fs2是scalaz-stream的最新版本，沿用了scalaz-stream被动式（pull model）数据流原理但采用了全新的实现方法。fs2比较scalaz-stream而言具备了：更精简的基础组件（combinator）、更安全的类型、资源使用（type safe, resource safety）、更高的运算效率。由于fs2基本沿用了scalaz-stream的原理，所以我们会在下面的讨论里着重介绍fs2的使用。根据fs2的官方文件，fs2具备了以下新的特点：
1、完全不含任何外部依赖（third-party dependency）
2、流元素增加了节组（chunk）类型和相关的操作方法
3、fs2不再只局限于Task一种副作用运算方式（effect）。用户可以提供自己的effect类型
4、更精简的流转换组件（stream transformation primitives）
5、增加了更多并行运算组件（concurrent primitives）
6、通过bracket函数增强了资源使用安全，特别是异线程资源占用的事后处理过程。用onFinalize取代了onComplete
7、stream状态转换采用了全新的实现方式，使用了新的数据结构：Pull
8、Stream取代了Process。fs2中再没有Process1、Tee、Wye、Channel这些类型别名，取而代之的是：
- type Pipe[F,A,B] = Stream[F,A] => Stream[F,B]
- type Pipe2[F,A,B,C] = (Stream[F,A], Stream[F,B]) => Stream[F,C]
- Pipe 替代了 Channel 和 Process1
- Pipe2 替代了 Tee 和 Wye
下面我们来看看fs2的一些基本操作：
1 Stream() //> res0: fs2.Stream[Nothing,Nothing] = Segment(Emit(Chunk()))
2 Stream(1,2,3) //> res1: fs2.Stream[Nothing,Int] = Segment(Emit(Chunk(1, 2, 3)))
3 Stream.emit(4) //> res2: fs2.Stream[Nothing,Int] = Segment(Emit(Chunk(4)))
4 Stream.emits(Seq(1,2,3)) //> res3: fs2.Stream[Nothing,Int] = Segment(Emit(Chunk(1, 2, 3)))
Stream的类型款式是：Stream[F[_],A]。从上面的例子我们看到所有的F[_]都是Nothing，我们称这样的流为纯数据流（pure stream）。再值得注意的是每个流构建都形成了一个Chunk，代表一节元素。fs2增加了Chunk类型来提高数据元素处理效率。这是fs2的一项新功能。
我们可以用toList或者toVector来运算纯数据流中的元素值：
1 Stream.emits(Seq(1,2,3)).toList //> res3: List[Int] = List(1, 2, 3)
2 Stream.emits(Seq(1,2,3)).toVector //> res4: Vector[Int] = Vector(1, 2, 3)
纯数据流具备了许多与List相似的操作函数：
1 (Stream(1,2,3) ++ Stream(4,5)).toList //> res5: List[Int] = List(1, 2, 3, 4, 5)
2 Stream(1,2,3).map { _ + 1}.toList //> res6: List[Int] = List(2, 3, 4)
3 Stream(1,2,3).filter { _ % 2 == 0}.toList //> res7: List[Int] = List(2)
4 Stream(1,2,3).fold(0)(_ + _).toList //> res8: List[Int] = List(6)
5 Stream(None,Some(1),Some(3),None).collect {
6 case None => 0
7 case Some(i) => i
8 }.toList //> res9: List[Int] = List(0, 1, 3, 0)
9 Stream.range(1,5).intersperse(42).toList //> res10: List[Int] = List(1, 42, 2, 42, 3, 42, 4)
10 Stream(1,2,3).flatMap {x => Stream(x,x)}.toList //> res11: List[Int] = List(1, 1, 2, 2, 3, 3)
11 Stream(1,2,3).repeat.take(5).toList //> res12: List[Int] = List(1, 2, 3, 1, 2)
以上都是一些基本的List操作函数示范。
我们知道，纯数据流就是scalaz-stream里的Process1，即transducer，是负责对流进行状态转换的。在fs2里transducer就是Pipe（也是channel），我们一般用through来连接transducer。上面示范中的take,filter等都是transducer，我们可以在object pipe里找到这些函数：
1 object pipe {
2 ...
