`IO`s created with `par` fail with `ClassCastException` if performed within loop

Question

`IO`s created with `par` fail with `ClassCastException` if performed within loop

ktonga opened this issue 6 years ago · comments

So it seems something is going on wrong with par implementation, when performed in isolation it seems to work but it starts having unpredictable behavior when performed multiple consecutive times, I realized about this since I have some property-based test with default sampling of 100 performing some IO using par combinator.

You can find the test i'm talking about here: https://github.com/ktonga/scala-poc/blob/topic/parallel-io/src/test/scala/com/brickx/autoinvest/ProgramSpec.scala#L26

This tests used to run ok when IOs were combined using apply2 from default Applicative instance but started failing after I introduced the use of par in this WIP PR: ktonga/scala-poc#6

I'll try to add a minimal example that reproduces the bug to this project for easy debugging.

Cheers.

John A. De Goes · Answer 1 · Sat May 05 2018 17:16:11 GMT+0800 (China Standard Time)

@ktonga Thanks, I will investigate and fix (assuming it's in 8 which has some bug fixes relative to this version).

Gastón Tonietti · Answer 2 · Sat May 05 2018 17:19:47 GMT+0800 (China Standard Time)

Thanks John! I haven't tested it on 8 since i found it in an app where I'm only using the port.

Jed Wesley-Smith · Answer 3 · Sun May 06 2018 12:17:26 GMT+0800 (China Standard Time)

@jdegoes I added to scalaz8's RTSSpec and it failed:

[error]     ! multiple par of two values
[error]  java.lang.ClassCastException: scala.Tuple2 cannot be cast to java.lang.Integer (file:1)

John A. De Goes · Answer 4 · Sun May 06 2018 14:42:40 GMT+0800 (China Standard Time)

Awesome, thanks—fix coming shortly.

John A. De Goes · Answer 5 · Tue May 08 2018 01:46:39 GMT+0800 (China Standard Time)

Fixed here

>>= · Answer 6 · Thu May 10 2018 19:00:56 GMT+0800 (China Standard Time)

Apparently not? Latest fixes were backported (I think)

Also, it is exceptionally annoying that specs2 swallows these things.

