Spark深入解读（十二）---- 有向无环图DAG(代码及图解)

“无意中发现了一个巨牛的人工智能教程，忍不住分享一下给大家。教程不仅是零基础，通俗易懂，而且非常风趣幽默，像看小说一样！觉得太牛了，所以分享给大家。点这里可以跳转到教程。”。

目录：

1、有向无环图

2、代码结构

3、代码学习步鄹及方法

4、重点代码讲解

5、代码展现

6、运行结果

---

---

1、有向无环图

在图论中，如果一个有向图无法从某个顶点出发经过若干条边回到该点，则这个图是一个有向无环图（DAG图）。
因为有向图中一个点经过两种路线到达另一个点未必形成环，因此有向无环图未必能转化成树，但任何有向树均为有向无环图。
性质：有向无环图的生成树个数等于入度非零的节点的入度积。

2、代码结构

3、代码学习步鄹及方法

1、本文中涉及到了Spark的Dag和设计模式中的命令
2、Dag学习步鄹：task --> Node --> DAG --> DAGExecutor
3、设计模式命令 http://blog.csdn.net/silentwolfyh/article/details/52839495
4、图解Dag类的学习步鄹

4、重点代码讲解

下面这段代码是核心也是最难的，如何找到父节点

  //判断Node的task节点的父节点运行状态（flase ，true）
  private def getPending: Option[T] = {
    _pending.find { name =>
      val parents = _nodes(name)
      !parents.exists(name => !_success.contains(name))
    }
  }

1、nodes没有父节点时，!parents.exists() 为true
2、parents.exists() 为flase时，!parents.exists() 为true

5、代码展现

DAG.scala

package com.yh.dag

import java.time.{Duration, LocalDate}
import com.yh.nodeexecutor._
import org.slf4j.LoggerFactory
import scala.collection.immutable.{ListMap, ListSet}

/**
  * Created by yuhui on 2016/8/25.
  * task --> Node --> DAG --> DAGExecutor
  */

case class Node[T](task: T, parent: T*) {
  override def toString: String = {
    s"$task(${parent.mkString(",")})"
  }
}

case class DAG[T](nodes: Node[T]*)

case class DAGExecutor[T](dag: DAG[T]) {
  private val LOG = LoggerFactory.getLogger(this.getClass)
  private val _nodes: Map[T, Seq[T]] = dag.nodes.map(node => (node.task, node.parent.filter(_ != null))).toMap
  private var _pending: Set[T] = ListSet()
  private var _fails = ListMap[T, String]()
  private var _success = Seq[T]()

  //判断Node的task节点的父节点运行状态（flase ，true）
  private def getPending: Option[T] = {
    _pending.find { name =>
      val parents = _nodes(name)
      !parents.exists(name => !_success.contains(name))
    }
  }

  private def fail(name: T, message: String): Unit = {
    _pending -= name
    _fails += name -> message
    for (child <- _pending.filter(child => _nodes(child).contains(name))) {
      fail(child, s"依赖的任务无法执行: $name")
    }
  }

  private def success(name: T): Unit = {
    _pending -= name
    _success = _success :+ name
  }

  def execute(func: T => Unit): Unit = {
    _pending = _nodes.keySet
    _fails = ListMap()
    _success = Seq()
    var running = true

    while (running) {
      val taskOpt = getPending
      if (taskOpt.nonEmpty) {
        val task = taskOpt.get
        val startMills = System.currentTimeMillis()
        LOG.info("start task {}", task)
        try {
          println("=============")
          func(task)  //执行executor方法
          println("+++++++++++++")
          val time = Duration.ofMillis(System.currentTimeMillis() - startMills)
          LOG.info(s"end task $task time=$time")
          success(task)
        } catch {
          case e: Throwable => fail(task, e.getMessage)
            LOG.error(e.getMessage, e)
            LOG.info(s"fail task $task")
        }
      } else {
        running = false
      }
    }

    for (name <- _success) {
      LOG.info(s"success task: $name")
    }
    for (name <- _fails) {
      LOG.info(s"fail task: ${name._1} - ${name._2}")
    }
  }
}

object DAG {
  val allSDKDAG = new DAG[Task](
    Node(UserDetailsExecutor, WebSdkparseExecutor),
    Node(UserTagExecutor, WebSdkparseExecutor,WebSdkparseExecutor),
    Node(WebSdkparseExecutor),
    Node(UserOverviewExecutor, WebSdkparseExecutor)
  )

  def main(args: Array[String]): Unit = {
       DAGExecutor(allSDKDAG).execute { task =>task.executor("appkey": String, LocalDate.now(), LocalDate.now())}
  }
}

Task.scala

package com.yh.dag

import java.time.LocalDate
import org.apache.spark.sql.SQLContext
import org.slf4j.LoggerFactory

/**
  * Created by yuhui on 2016/12/27.
  */
abstract class Task {

  protected val LOG = LoggerFactory.getLogger(this.getClass)

  def executor(appkey: String, startDay: LocalDate, endDay: LocalDate): Unit

  def run(appkey: String, startDay: LocalDate, endDay: LocalDate): Unit = {
    executor(appkey, startDay, endDay)
  }

}

abstract class Executor extends Task with SQLContextAware {

  override def run(appkey: String, startDay: LocalDate, endDay: LocalDate)={}

}

trait SQLContextAware {
  implicit var ctx: SQLContext = _
}

UserDetailsExecutor.scala

package com.yh.nodeexecutor

import java.time.LocalDate
import com.yh.dag.Executor

object UserDetailsExecutor extends Executor{

  override def executor(appkey: String, startDay: LocalDate, endDay: LocalDate): Unit = {

    println("++++我的UserDetailsProcessor的执行过程++++")

  }

}

UserOverviewExecutor.scala

package com.yh.nodeexecutor

import java.time.LocalDate
import com.yh.dag.Executor

/**
  * Created by yuhui on 2016/12/27.
  */
object UserOverviewExecutor extends Executor{

  override def executor(appkey: String, startDay: LocalDate, endDay: LocalDate): Unit = {
    println("++++我的UserOverviewProcessor的执行过程++++")
  }

}

UserTagExecutor.scala

package com.yh.nodeexecutor

import java.time.LocalDate
import com.yh.dag.Executor

/**
  * Created by yuhui on 2016/12/27.
  */
object UserTagExecutor extends Executor{

  override def executor(appkey: String, startDay: LocalDate, endDay: LocalDate): Unit = {
    println("++++我的UserTagProcessor的执行过程++++")
  }

}

WebSdkparseExecutor.scala

package com.yh.nodeexecutor

import java.time.LocalDate
import com.yh.dag.Executor

/**
  * Created by yuhui on 2016/12/27.
  */
object WebSdkparseExecutor extends Executor{

  override def executor(appkey: String, startDay: LocalDate, endDay: LocalDate): Unit = {
    println("++++我的WebSdkparseProcessor的执行过程++++")
  }

}

6、运行结果

=============
++++我的WebSdkparseProcessor的执行过程++++
+++++++++++++
=============
++++我的UserDetailsProcessor的执行过程++++
+++++++++++++
=============
++++我的UserTagProcessor的执行过程++++
+++++++++++++
=============
++++我的UserOverviewProcessor的执行过程++++
+++++++++++++

Process finished with exit code 0

北京小辉微信公众号

大数据资料分享请关注