HDPCD-Java-复习笔记（20）

客官°小女子只卖身不卖艺 2022-06-07 11:37 107阅读 0赞

**Defining Workflow**

**Overview of Workflow**

Workflow in Hadoop refers to the scheduling and coordination of MapReduce jobs.

Orchestration of MapReduce jobs can be accomplished in several ways, including:

<table> 
 <tbody> 
  <tr> 
   <td> <p><strong>Linear Chain of MapReduce Jobs</strong></p> </td> 
   <td> <p>Use the return value of&nbsp;<span style="color:#666666">waitForCompletion</span>&nbsp;to determine the success of each job in the chain</p> </td> 
  </tr> 
  <tr> 
   <td> <p><strong>JobControl</strong></p> </td> 
   <td> <p>A class in the MapReduce API that coordinates multiple jobs</p> </td> 
  </tr> 
  <tr> 
   <td> <p><strong>Apache Oozie</strong></p> </td> 
   <td> <p>A workflow scheduler system for managing jobs and also for scheduling recurring jobs</p> </td> 
  </tr> 
 </tbody> 
</table>

**Linear Chain**

**![Center][]**

The linear chain workflow is straightforward. Start a job in a run method of a Tool instance, wait for the job to complete, and then start subsequent jobs in the same run method. Continue until all jobs are complete.

For example:

Job job1 = Job.getInstance(getConf(), "CreateBloomFilter");
    boolean job1success = job1.waitForCompletion(true);
    if(!job1success) {
      System.out.println("The CreateBloomFilter job failed!");
      return -1;
    }
    Job job2 = Job.getInstance(conf, "FilterStocksJob");
    boolean job2success = job2.waitForCompletion(true);
    if(!job2success) {
      System.out.println("The FilterStocksJob failed!");
      return -1;
    }
    return 1;

**The JobControl Class**

The org.apache.hadoop.mapreduce.lib.jobcontrol.JobControl class provides the ability to coordinate MapReduce jobs and define dependencies.  
Instead of invoking run on a Job instance**, you add a Job instance to a JobControl instance, along with defining any dependencies for the Job.**  
For example:

JobControl control = new JobControl("bloom");
    ControlledJob cjob1 = new ControlledJob(job1, null);
    List<ControlledJob> dependentJobs =
    	   new ArrayList<ControlledJob>();
    dependentJobs.add(cjob1);
    ControlledJob cjob2 =
    	  new ControlledJob(job2, dependentJobs);
    control.addJob(cjob1);
    control.addJob(cjob2);
    new Thread(control).start();
       
    while (!control.allFinished()) {
       Thread.sleep(1000);
    }
    control.stop();

• A JobControl manages ControlledJob instances.  
• The addJob method is used to add a ControlledJob to a JobControl.  
• A ControlledJob has two main components: a Job, and a List of that Job’s dependencies.  
• A Job will not be started until all its dependencies are in the SUCCESS state.  
• The JobControl class implements Runnable. To run the jobs in the JobControl instance, wrap it in a Thread and then start the Thread.

ControlledJob has a static inner enum named State for the various states that a ControlledJob can be in. The enum values are:

•WAITING  
•READY  
•RUNNING  
•SUCCESS  
•FAILED  
•DEPENDENT\_FAILED  
Use the getJobState method to retrieve the state of a ControlledJob.

**Overview of Oozie**

Oozie provides a framework for coordinating and scheduling Hadoop jobs. Oozie is not restricted to just **MapReduce** jobs; Oozie can schedule **Pig**, **Hive**, **Sqoop**, **streaming jobs**, and even **Java programs**.

Oozie has two main capabilities:

**Oozie Workflow**\-- **A collection of actions**, defined in a **workflow.xml**file  
**Oozie Coordinator**\-- **A recurring workflow**, defined in a**coordinator.xml** file

Oozie is a Java web application that runs in a Tomcat instance. Oozie is run as a service, and workflows are started using the oozie command.

**Defining an Oozie Workflow**  
![Center 1][]

An Oozie workflow consists of **a workflow.xml file and the files required by the workflow**.The workflow is put into HDFS with the following directory structure:

*  /appdir/workflow.xml
 *  /appdir/config-default.xml
 *  /appdir/lib/files.jar

The config-default.xml file is optional, and contains properties shared by all workflows.

Each workflow can also have a job.properties file (not put into HDFS) for job-specific properties.

A workflow definition consists of **two main entries**:

**Control flow nodes** \--- Jobs can contain **fork or join nodes**. A **fork node** splits one path into multiple paths. A **join node** waits until every path of a previous fork node arrives to it.  
**Action nodes** \--- For executing a job or task.