java.lang.ClassCastException: scala.Tuple2 cannot be cast to java.lang.Integer
	at scala.runtime.BoxesRunTime.unboxToInt(BoxesRunTime.java:101)
	at scalaz.ioeffect.RTSSpec.$anonfun$testPar$2(RTSTest.scala:376)
	at scala.runtime.java8.JFunction0$mcI$sp.apply(JFunction0$mcI$sp.java:12)
	at org.specs2.matcher.Expectable.value$lzycompute(Expectable.scala:21)
	at org.specs2.matcher.Expectable.value(Expectable.scala:21)
	at org.specs2.matcher.EqualityMatcher.apply(EqualityMatcher.scala:27)
	at org.specs2.matcher.Expectable.applyMatcher(Expectable.scala:51)
	at org.specs2.matcher.MustExpectable.must_$eq$eq$eq(MustExpectable.scala:20)
	at scalaz.ioeffect.RTSSpec.$anonfun$testPar$1(RTSTest.scala:377)
	at scalaz.ioeffect.RTSSpec.$anonfun$testPar$1$adapted(RTSTest.scala:372)
	at scala.collection.TraversableLike.$anonfun$map$1(TraversableLike.scala:234)
	at scala.collection.immutable.Range.foreach(Range.scala:156)
	at scala.collection.TraversableLike.map(TraversableLike.scala:234)
	at scala.collection.TraversableLike.map$(TraversableLike.scala:227)
	at scalaz.ioeffect.RTSSpec.testPar(RTSTest.scala:372)
	at scalaz.ioeffect.RTSSpec.$anonfun$is$46(RTSTest.scala:84)
	at org.specs2.specification.AroundTimeout$$anon$1.result$lzycompute$1(ExamplesTimeout.scala:37)
	at org.specs2.specification.AroundTimeout$$anon$1.result$1(ExamplesTimeout.scala:37)
	at org.specs2.specification.AroundTimeout$$anon$1.$anonfun$around$3(ExamplesTimeout.scala:40)
	at org.specs2.matcher.MatchResult$$anon$13.asResult(MatchResult.scala:349)
	at org.specs2.execute.AsResult$.apply(AsResult.scala:32)
	at org.specs2.specification.AroundTimeout$$anon$1.around(ExamplesTimeout.scala:40)
	at org.specs2.specification.Around.apply(Context.scala:137)
	at org.specs2.specification.Around.apply$(Context.scala:137)
	at org.specs2.specification.AroundTimeout$$anon$1.apply(ExamplesTimeout.scala:35)
	at org.specs2.specification.AroundTimeout.upTo(ExamplesTimeout.scala:32)
	at org.specs2.specification.AroundTimeout.upTo$(ExamplesTimeout.scala:31)
	at scalaz.ioeffect.RTSSpec.upTo(RTSTest.scala:15)
	at scalaz.ioeffect.RTSSpec.$anonfun$is$45(RTSTest.scala:84)
	at org.specs2.execute.ResultExecution.execute(ResultExecution.scala:23)
	at org.specs2.execute.ResultExecution.execute$(ResultExecution.scala:21)
	at org.specs2.execute.ResultExecution$.execute(ResultExecution.scala:119)
	at org.specs2.execute.Result$$anon$11.asResult(Result.scala:246)
	at org.specs2.execute.AsResult$.apply(AsResult.scala:32)
	at org.specs2.specification.core.AsExecution$$anon$1.$anonfun$execute$1(AsExecution.scala:15)
	at org.specs2.execute.ResultExecution.execute(ResultExecution.scala:23)
	at org.specs2.execute.ResultExecution.execute$(ResultExecution.scala:21)
	at org.specs2.execute.ResultExecution$.execute(ResultExecution.scala:119)
	at org.specs2.execute.Result$$anon$11.asResult(Result.scala:246)
	at org.specs2.execute.AsResult$.apply(AsResult.scala:32)
	at org.specs2.execute.AsResult$.$anonfun$safely$1(AsResult.scala:40)
	at org.specs2.execute.ResultExecution.execute(ResultExecution.scala:23)
	at org.specs2.execute.ResultExecution.execute$(ResultExecution.scala:21)
	at org.specs2.execute.ResultExecution$.execute(ResultExecution.scala:119)
	at org.specs2.execute.AsResult$.safely(AsResult.scala:40)
	at org.specs2.specification.core.Execution$.$anonfun$result$1(Execution.scala:305)
	at org.specs2.specification.core.Execution$.$anonfun$withEnvSync$3(Execution.scala:323)
	at org.specs2.execute.ResultExecution.execute(ResultExecution.scala:23)
	at org.specs2.execute.ResultExecution.execute$(ResultExecution.scala:21)
	at org.specs2.execute.ResultExecution$.execute(ResultExecution.scala:119)
	at org.specs2.execute.Result$$anon$11.asResult(Result.scala:246)
	at org.specs2.execute.AsResult$.apply(AsResult.scala:32)
	at org.specs2.execute.AsResult$.$anonfun$safely$1(AsResult.scala:40)
	at org.specs2.execute.ResultExecution.execute(ResultExecution.scala:23)
	at org.specs2.execute.ResultExecution.execute$(ResultExecution.scala:21)
	at org.specs2.execute.ResultExecution$.execute(ResultExecution.scala:119)
	at org.specs2.execute.AsResult$.safely(AsResult.scala:40)
	at org.specs2.specification.core.Execution$.$anonfun$withEnvSync$2(Execution.scala:323)
	at org.specs2.specification.core.Execution.$anonfun$startExecution$3(Execution.scala:135)
	at scala.util.Success.$anonfun$map$1(Try.scala:251)
	at scala.util.Success.map(Try.scala:209)
	at scala.concurrent.Future.$anonfun$map$1(Future.scala:289)
	at scala.concurrent.impl.Promise.liftedTree1$1(Promise.scala:29)
	at scala.concurrent.impl.Promise.$anonfun$transform$1(Promise.scala:29)
	at scala.concurrent.impl.CallbackRunnable.run(Promise.scala:60)
	at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149)
	at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624)
	at java.lang.Thread.run(Thread.java:748)

John A. De Goes · Answer 7 · Thu May 10 2018 19:02:40 GMT+0800 (China Standard Time)

Hmm, this is really weird since it passes locally and in Travis CI (in fact I have two tests now that test essentially the same thing). I'll work harder to reproduce this locally.

