Added the System trait

[kvark]

Closes #30

The new code somehow introduces a deadlock in the tests. AFAIK, all I did was changing a pulse/barrier into a channel.

[kvark]

Weird, if I move the same test code into the example, it works.
@csherratt do you see anything wrong here?

[ghost]

The test case that is failing is test_panics. I probably should have used something like timebomb to check for tests like this.
The test case is checking to see if the task panicking propagates to the dispatch thread. With pulse a pulse would be dropped forcing it into an error state. This would propagate to anything waiting on the pulse.
My guess is this is not the behavior of mpsc, it probably only enters an error state if all produces have been hung up. Which can't happen since you hold your own sender side.

[kvark]

@csherratt thanks for figuring this out! So what can be done to fix it nicely then?

[ghost]

Easiest way, add a drop guard to do the send in the case of a task panicking. This is the same technique the threadpool is doing internally. Well they spawn a thread, you would transmit the message.

https://github.com/frewsxcv/rust-threadpool/blob/master/src/lib.rs#L29-L60

[kvark]

@csherratt the more I look at threadpool the more I realize that we are essentially building on top of it without actually hooking up to the same internal mechanisms. We'd benefit from tighter integration. so perhaps we should just hand-craft something similar manually.

[ghost]

@kvark I don't disagree on the threadpool Their implementation is pretty much wait on a mutex for work to arrive. It's simple, but obviously sub-optimal.