Akka(26): Stream:异常处理-Exception handling

   akka-stream是基于Actor模式的,所以也继承了Actor模式的“坚韧性(resilient)”特点,在任何异常情况下都有某种整体统一的异常处理策略和具体实施方式。在akka-stream的官方文件中都有详细的说明和示范例子。我们在这篇讨论里也没有什么更好的想法和范例,也只能略做一些字面翻译和分析理解的事了。下面列出了akka-stream处理异常的一些实用方法:

1、recover:这是一个函数,发出数据流最后一个元素然后根据上游发生的异常终止当前数据流

2、recoverWithRetries:也是个函数,在上游发生异常后改选用后备数据流作为上游继续运行

3、Backoff restart strategy:是RestartSource,RestartFlow,RestartSink的一个属性。为它们提供“逐步延迟重启策略”

4、Supervision strategy:是数据流构件的“异常监管策略”属性。为发生异常的功能阶段Stage提供异常情况处理方法

下面我们就用一些代码例子来示范它们的使用方法:

1、recover:Flow[T].recover函数的款式如下:

  /**
   * Recover allows to send last element on failure and gracefully complete the stream
   * Since the underlying failure signal onError arrives out-of-band, it might jump over existing elements.
   * This stage can recover the failure signal, but not the skipped elements, which will be dropped.
   *
   * Throwing an exception inside `recover` _will_ be logged on ERROR level automatically.
   *
   * '''Emits when''' element is available from the upstream or upstream is failed and pf returns an element
   *
   * '''Backpressures when''' downstream backpressures
   *
   * '''Completes when''' upstream completes or upstream failed with exception pf can handle
   *
   * '''Cancels when''' downstream cancels
   *
   */
  def recover[T >: Out](pf: PartialFunction[Throwable, T]): Repr[T] = via(Recover(pf))

下面是一个用例:

  Source(0 to 10).map { n =>
    if (n < 5) n.toString
    else throw new Exception("Boooommm!")
  }.recover{
    case e: Exception => s"truncate stream: ${e.getMessage}"
  }.runWith(Sink.foreach(println))

运算结果:

0
1
2
3
4
truncate stream: Boooommm!

2、recoverWithRetries:看看它的函数款式:

 /**
   * RecoverWithRetries allows to switch to alternative Source on flow failure. It will stay in effect after
   * a failure has been recovered up to `attempts` number of times so that each time there is a failure
   * it is fed into the `pf` and a new Source may be materialized. Note that if you pass in 0, this won't
   * attempt to recover at all.
   *
   * A negative `attempts` number is interpreted as "infinite", which results in the exact same behavior as `recoverWith`.
   *
   * Since the underlying failure signal onError arrives out-of-band, it might jump over existing elements.
   * This stage can recover the failure signal, but not the skipped elements, which will be dropped.
   *
   * Throwing an exception inside `recoverWithRetries` _will_ be logged on ERROR level automatically.
   *
   * '''Emits when''' element is available from the upstream or upstream is failed and element is available
   * from alternative Source
   *
   * '''Backpressures when''' downstream backpressures
   *
   * '''Completes when''' upstream completes or upstream failed with exception pf can handle
   *
   * '''Cancels when''' downstream cancels
   *
   * @param attempts Maximum number of retries or -1 to retry indefinitely
   * @param pf Receives the failure cause and returns the new Source to be materialized if any
   * @throws IllegalArgumentException if `attempts` is a negative number other than -1
   *
   */
  def recoverWithRetries[T >: Out](attempts: Int, pf: PartialFunction[Throwable, Graph[SourceShape[T], NotUsed]]): Repr[T] =
    via(new RecoverWith(attempts, pf))

attempts代表发生异常过程中尝试恢复次数,0代表不尝试恢复,直接异常中断。<0代表无限尝试次数。下面是一个用例示范: 

 val backupSource = Source(List("five","six","seven","eight","nine"))
  Source(0 to 10).map { n =>
    if (n < 5) n.toString
    else throw new RuntimeException("Boooommm!")
  }.recoverWithRetries(attempts = 1, {
      case e: RuntimeException => backupSource
    }
  ).runWith(Sink.foreach(println))

