upstream.

We don't get an infinite loop because Pauser explicitly prevents reentrancy of pausing, but once any of these have been paused, it will remain paused forever.

It might be nice to have some kind of cycle detection so that you get an error if you try to set this up without appropriate precautions.

The right solution here would probably be to implement backpressure policy, specifically #37. If we had a Roll somewhere in the cycle, with the ability to buffer even a single element, that ought to un-stick the pause loop, at least at the start. I will need to think a bit about what should happen when that buffer fills up though...

[glyph]

OK, we could also just eliminate the cycle, which is apparently unnecessary and redundant.

I have no idea why this shows coverage dropping. We should really switch to codecov, which seems to have fewer random bugs.

[glyph]

Reviewed 3 of 4 files at r1, 1 of 1 files at r2, 4 of 4 files at r3.
Review status: all files reviewed at latest revision, all discussions resolved.

---

Comments from the review on Reviewable.io