Port memory leak tests to integration tests, runnable via it:test

[mpilquist]

Bad news: two tests are failing - parJoin and `signal continuous'.

[SystemFw]

The silver lining is that this test strategy seems to be very useful :)

[mpilquist]

Both leaks appear to have existed since 1.0.0 and didn't exist in 0.9.7 (this is definitely true for parJoin and I think it's true for the other, though I started focusing on parJoin).

I updated the PR to write a heap dump upon test failure. Analyzing the parJoin failure points to a leak in PubSub.Strategy.Discrete -- looks like the outOfOrder map is huge.

[SystemFw]

btw I'm getting a bit worried about PubSub in general.
I appreciate the interface reusability that it gives, but we already know for sure that Topic has a contention issue that isn't immediately fixable with PubSub, and now this leak.

Hopefully Pavel can save the day :)
But if not, do you think it's something that needs addressing at the root, or should we just fix it as we go?

[mpilquist]

I rolled SignallingRef back to its 0.9 implementation, which fixed both failing leak tests.

The discrete pubsub strategy definitely seems to have a leak where the outOfOrder map grows very very large, though I wasn't able to work out the exact case in which it happens. I find the pubsub stuff hard to reason about, but perhaps that's just unfamiliarity. I have no strong opinions on pubsub going forward, though I don't think we should be shy about writing custom concurrent data structures if we get stuff in exchange -- i.e., maximally reusing pubsub is not a goal.

[SystemFw]

i.e., maximally reusing pubsub is not a goal.

Right, cause we already know that for Topic is for sure a regression