运算结果:

0
1
2
3
4
five
six
seven
eight
nine

3、Backoff-Restart-Strategy:aka-stream预设定了RestartSource,RestartFlow,RestartSink来在Source,Flow,Sink节点实现“逐步延迟重启策略”,即采取一种逐步延后重启时间点的方式来避免多个进程同时争取某一项资源。下面是这三个类型的定义:

/**
 * A RestartSource wraps a [[Source]] that gets restarted when it completes or fails.
 *
 * They are useful for graphs that need to run for longer than the [[Source]] can necessarily guarantee it will, for
 * example, for [[Source]] streams that depend on a remote server that may crash or become partitioned. The
 * RestartSource ensures that the graph can continue running while the [[Source]] restarts.
 */
object RestartSource {

  /**
   * Wrap the given [[Source]] with a [[Source]] that will restart it when it fails or complete using an exponential
   * backoff.
   *
   * This [[Source]] will never emit a complete or failure, since the completion or failure of the wrapped [[Source]]
   * is always handled by restarting it. The wrapped [[Source]] can however be cancelled by cancelling this [[Source]].
   * When that happens, the wrapped [[Source]], if currently running will be cancelled, and it will not be restarted.
   * This can be triggered simply by the downstream cancelling, or externally by introducing a [[KillSwitch]] right
   * after this [[Source]] in the graph.
   *
   * This uses the same exponential backoff algorithm as [[akka.pattern.Backoff]].
   *
   * @param minBackoff minimum (initial) duration until the child actor will
   *   started again, if it is terminated
   * @param maxBackoff the exponential back-off is capped to this duration
   * @param randomFactor after calculation of the exponential back-off an additional
   *   random delay based on this factor is added, e.g. `0.2` adds up to `20%` delay.
   *   In order to skip this additional delay pass in `0`.
   * @param sourceFactory A factory for producing the [[Source]] to wrap.
   */
  def withBackoff[T](minBackoff: FiniteDuration, maxBackoff: FiniteDuration, randomFactor: Double)(sourceFactory: () ⇒ Source[T, _]): Source[T, NotUsed] = {
    Source.fromGraph(new RestartWithBackoffSource(sourceFactory, minBackoff, maxBackoff, randomFactor))
  }
}

/**
 * A RestartFlow wraps a [[Flow]] that gets restarted when it completes or fails.
 *
 * They are useful for graphs that need to run for longer than the [[Flow]] can necessarily guarantee it will, for
 * example, for [[Flow]] streams that depend on a remote server that may crash or become partitioned. The
 * RestartFlow ensures that the graph can continue running while the [[Flow]] restarts.
 */
object RestartFlow {

  /**
   * Wrap the given [[Flow]] with a [[Flow]] that will restart it when it fails or complete using an exponential
   * backoff.
   *
   * This [[Flow]] will not cancel, complete or emit a failure, until the opposite end of it has been cancelled or
   * completed. Any termination by the [[Flow]] before that time will be handled by restarting it. Any termination
   * signals sent to this [[Flow]] however will terminate the wrapped [[Flow]], if it's running, and then the [[Flow]]
   * will be allowed to terminate without being restarted.
   *
   * The restart process is inherently lossy, since there is no coordination between cancelling and the sending of
   * messages. A termination signal from either end of the wrapped [[Flow]] will cause the other end to be terminated,
   * and any in transit messages will be lost. During backoff, this [[Flow]] will backpressure.
   *
   * This uses the same exponential backoff algorithm as [[akka.pattern.Backoff]].
   *
   * @param minBackoff minimum (initial) duration until the child actor will
   *   started again, if it is terminated
   * @param maxBackoff the exponential back-off is capped to this duration
   * @param randomFactor after calculation of the exponential back-off an additional
   *   random delay based on this factor is added, e.g. `0.2` adds up to `20%` delay.
   *   In order to skip this additional delay pass in `0`.
   * @param flowFactory A factory for producing the [[Flow]] to wrap.
   */
  def withBackoff[In, Out](minBackoff: FiniteDuration, maxBackoff: FiniteDuration, randomFactor: Double)(flowFactory: () ⇒ Flow[In, Out, _]): Flow[In, Out, NotUsed] = {
    Flow.fromGraph(new RestartWithBackoffFlow(flowFactory, minBackoff, maxBackoff, randomFactor))
  }
}

