背景
如this question所述,我正在使用Scalaz 7迭代器处理恒定堆空间中的大型(即无界)数据流。
我的代码如下所示:
type ErrorOrT[M[+_], A] = EitherT[M, Throwable, A]
type ErrorOr[A] = ErrorOrT[IO, A]
def processChunk(c: Chunk, idx: Long): Result
def process(data: EnumeratorT[Chunk, ErrorOr]): IterateeT[Vector[(Chunk, Long)], ErrorOr, Vector[Result]] =
Iteratee.fold[Vector[(Chunk, Long)], ErrorOr, Vector[Result]](Nil) { (rs, vs) =>
rs ++ vs map {
case (c, i) => processChunk(c, i)
}
} &= (data.zipWithIndex mapE Iteratee.group(P))
问题
我似乎遇到了内存泄漏,但我对Scalaz / FP不太熟悉,不知道该错误是在Scalaz中还是在我的代码中。直观地说,我希望这段代码只需要(大约为) P 乘以Chunk
- 大小的空间。
注意:我发现a similar question遇到了OutOfMemoryError
,但我的代码未使用consume
。
测试
我运行了一些测试来尝试找出问题所在。总而言之,只有在使用zipWithIndex
和group
时才会出现泄漏。
// no zipping/grouping
scala> (i1 &= enumArrs(1 << 25, 128)).run.unsafePerformIO
res47: Long = 4294967296
// grouping only
scala> (i2 &= (enumArrs(1 << 25, 128) mapE Iteratee.group(4))).run.unsafePerformIO
res49: Long = 4294967296
// zipping and grouping
scala> (i3 &= (enumArrs(1 << 25, 128).zipWithIndex mapE Iteratee.group(4))).run.unsafePerformIO
java.lang.OutOfMemoryError: Java heap space
// zipping only
scala> (i4 &= (enumArrs(1 << 25, 128).zipWithIndex)).run.unsafePerformIO
res51: Long = 4294967296
// no zipping/grouping, larger arrays
scala> (i1 &= enumArrs(1 << 27, 128)).run.unsafePerformIO
res53: Long = 17179869184
// zipping only, larger arrays
scala> (i4 &= (enumArrs(1 << 27, 128).zipWithIndex)).run.unsafePerformIO
res54: Long = 17179869184
测试代码:
import scalaz.iteratee._, scalaz.effect.IO, scalaz.std.vector._
// define an enumerator that produces a stream of new, zero-filled arrays
def enumArrs(sz: Int, n: Int) =
Iteratee.enumIterator[Array[Int], IO](
Iterator.continually(Array.fill(sz)(0)).take(n))
// define an iteratee that consumes a stream of arrays
// and computes its length
val i1 = Iteratee.fold[Array[Int], IO, Long](0) {
(c, a) => c + a.length
}
// define an iteratee that consumes a grouped stream of arrays
// and computes its length
val i2 = Iteratee.fold[Vector[Array[Int]], IO, Long](0) {
(c, as) => c + as.map(_.length).sum
}
// define an iteratee that consumes a grouped/zipped stream of arrays
// and computes its length
val i3 = Iteratee.fold[Vector[(Array[Int], Long)], IO, Long](0) {
(c, vs) => c + vs.map(_._1.length).sum
}
// define an iteratee that consumes a zipped stream of arrays
// and computes its length
val i4 = Iteratee.fold[(Array[Int], Long), IO, Long](0) {
(c, v) => c + v._1.length
}
问题
答案 0 :(得分:4)
对于那些坚持使用较早的iteratee
API的人来说,这几乎没什么安慰,但我最近验证了对scalaz-stream API的等效测试。这是一个较新的流处理API,旨在替换iteratee
。
为了完整性,这是测试代码:
// create a stream containing `n` arrays with `sz` Ints in each one
def streamArrs(sz: Int, n: Int): Process[Task, Array[Int]] =
(Process emit Array.fill(sz)(0)).repeat take n
(streamArrs(1 << 25, 1 << 14).zipWithIndex
pipe process1.chunk(4)
pipe process1.fold(0L) {
(c, vs) => c + vs.map(_._1.length.toLong).sum
}).runLast.run
这应该适用于n
参数的任何值(假设你愿意等待足够长的时间) - 我测试了2 ^ 14个32MiB阵列(即,总共分配了一半的内存TiB)随着时间的推移)。