3 /** Drop `n` elements of the input, then echo the rest. */
4 def drop[F[_],I](n: Long): Stream[F,I] => Stream[F,I] =
5 _ pull (h => Pull.drop(n)(h) flatMap Pull.echo)
6 ...
7 /** Emits `true` as soon as a matching element is received, else `false` if no input matches */
8 def exists[F[_], I](p: I => Boolean): Stream[F, I] => Stream[F, Boolean] =
9 _ pull { h => Pull.forall[F,I](!p(_))(h) flatMap { i => Pull.output1(!i) }}
10
11 /** Emit only inputs which match the supplied predicate. */
12 def filter[F[_], I](f: I => Boolean): Stream[F,I] => Stream[F,I] =
13 mapChunks(_ filter f)
14
15 /** Emits the first input (if any) which matches the supplied predicate, to the output of the returned `Pull` */
16 def find[F[_],I](f: I => Boolean): Stream[F,I] => Stream[F,I] =
17 _ pull { h => Pull.find(f)(h).flatMap { case o #: h => Pull.output1(o) }}
18
19
20 /**
21 * Folds all inputs using an initial value `z` and supplied binary operator,
22 * and emits a single element stream.
23 */
24 def fold[F[_],I,O](z: O)(f: (O, I) => O): Stream[F,I] => Stream[F,O] =
25 _ pull { h => Pull.fold(z)(f)(h).flatMap(Pull.output1) }
26 ...
27 /** Emits all elements of the input except the first one. */
28 def tail[F[_],I]: Stream[F,I] => Stream[F,I] =
29 drop(1)
30
31 /** Emit the first `n` elements of the input `Handle` and return the new `Handle`. */
32 def take[F[_],I](n: Long): Stream[F,I] => Stream[F,I] =
33 _ pull Pull.take(n)
34 ...
我们可以用through来连接这些transducer：
1 Stream(1,2,3).repeat
2 .throughPure(pipe.take(10))
3 .throughPure(pipe.filter(_ % 2 == 0))
4 .toList //> res13: List[Int] = List(2, 2, 2)
以上的throughPure等于是through + pure。Pure是没有任何作用的F[_]，是专门为帮助compiler进行类型推导的类型。其实我们可以用pure先把纯数据流升格后再用through：
1 Stream(1,2,3).repeat.pure 2 .through(pipe.take(10)) 3 .through(pipe.filter(_ % 2 == 0)) 4 .toList //> res14: List[Int] = List(2, 2, 2)
这时compiler不再出错误信息了。在fs2 pipe对象里的函数通过方法注入或者类型继承变成了Stream的自身函数，所以我们也可以直接在Stream类型上使用这些transducer：
1 Stream(1,2,3).repeat.take(10).filter(_ % 2 == 0).toList
2 //> res15: List[Int] = List(2, 2, 2)
我们在前面提到过fs2使用了全新的方法和数据类型来实现transducer。transducer的类型是Pipe，即：
type Pipe[F[_],-I,+O] = Stream[F,I] => Stream[F,O]
我们看到Pipe就是一个Function1的类型别名，一个lambda：提供一个Stream[F,I]，返回Stream[F,O]。那么在fs2里是如何读取一个Stream[F,I]里的元素呢？我们前面提到是通过一个新的数据结构Pull来实现的，先来看看fs2是如何实现Stream >> Pull >> Stream转换的：
1 val pll = Stream(1,2,3).pure.open //> pll : fs2.Pull[fs2.Pure,Nothing,fs2.Stream.Handle[fs2.Pure,Int]] = fs2.Pull
2 de5031f
3 val strm = pll.close //> strm : fs2.Stream[fs2.Pure,Nothing] = evalScope(Scope(Bind(Eval(Snapshot),<
4 function1>))).flatMap(<function1>)
对一个Stream施用open后得到一个Pull类型。pll是个Pull数据结构，它的类型定义如下：