/**
 * A RestartSink wraps a [[Sink]] that gets restarted when it completes or fails.
 *
 * They are useful for graphs that need to run for longer than the [[Sink]] can necessarily guarantee it will, for
 * example, for [[Sink]] streams that depend on a remote server that may crash or become partitioned. The
 * RestartSink ensures that the graph can continue running while the [[Sink]] restarts.
 */
object RestartSink {

  /**
   * Wrap the given [[Sink]] with a [[Sink]] that will restart it when it fails or complete using an exponential
   * backoff.
   *
   * This [[Sink]] will never cancel, since cancellation by the wrapped [[Sink]] is always handled by restarting it.
   * The wrapped [[Sink]] can however be completed by feeding a completion or error into this [[Sink]]. When that
   * happens, the [[Sink]], if currently running, will terminate and will not be restarted. This can be triggered
   * simply by the upstream completing, or externally by introducing a [[KillSwitch]] right before this [[Sink]] in the
   * graph.
   *
   * The restart process is inherently lossy, since there is no coordination between cancelling and the sending of
   * messages. When the wrapped [[Sink]] does cancel, this [[Sink]] will backpressure, however any elements already
   * sent may have been lost.
   *
   * This uses the same exponential backoff algorithm as [[akka.pattern.Backoff]].
   *
   * @param minBackoff minimum (initial) duration until the child actor will
   *   started again, if it is terminated
   * @param maxBackoff the exponential back-off is capped to this duration
   * @param randomFactor after calculation of the exponential back-off an additional
   *   random delay based on this factor is added, e.g. `0.2` adds up to `20%` delay.
   *   In order to skip this additional delay pass in `0`.
   * @param sinkFactory A factory for producing the [[Sink]] to wrap.
   */
  def withBackoff[T](minBackoff: FiniteDuration, maxBackoff: FiniteDuration, randomFactor: Double)(sinkFactory: () ⇒ Sink[T, _]): Sink[T, NotUsed] = {
    Sink.fromGraph(new RestartWithBackoffSink(sinkFactory, minBackoff, maxBackoff, randomFactor))
  }
}

注意这些withBackoff[T]中的sourceFactor,flowFactor,sinkFactory,是它们构建了目标构件。下面我们就虚构一个由RestartSource,RestartFlow,RestartSink合组成的数据流:

  val backoffSource = RestartSource.withBackoff(
    minBackoff = 3.seconds,
    maxBackoff = 30.seconds,
    randomFactor = 0.2
  ){ () => Source(List("FileA","FileB","FileC"))}

  val backoffFlow = RestartFlow.withBackoff(
    minBackoff = 3.seconds,
    maxBackoff = 30.seconds,
    randomFactor = 0.2
  ){ () => Flow[String].map(_.toUpperCase())}

  val backoffSink = RestartSink.withBackoff(
    minBackoff = 3.seconds,
    maxBackoff = 30.seconds,
    randomFactor = 0.2
  ){ () => Sink.foreach(println)}

  backoffSource.via(backoffFlow).to(backoffSink).run()

当然,在现实应用中这几个构件都可能涉及到一些资源的占用,如数据库、网络服务等。下面是运算结果显示:

FILEA
FILEB
FILEC
FILEA
FILEB
FILEC
FILEA
FILEB
FILEC

这个stream是重复循环的。我们只有通过KillSwitch来手动终止它:

  val killSwitch = backoffSource.viaMat(KillSwitches.single)(Keep.right)
    .viaMat(backoffFlow)(Keep.left)
    .toMat(backoffSink)(Keep.left)
    .run()

  Thread.sleep(1000)
  killSwitch.shutdown()