Consider this example:

<?xml version="1.0" encoding="utf-8"?>
    
    <workflow-app xmlns="uri:oozie:workflow:0.2" name="payroll-workflow">  
      <start to="payroll-job"/>  
      <action name="payroll-job"> 
        <map-reduce> 
          <job-tracker>${resourceManager}</job-tracker>  
          <name-node>${nameNode}</name-node>  
          <prepare> 
            <delete path="${nameNode}/user/${wf:user()}/payroll/result"/> 
          </prepare>  
          <configuration> 
            <property> 
              <name>mapreduce.job.queuename</name>  
              <value>${queueName}</value> 
            </property>  
            <property> 
              <name>mapred.mapper.new-api</name>  
              <value>true</value> 
            </property>  
            <property> 
              <name>mapred.reducer.new-api</name>  
              <value>true</value> 
            </property>  
            <property> 
              <name>mapreduce.job.map.class</name>  
              <value>payroll.PayrollMapper</value> 
            </property>  
            <property> 
              <name>mapreduce.job.reduce.class</name>  
              <value>payroll.PayrollReducer</value> 
            </property>  
            <property> 
              <name>mapreduce.job.inputformat.class</name>  
              <value>org.apache.hadoop.mapreduce.lib.input.TextInputFormat</value> 
            </property>  
            <property> 
              <name>mapreduce.job.outputformat.class</name>  
              <value>org.apache.hadoop.mapreduce.lib.output.NullOutputFormat</value> 
            </property>  
            <property> 
              <name>mapreduce.job.output.key.class</name>  
              <value>payroll.EmployeeKey</value> 
            </property>  
            <property> 
              <name>mapreduce.job.output.value.class</name>  
              <value>payroll.Employee</value> 
            </property>  
            <property> 
              <name>mapreduce.job.reduces</name>  
              <value>20</value> 
            </property>  
            <property> 
              <name>mapreduce.input.fileinputformat.inputdir</name>  
              <value>${nameNode}/user/${wf:user()}/payroll/input</value> 
            </property>  
            <property> 
              <name>mapreduce.output.fileoutputformat.outputdir</name>  
              <value>${nameNode}/user/${wf:user()}/payroll/result</value> 
            </property>  
            <property> 
              <name>taxCode</name>  
              <value>${taxCode}</value> 
            </property> 
          </configuration> 
        </map-reduce>  
        <ok to="compute-tax"/>  
        <error to="fail"/> 
      </action>  
      <action name="compute-tax"> 
        <map-reduce> 
           
        </map-reduce>  
        <ok to="end"/>  
        <error to="fail"/> 
      </action>  
      <kill name="fail"> 
        <message>Job failed, error message[${wf:errorMessage(wf:lastErrorNode())}]</message> 
      </kill>  
      <end name="end"/> 
    </workflow-app>

•  Every workflow must define a <start> and <end>.  
•  This workflow has two action nodes: payroll-job and compute-tax.  
•  Notice a workflow looks almost identical to a run method of a MapReduce job, except that the job properties are specified in XML.  
•  The <delete> function is a convenient way to delete an existing output folder.

Parameters use the**$\{\}** syntax and represent values defined outside of workflow.xml. For example, $\{resourceManager\} is the server name and port number where the ResourceManager is running. Instead of hard-coding this value, **you define it in** an external properties file (the **job.properties** file).

wf:user(), which returns the name of the user running the job, and wf:lastErrorNode(), which returns the NodeManager where the mostrecent error occurred.

Using the New MapReduce API in an Oozie Workflow

•<property>  
• <name>mapred.mapper.new-api</name>  
•</property>  
•<property>  
• <name>mapred.reducer.new-api</name>  
• <value>true</value>  
•</property>

**Pig Actions**

Here is an example of a simple workflow that only contains a single Pig script asone of its actions:

<?xml version="1.0" encoding="utf-8"?>
    
    <workflow-app xmlns="uri:oozie:workflow:0.2" name="whitehouse-workflow">  
      <start to="transform_whitehouse_visitors"/>  
      <action name="transform_whitehouse_visitors"> 
        <pig> 
          <job-tracker>${resourceManager}</job-tracker>  
          <name-node>${nameNode}</name-node>  
          <prepare> 
            <delete path="wh_visits"/> 
          </prepare>  
          <script>whitehouse.pig</script> 
        </pig>  
        <ok to="end"/>  
        <error to="fail"/> 
      </action>  
      <kill name="fail"> 
        <message>Job failed, error message[${wf:errorMessage(wf:lastErrorNode())}]</message> 
      </kill>  
      <end name="end"/> 
    </workflow-app>

Hive Actions

<?xml version="1.0" encoding="utf-8"?>
    
    <action name="find_congress_visits"> 
      <hive xmlns="uri:oozie:hive-action:0.5">  
        <job-tracker>${resourceManager}</job-tracker>  
        <name-node>${nameNode}</name-node>  
        <prepare> 
          <delete path="congress_visits"/> 
        </prepare>  
        <job-xml>hive-site.xml</job-xml>  
        <configuration> 
          <property> 
            <name>mapreduce.map.output.compress</name>  
            <value>true</value> 
          </property> 
        </configuration>  
        <script>congress.hive</script> 
      </hive>  
      <ok to="end"/>  
      <error to="fail"/> 
    </action>

**Submitting a Workflow Job**

Oozie has a command line tool named oozie for submitting and executing workflows. The command looks like:

• \# oozie job -config job.properties -run

Note that the workflow is typically deployed in HDFS, while job.properties, which contains the properties passed into the workflow, is typically kept on the local filesystem.