class Pull[+F[_],+O,+R](private[fs2] val get: Free[P[F,O]#f,Option[Either[Throwable,R]]])
在Pull的类型参数中F是一个运算，O代表输出元素类型，R代表Pull里的数据资源。我们可以从R读取元素。在上面的例子里pll的R值是个Handle类型。这个类型里应该提供了读取元素的方法：
implicit class HandleOps[+F[_],+A](h: Handle[F,A]) {
def push[A2>:A](c: Chunk[A2])(implicit A2: RealSupertype[A,A2]): Handle[F,A2] =
self.push(h: Handle[F,A2])(c)
def push1[A2>:A](a: A2)(implicit A2: RealSupertype[A,A2]): Handle[F,A2] =
self.push1(h: Handle[F,A2])(a)
def #:[H](hd: H): Step[H, Handle[F,A]] = Step(hd, h)
def await: Pull[F, Nothing, Step[Chunk[A], Handle[F,A]]] = self.await(h)
def await1: Pull[F, Nothing, Step[A, Handle[F,A]]] = self.await1(h)
def awaitNonempty: Pull[F, Nothing, Step[Chunk[A], Handle[F,A]]] = Pull.awaitNonempty(h)
def echo1: Pull[F,A,Handle[F,A]] = Pull.echo1(h)
def echoChunk: Pull[F,A,Handle[F,A]] = Pull.echoChunk(h)
def peek: Pull[F, Nothing, Step[Chunk[A], Handle[F,A]]] = self.peek(h)
def peek1: Pull[F, Nothing, Step[A, Handle[F,A]]] = self.peek1(h)
def awaitAsync[F2[_],A2>:A](implicit S: Sub1[F,F2], F2: Async[F2], A2: RealSupertype[A,A2]):
Pull[F2, Nothing, AsyncStep[F2,A2]] = self.awaitAsync(Sub1.substHandle(h))
def await1Async[F2[_],A2>:A](implicit S: Sub1[F,F2], F2: Async[F2], A2: RealSupertype[A,A2]):
Pull[F2, Nothing, AsyncStep1[F2,A2]] = self.await1Async(Sub1.substHandle(h))
def covary[F2[_]](implicit S: Sub1[F,F2]): Handle[F2,A] = Sub1.substHandle(h)
}
implicit class HandleInvariantEffectOps[F[_],+A](h: Handle[F,A]) {
def invAwait1Async[A2>:A](implicit F: Async[F], A2: RealSupertype[A,A2]):
Pull[F, Nothing, AsyncStep1[F,A2]] = self.await1Async(h)
def invAwaitAsync[A2>:A](implicit F: Async[F], A2: RealSupertype[A,A2]):
Pull[F, Nothing, AsyncStep[F,A2]] = self.awaitAsync(h)
def receive1[O,B](f: Step[A,Handle[F,A]] => Pull[F,O,B]): Pull[F,O,B] = h.await1.flatMap(f)
def receive[O,B](f: Step[Chunk[A],Handle[F,A]] => Pull[F,O,B]): Pull[F,O,B] = h.await.flatMap(f)
}
果然在Handle提供的函数里有await,receive等这些读取函数。我们试着来实现一个简单的transducer：一个filter函数：
1 import scala.language.higherKinds
2 def myFilter[F[_],A](f: A => Boolean): Pipe[F, A, A] = {
3 def go(h: Stream.Handle[F,A]): Pull[F,A,Unit] = {
4 // h.receive1 {case Step(a,h) => if(f(a)) Pull.output1(a) >> go(h) else go(h)}
5 h.await1.flatMap { case Step(a,h) => if(f(a)) Pull.output1(a) >> go(h) else go(h)}
6 }
7 // sin => sin.open.flatMap {h => go(h)}.close
8 sin => sin.pull(go _)
9 } //> myFilter: [F[_], A](f: A => Boolean)fs2.Pipe[F,A,A]
10
11 Stream.range(0,10).pure.through(myFilter(_ % 2 == 0)).toList
12 //> res17: List[Int] = List(0, 2, 4, 6, 8)
我们从Pull里用await1或者receive1把一个Step数据结构从Handle里扯（pull）出来然后再output到Pull结构里。把这个Pull close后得到我们需要的Stream。我们把例子使用的类型及函数款式陈列在下面：