4、Supervisor-Strategy:这种模式是受Actor监管策略模式的启发,在aka-stream的一些功能节点Stage上实现的。对于某些功能节点Stage来说,可能这种监管模式就根本不适用,如连接外部系统的Stage,因为造成异常失败的因素可能还是会重复造成异常。对于出现异常的stream,Supervisor-Strategy提供了三种处理方法:

Stop:终结stream,返回异常

Resume:越过当前元素,继续运行

Restart:重新启动、越过当前元素、清除任何内部状态

akka-stream的默认异常处理方式是Stop,即立即终止数据流,返回异常。

我们可以通过ActorMaterializerSettings().withSupervisionStrategy以及Flow[T].withAttributes(ActorAttributes.withSupervisionStrategy来设定异常监管策略。下面这个例子使用了ActorMaterializerSettings来设定Supervision:

 implicit val mat2 = ActorMaterializer(
    ActorMaterializerSettings(sys).withSupervisionStrategy(decider)
      .withInputBuffer(initialSize = 16, maxSize = 16)
  )

  Source(1 to 5).map { n =>
    if (n != 3) n.toString
    else throw new ArithmeticException("no 3 please!")
  }.runWith(Sink.foreach(println))

  Thread.sleep(1000)
  println("")
  Thread.sleep(1000)

  Source(1 to 5).map { n =>
    if (n != 5) n.toString
    else throw new Exception("no 3 please!")
  }.runWith(Sink.foreach(println))

上面两个stream分别示范了Resume和Stop策略的效果,如下:

1
2
4
5

1
2
3
4

在下面的这个例子里我们在Flow构件的属性Attributes里设定了SupervisionStrategy:

  val decider : Supervision.Decider = {
    case _: IllegalArgumentException => Supervision.Restart
    case _ => Supervision.Stop
  }
  val flow = Flow[Int]
    .scan(0) { (acc, elem) =>
      if (elem < 0) throw new IllegalArgumentException("negative not allowed")
      else acc + elem
    }.withAttributes(ActorAttributes.supervisionStrategy(decider))

  Source(List(1, 3, -1, 5, 7)).via(flow)
    .runWith(Sink.foreach(println))

以上例子中对异常采用了Restart。从下面的运算结果中我们确定了Restart在重启过程中清除了内部状态,也就是说从发生异常的位置开始重新进行计算了:

0
1
4
0
5
12

好了,下面是这次示范涉及的完整源代码:

import akka.actor._
import akka.stream._
import akka.stream.scaladsl._
import scala.concurrent.duration._

object ExceptionHandling extends App {
  implicit val sys = ActorSystem("demoSys")
  implicit val ec = sys.dispatcher
  implicit val mat = ActorMaterializer(
    ActorMaterializerSettings(sys)
      .withInputBuffer(initialSize = 16, maxSize = 16)
  )

/*
  Source(0 to 10).map { n =>
    if (n < 5) n.toString
    else throw new Exception("Boooommm!")
  }.recover{
    case e: Exception => s"truncate stream: ${e.getMessage}"
  }.runWith(Sink.foreach(println))
*/
/*
  val backupSource = Source(List("five","six","seven","eight","nine"))
  Source(0 to 10).map { n =>
    if (n < 5) n.toString
    else throw new RuntimeException("Boooommm!")
  }.recoverWithRetries(attempts = 0, {
      case e: RuntimeException => backupSource
    }
  ).runWith(Sink.foreach(println))

  val backoffSource = RestartSource.withBackoff(
    minBackoff = 3.seconds,
    maxBackoff = 30.seconds,
    randomFactor = 0.2
  ){ () => Source(List("FileA","FileB","FileC"))}

  val backoffFlow = RestartFlow.withBackoff(
    minBackoff = 3.seconds,
    maxBackoff = 30.seconds,
    randomFactor = 0.2
  ){ () => Flow[String].map(_.toUpperCase())}

  val backoffSink = RestartSink.withBackoff(
    minBackoff = 3.seconds,
    maxBackoff = 30.seconds,
    randomFactor = 0.2
  ){ () => Sink.foreach(println)}

