Use correct scheduler address to connect to when using interface keyword

[guillaumeeb]

Closes #126.

I'm not completely convinced about the solution here. The thing is, we need to keep binding to all network interface, at least for Bokeh Dashboard, as demanded here: #90. Keeping listening to all network interface for distributed scheduler is nice too I think, but not mandatory.

So another solution from what is proposed here could be to go with a solution as proposed by @mrocklin in #90 (comment), and stick to the use of self.scheduler.address as a URL to point workers to.

Any advice on this welcome.

[lesteve]

So another solution from what is proposed here could be to go with a solution as proposed by @mrocklin in #90 (comment), and stick to the use of self.scheduler.address as a URL to point workers to.

@guillaumeeb I opened #135 which is closer to the spirit of #90 (comment). I am not sure this is a better option to be honest, let me know if you have any thoughts on it!

Side-comment: it'd be great to have tests for this but I don't really know how hard it is to have a IP address setup in docker containers that looks similar to the one of a typical cluster ... edit: I guess having two different network interfaces as well may be useful too for the problem reported in #126.

[guillaumeeb]

No idea for more complete testing, there is some though already in test_jobqueue_core.py.

Closing this one in favor of #135 that I found cleaner.