type Pipe[F[_],-I,+O] = Stream[F,I] => Stream[F,O]
def await1[F[_],I]: Handle[F,I] => Pull[F,Nothing,Step[I,Handle[F,I]]] = {...}
def receive1[F[_],I,O,R](f: Step[I,Handle[F,I]] => Pull[F,O,R]): Handle[F,I] => Pull[F,O,R] =
_.await1.flatMap(f)
def pull[F[_],F2[_],A,B](s: Stream[F,A])(using: Handle[F,A] => Pull[F2,B,Any])(implicit S: Sub1[F,F2])
: Stream[F2,B] =
Pull.close { Sub1.substPull(open(s)) flatMap (h => Sub1.substPull(using(h))) }
再示范另一个Pipe的实现：take
1 def myTake[F[_],A](n: Int): Pipe[F,A,A] = {
2 def go(n: Int): Stream.Handle[F,A] => Pull[F,A,Unit] = h => {
3 if (n <= 0) Pull.done
4 else h.receive1 { case a #: h => Pull.output1(a).flatMap{_ => go(n-1)(h)}}
5 }
6 sin => sin.pull(go(n))
7 } //> myTake: [F[_], A](n: Int)fs2.Pipe[F,A,A]
8 Stream.range(0,10).pure.through(myTake(3)).toList //> res18: List[Int] = List(0, 1, 2)
我们曾经提过fs2功能提升的其中一项是增加了节组（Chunk）数据类型和相关的操作函数。Chunk是fs2内部使用的一种集合，这样fs2就可以一节一节（by chunks）来处理数据了。Chunk本身具备了完整的集合函数：
/**
* Chunk represents a strict, in-memory sequence of `A` values.
*/
trait Chunk[+A] { self =>
def size: Int
def uncons: Option[(A, Chunk[A])] =
if (size == 0) None
else Some(apply(0) -> drop(1))
def apply(i: Int): A
def copyToArray[B >: A](xs: Array[B]): Unit
def drop(n: Int): Chunk[A]
def take(n: Int): Chunk[A]
def filter(f: A => Boolean): Chunk[A]
def foldLeft[B](z: B)(f: (B,A) => B): B
def foldRight[B](z: B)(f: (A,B) => B): B
def indexWhere(p: A => Boolean): Option[Int] = {
val index = iterator.indexWhere(p)
if (index < 0) None else Some(index)
}
def isEmpty = size == 0
def toArray[B >: A: ClassTag]: Array[B] = {
val arr = new Array[B](size)
copyToArray(arr)
arr
}
def toList = foldRight(Nil: List[A])(_ :: _)
def toVector = foldLeft(Vector.empty[A])(_ :+ _)
def collect[B](pf: PartialFunction[A,B]): Chunk[B] = {
val buf = new collection.mutable.ArrayBuffer[B](size)
iterator.collect(pf).copyToBuffer(buf)
Chunk.indexedSeq(buf)
}
def map[B](f: A => B): Chunk[B] = {
val buf = new collection.mutable.ArrayBuffer[B](size)
iterator.map(f).copyToBuffer(buf)
Chunk.indexedSeq(buf)
}
def mapAccumulate[S,B](s0: S)(f: (S,A) => (S,B)): (S,Chunk[B]) = {
val buf = new collection.mutable.ArrayBuffer[B](size)
var s = s0
for { c <- iterator } {
val (newS, newC) = f(s, c)
buf += newC
s = newS
}
(s, Chunk.indexedSeq(buf))
}
def scanLeft[B](z: B)(f: (B, A) => B): Chunk[B] = {
val buf = new collection.mutable.ArrayBuffer[B](size + 1)
iterator.scanLeft(z)(f).copyToBuffer(buf)
Chunk.indexedSeq(buf)
}
def iterator: Iterator[A] = new Iterator[A] {
var i = 0
def hasNext = i < self.size
def next = { val result = apply(i); i += 1; result }
}
...
