Flakiness in IOSpec

[durban]

https://github.com/typelevel/cats-effect/actions/runs/7732197317/job/21083847462#step:24:3685

...
[error]     x not lose cedeing threads from the bypass when blocker transitioning
[error]      List(3714008, 3714008, 3714008) == 'List(3714008, 3714008, 3714008)' (IOPlatformSpecification.scala:459)
...
[info] Total for specification IOSpec
[info] Finished in 17 seconds, 346 ms
[info] 346 examples, 13123 expectations, 1 failure, 0 error

[djspiewak]

List(3714008, 3714008, 3714008) == 'List(3714008, 3714008, 3714008)'

Oh the assertion is as follows:

results must not[List[Long]](beEqualTo(List.fill(3)(results.head)))

Technically this is a fairly timing-sensitive test. The assertion is right here:

            val results = 0.until(3).toList map { _ =>
              Thread.sleep(100L)
              ctr.get()
            }

Specifically, we're sampling ctr (an AtomicLong) once every 100ms and asserting that it's changing. The changes are happening here:

val tsk1 = IO { ctr.incrementAndGet() }.foreverM

So the idea is that we're asserting that this fiber is still running somewhere by observing the changes it is making to the counter. The timing sensitivity is, if the fiber is in fact still running but somehow doesn't get scheduled even once in 300ms of wall clock time, we won't see ctr change and we'll have a spurious failure. We can either increase the sample iteration count or increase the sleep within the sample iterations. Also it's not totally clear to me why we need more than two samples, so that might be a nice simplification.