>>= · Answer 8 · Thu May 10 2018 19:04:39 GMT+0800 (China Standard Time)

TIL testOnly -- fullstacktrace is how to surface stack traces in specs2 .

>>= · Answer 9 · Thu May 10 2018 19:10:13 GMT+0800 (China Standard Time)

To reproduce for me

ioeffect/testOnly *RTS* -- fullStackTrace

  @inline
  final def nextInstr[E](value: Any, stack: Stack): IO[E, Any] =
    if (!stack.isEmpty()) stack.pop()(value).asInstanceOf[IO[E, Any]] else null

is where it's happening. I'm guessing a tuple is added to the stack incorrectly.

>>= · Answer 10 · Thu May 10 2018 19:15:25 GMT+0800 (China Standard Time)

Some printlns to help...

  @inline
  final def nextInstr[E](value: Any, stack: Stack): IO[E, Any] = try {
    if (!stack.isEmpty()) stack.pop()(value).asInstanceOf[IO[E, Any]] else null
  } catch {
    case t: ClassCastException =>
      println(s"GOT $value in stack $stack")

      throw t
  }

GOT 1 in stack scalaz.ioeffect.RTS$Stack@5ad19e47
GOT (1,\/-(2)) in stack scalaz.ioeffect.RTS$Stack@5e0b8c09
GOT FlatMap(FlatMap(AsyncEffect(scalaz.ioeffect.RTS$FiberContext$$Lambda$863/110484312@431ccd3f),scalaz.ioeffect.IO$$$Lambda$656/285346270@479d6400),scalaz.ioeffect.IO$$Lambda$926/1075799910@120b03a7) in stack scalaz.ioeffect.RTS$Stack@5e0b8c09

>>= · Answer 11 · Thu May 10 2018 19:17:26 GMT+0800 (China Standard Time)

btw, should those vars not be marked @volatile?

Gastón Tonietti · Answer 12 · Thu May 10 2018 19:54:32 GMT+0800 (China Standard Time)

It was really easy for me to reproduce it, i just ran that single test a few times in a row in both repos and the test failed every single time with class cast exception. Do we have real parallelism in CI? some times with the vm has a tiny cpu the thread pool has only a couple threads. This seems to happen only when a lot of stuff is going on at the same time.

John A. De Goes · Answer 13 · Fri May 11 2018 01:33:23 GMT+0800 (China Standard Time)

Thanks, this is helpful. I'll make sure I actually fix it this time. 😄

John A. De Goes · Answer 14 · Fri May 11 2018 03:09:38 GMT+0800 (China Standard Time)

Oh my, this was one of the most fun bugs I've fixed in a long time. PR coming soon. 😄

John A. De Goes · Answer 15 · Fri May 11 2018 03:26:46 GMT+0800 (China Standard Time)

Fixed for realz now

John A. De Goes · Answer 16 · Fri May 11 2018 04:57:22 GMT+0800 (China Standard Time)

@ktonga Can you verify?

Gastón Tonietti · Answer 17 · Fri May 11 2018 05:55:01 GMT+0800 (China Standard Time)

@jdegoes will do! Thanks.

>>= · Answer 18 · Fri May 11 2018 06:35:01 GMT+0800 (China Standard Time)

Raise a PR with the backport and I'll release tomorrow as 2.2.0

Gastón Tonietti · Answer 19 · Fri May 11 2018 08:34:31 GMT+0800 (China Standard Time)

Can confirm, bug seems to be fixed in z8. Will work on the backport. Thanks John!

Gastón Tonietti · Answer 20 · Mon May 14 2018 15:54:41 GMT+0800 (China Standard Time)

Fixed by #34