fs2的大部分转换函数都考虑了对Chunk数据的处理机制。我们先看看fs2是如何表现Chunk数据的：
1 (Stream(1,2) ++ Stream(3,4,5) ++ Stream(6,7)).chunks.toList
2 //> res16: List[fs2.Chunk[Int]] = List(Chunk(1, 2), Chunk(3, 4, 5), Chunk(6, 7))
fs2是按照Stream的构建批次来分节的。我们来示范一下如何使用Pull的Chunk机制：
1 def myTakeC[F[_],A](n: Int): Pipe[F,A,A] = {
2 def go(n: Int): Stream.Handle[F,A] => Pull[F,A,Unit] = h => {
3 if ( n <= 0 ) Pull.done
4 else Pull.awaitLimit(n)(h).flatMap {case Step(chunk,h) =>
5 if (chunk.size <= n) Pull.output(chunk) >> go(n-chunk.size)(h)
6 else Pull.output(chunk.take(n)) }
7 }
8 sin => sin.pull(go(n))
9 } //> myTakeC: [F[_], A](n: Int)fs2.Pipe[F,A,A]
10 val s1 = (Stream(1,2) ++ Stream(3,4,5) ++ Stream(6,7))
11 //> s1 : fs2.Stream[Nothing,Int] = append(append(Segment(Emit(Chunk(1, 2))), S
12 egment(Emit(Chunk(()))).flatMap(<function1>)), Segment(Emit(Chunk(()))).fla
13 tMap(<function1>))
14 s1.pure.through(myTake(4)).chunks.toList //> res20: List[fs2.Chunk[Int]] = List(Chunk(1), Chunk(2), Chunk(3), Chunk(4))
15 s1.pure.through(myTakeC(4)).chunks.toList //> res21: List[fs2.Chunk[Int]] = List(Chunk(1, 2), Chunk(3, 4))
myTake和myTakeC产生了不同的结果。
fs2的特长应该是多线程编程了。在Stream的类型款式中：Stream[F[_],A]，F[_]是一种可能产生副作用的运算方式，当F[_]等于Nothing时，Stream[Nothing,A]是一种纯数据流，而Stream[F[_],A]就是一种运算流了。我们可以在对运算流进行状态转换的过程中进行运算来实现F的副作用如：数据库读写、IO操作等。fs2不再绑定Task一种运算方式了。任何有Catchable实例的Monad都可以成为Stream的运算方式。但是，作为一种以多线程编程为主导的工具库，没有什么运算方式会比Task更合适了。
我们可以把一个纯数据流升格成运算流：
1 val s2 = Stream.emits(Seq(1,2,3)).covary[Task] //> s2 : fs2.Stream[fs2.Task,Int] = Segment(Emit(Chunk(1, 2, 3)))
我们先运算这个运算流，结果为一个Task，然后再运算Task来获取运算值：
1 val s2 = Stream.emits(Seq(1,2,3)).covary[Task] //> s2 : fs2.Stream[fs2.Task,Int] = Segment(Emit(Chunk(1, 2, 3)))
2 val t2 = s2.runLog //> t2 : fs2.Task[Vector[Int]] = Task
3 t2.unsafeRun //> res22: Vector[Int] = Vector(1, 2, 3)
现在使用myTake和myFilter就不需要pure升格了：
1 s3.through(myFilter(_ % 2 == 0)).through(myTake(3)).runLog.unsafeRun
2 //> res23: Vector[Int] = Vector(2, 2, 2)
下面的例子里展示了fs2的运算流从源头（Source）到传换（Transducer）一直到终点（Sink）的使用示范：
1 def stdOut: Sink[Task,String] =
2 _.evalMap { x => Task.delay{ println(s"milli: $x")}}
3 //> stdOut: => fs2.Sink[fs2.Task,String]
4 Stream.repeatEval(Task.delay{System.currentTimeMillis})
5 .map(_.toString)
6 .through(myTake(3))
7 .to(stdOut)
8 .run.unsafeRun //> milli: 1472001934708
9 //| milli: 1472001934714
10 //| milli: 1472001934714
在上面的例子里我们使用了through,to等连接函数。由于数据最终发送到终点stdOut，我们无须用runLog来记录运算结果。
原创文章，作者：奋斗，如若转载，请注明出处：https://blog.ytso.com/12895.html