  //backoffSource.via(backoffFlow).to(backoffSink).run()

  val killSwitch = backoffSource.viaMat(KillSwitches.single)(Keep.right)
    .viaMat(backoffFlow)(Keep.left)
    .toMat(backoffSink)(Keep.left)
    .run()

  Thread.sleep(1000)
  killSwitch.shutdown()
*/
  /*
  val decider: Supervision.Decider = {
    case _: ArithmeticException => Supervision.Resume
    case _ => Supervision.Stop
  }

  implicit val mat2 = ActorMaterializer(
    ActorMaterializerSettings(sys).withSupervisionStrategy(decider)
      .withInputBuffer(initialSize = 16, maxSize = 16)
  )

  Source(1 to 5).map { n =>
    if (n != 3) n.toString
    else throw new ArithmeticException("no 3 please!")
  }.runWith(Sink.foreach(println))

  Thread.sleep(1000)
  println("")
  Thread.sleep(1000)

  Source(1 to 5).map { n =>
    if (n != 5) n.toString
    else throw new Exception("no 3 please!")
  }.runWith(Sink.foreach(println))
*/
  val decider : Supervision.Decider = {
    case _: IllegalArgumentException => Supervision.Restart
    case _ => Supervision.Stop
  }
  val flow = Flow[Int]
    .scan(0) { (acc, elem) =>
      if (elem < 0) throw new IllegalArgumentException("negative not allowed")
      else acc + elem
    }.withAttributes(ActorAttributes.supervisionStrategy(decider))

  Source(List(1, 3, -1, 5, 7)).via(flow)
    .runWith(Sink.foreach(println))


  scala.io.StdIn.readLine()
  sys.terminate()

}

本文参与腾讯云自媒体分享计划,欢迎正在阅读的你也加入,一起分享。

发表于

我来说两句

0 条评论
登录 后参与评论

相关文章

来自专栏算法与数据结构

CC150--确定两串乱序同构

给定两个字符串,请编写程序,确定其中一个字符串的字符重新排列后,能否变成另一个字符串。这里规定大小写为不同字符,且考虑字符串中的空格。

1083
来自专栏别先生

c3p0和QueryRunner的结合使用,让开发更加简便

1:DBUtils中的QueryRunner的使用:   1.1:QueryRunner中提供了对SQL语句操作的api;   1.2:主要有三个方法:    ...

3488
来自专栏King_3的技术专栏

leetcode-482-License Key Formatting

1233
来自专栏小二的折腾日记

LeetCode-49-Group-Anagrams

输入一个字符串数组,输出的是:将相同字符的字符串放在一个数组的二维数组。相同字符的处理,基本就是要对字符串排序的。然后需要考虑的就是排序好的那一个字符串怎么存的...

601
来自专栏算法修养

LeetCode 102. Binary Tree Level Order Traversal

1354
来自专栏吴小龙同學

Android MVP 实例

Android MVP 实例 简单的请求天气功能,演示Android MVP是如何使用的 效果预览 image.png 准备 MVP流程 ? 说明: 步骤...

2895
来自专栏三流程序员的挣扎

Navigation 详解二

BottomNavigationView 以更简洁的方式来实现过去的 BottomNavigationBar 的样式。Android Studio 中创建一个 ...

942
来自专栏mathor

LeetCode84.柱状图中最大的矩形

 单调栈板子题,创建一个单调递增栈(栈底到栈顶是递增的),栈内存放的数组下标,遍历数组,将下标存进栈内,以样例来说  首先栈空,0直接进栈;然后因为num...

662
来自专栏desperate633

LintCode 平面列表题目分析代码

给定一个列表,该列表中的每个要素要么是个列表,要么是整数。将其变成一个只包含整数的简单列表。 ** 注意事项 ** 如果给定的列表中的要素本身也是一个列表,...

552
来自专栏Golang语言社区

【Go 语言社区】Golang中interface判断nil问题

在示例中,我们定义一个interface名为Stringer,同时定义一个符合其定义的Binary类型: type Stringer interface { ...

43718

扫码关注云+社区