Spark2.3.2源码解析： 7.3. SparkContext源码分析（二）：调度系统 DAGScheduler

旧城等待， 2022-04-03 02:52 178阅读 0赞

简介：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 1][]

**本质上在Actions算子中通过SparkContext执行提交作业的runJob操作，触发了RDD DAG的执行。**   
所以本文从**runJob方法进行解析 从而引出DAGScheduler、TaskScheduler 。。。。**

代码部分：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 2][]

提交流程图：

![webp][]

调度流程：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 3][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 4][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 5][]

源码解析：

直接看RDD的 count()方法吧

<table> 
 <tbody> 
  <tr> 
   <td> <pre><code>/**
 * Return the number of elements in the RDD.
 */
def count(): Long = sc.runJob(this, Utils.getIteratorSize _).sum
</code></pre> </td> 
  </tr> 
 </tbody> 
</table>

经历的步骤：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 6][]

跳转了好几个**runJob 最后执行这个runJob方法**

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 7][]

核心 向scheduler 提交一个job ，并返回一个jobwaiter ：

val waiter = submitJob(rdd, func, partitions, callSite, resultHandler, properties)

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 8][]

private[scheduler] val eventProcessLoop = new DAGSchedulerEventProcessLoop(this)

eventProcessLoop.post(JobSubmitted(
      jobId, rdd, func2, partitions.toArray, callSite, waiter,
      SerializationUtils.clone(properties)))

eventProcessLoop接收各种消息并进行处理，处理的逻辑在其doOnReceive方法中：

向eventProcessLoop 申请调用**JobSubmitted 方法。**

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 9][]

接下来查看 handleJobSubmitted 这个方法。主要分三块：

# 1.创建ResultStage  #

# Stage划分过程是从最后一个Stage开始 。 #

# 反向驱动，正向提交 #

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 10][]

接下来看 ：

# createResultStage #

## 根据给定的jobid 创建ResultStage ##

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 11][]

## 这个方法核心的 就是创建getOrCreateParentStages 方法 ##

# #

根据给定的RDD获取或者创建父stages列表 ，新的stage会根据提供的firstJobId进行创建

## 接下来看getOrCreateShuffleMapStage 方法，这个有点复杂 ##

分两块，一个是

getShuffleDependencies

另一个是

getOrCreateShuffleMapStage(shuffleDep, firstJobId)

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 12][]

# 先看getShuffleDependencies，这个有点复杂，类似递归( 只返回到上一个shuffle rdd ，并不是所有的shuffle RDD #

# ) #

这是一个递归方法，获取shuffle依赖的集合 getShuffleDependencies

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 13][]

## ![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 14][] ##

## 创建ShuffleMapStage   ##

createShuffleMapStage

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 15][]

上面的方法，会依次递归，循环调用 建议用dub模式跟踪一下，后面我会写一篇文章，专门写stage划分的

### 回到createResultStage 方法 ###

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 16][]

回到 handleJobSubmitted 方法

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 17][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 18][]

为了更清晰的了解 ，划分了下面的图：

\------------------------------------------------------------------------------------------------------------------------------------

RDDs原始依赖图

![Image 1][]

RDDs原始依赖图

getShuffleDependencies  获取RDD的第一层直接宽依赖

![Image 1][]

getMissingAncestorShuffleDependencies

深度遍历顺序获取所有祖先的宽依赖

![Image 1][]

最后划分结果：

![Image 1][]

\---------------------------下面是提交任务， 正向提交--------------------------------------------------------------------------

提交任务。。。

### logInfo("DAGScheduler\#handleJobSubmitted：submitStage提交stage。。。。。。。")   submitStage(finalStage) ###

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 19][]

### 这里，就是依次将尚未提交的任务，加入任务队列waitingStages ###

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 20][]

### submitMissingTasks 这个方法比较长，直接看注释吧![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 21][] ###

这个是返回task的最佳位置， 其实就是那台机器

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 22][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 23][]

为每一个ShuffleMapStage的分区 创建一个 ShuffleMapTask

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 24][]

提交任务：

### submitWaitingChildStages ###

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 25][]

###  其实就是将waitingStages队列中的任务，递归提交 ###

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 26][]

递归提交，最终执行的方法是：

### org.apache.spark.scheduler.TaskSchedulerImpl\#submitTasks ###

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 27][]

最终交由taskScheduler进行提交

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 28][]

# 记住：反向划分，正向提交 #

迷糊迷糊^\_^

参考资源：

[https://www.jianshu.com/p/14355e250e2f][https_www.jianshu.com_p_14355e250e2f]

《spark内核设计的艺术架构与实现》 耿嘉安

[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70]: /images/20220403/a57fcde8291d4a3fb5bff0a5f4e60013.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 1]: /images/20220403/31ec748d4cd24dafbaf43fe063cbfa66.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 2]: /images/20220403/cf9c7efab9474f83a4c9796068443667.png
[webp]: /images/20220403/01a6142614e44db4b625e88917545704.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 3]: /images/20220403/1551f43af5af4d5394413991d40d3c76.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 4]: /images/20220403/2846d41fbdf1408194e9ed639eea9e39.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 5]: /images/20220403/cbfc6b6d58d24027936ab5982d9b5166.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 6]: /images/20220403/e75767a041714599898d95a649eb4634.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 7]: /images/20220403/f6fc8477c92d42609120ee07ea7b10cb.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 8]: /images/20220403/d91508beb1e84b858aaa3e2a84e40bdf.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 9]: /images/20220403/4d28cb3a3de142329481c5e50a599df6.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 10]: /images/20220403/ad8b86a5529640c38ec4f14fe4447299.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 11]: /images/20220403/54b2784d477c44ac9b62f00d4d591e9d.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 12]: /images/20220403/49b38f9f30b447f4aae4f84d3d73fae5.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 13]: /images/20220403/ab72bef057374b9b8adc87498ac60cc0.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 14]: /images/20220403/b8016f0905094a4f8e1bbec75207dabb.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 15]: /images/20220403/aa8080ad52db4f499eb33e235ce45e8b.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 16]: /images/20220403/77f385bf35bc42db81302dc13144b225.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 17]: /images/20220403/5899efc93ddc461da008ba131c06f171.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 18]: /images/20220403/0197a93f218444aab42d344908552090.png
[Image 1]: 
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 19]: /images/20220403/8107e0177dc246178d1dc11aac34e12d.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 20]: /images/20220403/296f750fdb424fe19ebb3916b3b31ad0.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 21]: /images/20220403/77996d19a8db473e93bccc7537058130.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 22]: /images/20220403/d4091baf81284a29970c1a84e50aa925.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 23]: /images/20220403/e4ff86509f2b41018cbe822b43eb6463.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 24]: /images/20220403/41bc031e258b4755bb3ba26d0728dab2.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 25]: /images/20220403/4053037c34044640a9478136b19e1a2f.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 26]: /images/20220403/3fe0f16f7adb4210ae32e163a4e86ee0.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 27]: /images/20220403/10c481aca663426383eabf7be6efd5a4.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3poYW5nbG9uZ180NDQ0_size_16_color_FFFFFF_t_70 28]: /images/20220403/ce2701e07b894ddba78fece42bc9d727.png
[https_www.jianshu.com_p_14355e250e2f]: https://www.jianshu.com/p/14355e250e2f