foreachParN should not use more than 'n' fibers. #1165
Description
The current implementation of foreachParN uses a semaphore to restrict the number of active fibers working concurrently. This doesn't restrict the number of fibers actually being created, and results in some surprising behavior. For example
val rt = new DefaultRuntime {}
val n = 10
val seq = 0 to 1000000
val boom = ZIO.foreachParN[Any,Nothing, Int, Int](n)(seq)(UIO.succeed(_))
rt.unsafeRun(boom)The expectation here would be that this works, but instead we run into an out of memory exception.
[error] (zio-default-async-268-1441532751) java.lang.OutOfMemoryError: GC overhead limit exceeded
[error] java.lang.OutOfMemoryError: GC overhead limit exceeded
[error] at java.lang.Integer.valueOf(Integer.java:832)
[error] at scala.runtime.BoxesRunTime.boxToInteger(BoxesRunTime.java:67)
[error] at zio.internal.PlatformLive$ExecutorUtil$.$anonfun$makeDefault$1$adapted(PlatformLive.scala:67)
[error] at zio.internal.PlatformLive$ExecutorUtil$$$Lambda$4633/467302532.apply(Unknown Source)
[error] at zio.internal.PlatformLive$ExecutorUtil$$anon$2.yieldOpCount(PlatformLive.scala:115)
[error] at zio.internal.FiberContext.evaluateNow(FiberContext.scala:217)
[error] at zio.internal.FiberContext.$anonfun$fork$1(FiberContext.scala:586)
[error] at zio.internal.FiberContext$$Lambda$4847/1031667563.run(Unknown Source)
[error] at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149)
[error] at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624)
[error] at java.lang.Thread.run(Thread.java:748)A possible fix is to change the implementation to use a queue with a worker pool instead of using semaphores.
The general idea would be something like
final def foreachParN[R >: LowerR, E <: UpperE, A, B](n: Long)(as: Iterable[A])(fn: A => ZIO[R, E, B]): ZIO[R, E, List[B]] =
for {
q <- Queue.bounded[(Promise[E, B], A)](n.toInt)
env <- ZIO.environment[R]
pairs <- ZIO.foreach(as)(a => Promise.make[E, B].map(p => (p, a))) // preserve order
_ <- ZIO.foreach(pairs)(pair => q.offer(pair)).fork // publish work
_ <- ZIO.collectAll(List.fill(n.toInt)(q.take.flatMap { case (p, a) => p.done(fn(a).provide(env).onError(_ => q.shutdown)) }.forever.fork)) // worker handles work
res <- ZIO.collectAll(pairs.map(_._1.await)).ensuring(q.shutdown) // wait for results
} yield resOther approaches than the above one are probably possible as well.