The command above does not specify which Oozie Workflow to execute. This isspecified by the oozie.wf.application.path property:

oozie.wf.application.path=hdfs://node:8020/path/to/app

Here is an example of a job.properties file:

*  oozie.wf.application.path=hdfs://node:8020/path/to/app
 *  resourceManager=node:8050
 *  \#Hadoop fs.defaultFS
 *  nameNode=hdfs://node:8020/
 *  \#Hadoop mapreduce.job.queuename
 *  queueName=default
 *  \#Application-specific properties
 *  taxCode=2012

**Fork and Join Nodes**

Oozie has fork and join nodes for controlling workflow. For example:

**Defining an Oozie Coordinator Job**

The Oozie Coordinator component enables recurring Oozie Workflows to be defined.These recurring jobs can be triggered by two types of events:

<table> 
 <tbody> 
  <tr> 
   <td> <p><strong>Time</strong></p> </td> 
   <td> <p>Similar to a&nbsp;<span style="color:#666666"><strong>cron</strong></span>&nbsp;job</p> </td> 
  </tr> 
  <tr> 
   <td> <p><strong>Data Availability</strong></p> </td> 
   <td> <p>The job triggers when a specified directory is created</p> </td> 
  </tr> 
 </tbody> 
</table>

An Oozie Coordinator job consists of two files:

<table> 
 <tbody> 
  <tr> 
   <td> <p><span style="color:#666666">coordinator.xml</span></p> </td> 
   <td> <p>The definition of the Coordinator application</p> </td> 
  </tr> 
  <tr> 
   <td> <p><span style="color:#666666">coordinator.properties</span></p> </td> 
   <td> <p>For defining the job’s properties</p> </td> 
  </tr> 
 </tbody> 
</table>

**Schedule a Job Based on Time**  
Consider this example of a coordinator.xml file.

<?xml version="1.0" encoding="utf-8"?>
    
    <coordinator-app xmlns="uri:oozie:coordinator:0.1"

name="tf-idf" frequency="1440" start="2013-01-01T00:00Z" end="2013-12-31T00:00Z" timezone="UTC">  
      <action> 
        <workflow> 
          <app-path>hdfs://node:8020/root/java/tfidf/workflow</app-path> 
        </workflow> 
      </action> 
    </coordinator-app>

The frequency is in minutes, so this job runs once a day.

You submit an Oozie Coordinator job similar to submitting a Workflow job:  
•\# oozie job -config coordinator.properties -run

The coordinator.properties file contains the path to thecoordinator app:

oozie.coord.application.path=hdfs://node:8020/path/to/app

**Schedule a Job Based on Data Availability**  
The following Coordinator application triggers a Workflow job when  **a directory** named hdfs://node :8020 /job/result/ **gets created**:

<?xml version="1.0" encoding="utf-8"?>
    
    <coordinator-app xmlns="uri:oozie:coordinator:0.1"

name="file_check"

frequency="1440" start="2012-01-01T00:00Z" end="2015-12-31T00:00Z" timezone="UTC">  
      <datasets> 
        <dataset name="input1"> 
          <uri-template>hdfs://node:8020/job/result/</uri-template> 
        </dataset> 
      </datasets>  
      <action> 
        <workflow> 
          <app-path>hdfs://node:8020/myapp/</app-path> 
        </workflow> 
      </action> 
    </coordinator-app>

• This Coordinator app is scheduled to run once a day.  
• If the folder hdfs://node :8020 /job/result/ exists, the <action> executes, which in this example is an Oozie Workflow deployed in the hdfs://node:8020/myapp folder.  
• The use-case scenario for this example posits that a separate MapReduce job executes once a day at an unspecified time. When that job runs, it deletes the hdfs://node :8020/job/result directory, and then creates a new one, which triggers the Coordinator to run.  
• This Coordinator runs once a day, and if /job/result exists, then the /myapp workflow will execute.

[Center]: /images/20220607/b8fc2690f811419e849a8f63ccf843a9.png
[Center 1]: /images/20220607/6d2d6a4eff8f49b690c2ab655262ce8b.png