如何将 Spark 客户端 submitApplication 转换为 Yarn Rest API?

How can I translate a Spark Client submitApplication to Yarn Rest API?

目前我有一个使用 spark.deploy.yarn.Client 向 Yarn 提交应用程序的工作代码实现。聚合此客户需要的所有参数很复杂,但提交申请很简单:

ClientArguments cArgs = new ClientArguments(args.toArray(new String[0]));
client = new Client(cArgs, sparkConf);
applicationID = client.submitApplication();

在此之前的大部分代码都在累积 sparkConfargs。现在我希望停用 Client 并仅使用 Rest。 Spark 提供完整的 REST api,包括提交应用程序 - 根据 Spark documentation,这很简单 json/xml post:

POST http://<rm http address:port>/ws/v1/cluster/apps
Accept: application/json
Content-Type: application/json
{
  "application-id":"application_1404203615263_0001",
  "application-name":"test",
  "am-container-spec":
{
  "local-resources":
  {
    "entry":
    [
      {
        "key":"AppMaster.jar",
        "value":
        {
          "resource":"hdfs://hdfs-namenode:9000/user/testuser/DistributedShell/demo-app/AppMaster.jar",
          "type":"FILE",
          "visibility":"APPLICATION",
          "size": 43004,
          "timestamp": 1405452071209
        }
      }
    ]
  },
  "commands":
  {
    "command":"{{JAVA_HOME}}/bin/java -Xmx10m org.apache.hadoop.yarn.applications.distributedshell.ApplicationMaster --container_memory 10 --container_vcores 1 --num_containers 1 --priority 0 1><LOG_DIR>/AppMaster.stdout 2><LOG_DIR>/AppMaster.stderr"
  },
  "environment":
  {
    "entry":
    [
      {
        "key": "DISTRIBUTEDSHELLSCRIPTTIMESTAMP",
        "value": "1405459400754"
      },
      {
        "key": "CLASSPATH",
        "value": "{{CLASSPATH}}<CPS>./*<CPS>{{HADOOP_CONF_DIR}}<CPS>{{HADOOP_COMMON_HOME}}/share/hadoop/common/*<CPS>{{HADOOP_COMMON_HOME}}/share/hadoop/common/lib/*<CPS>{{HADOOP_HDFS_HOME}}/share/hadoop/hdfs/*<CPS>{{HADOOP_HDFS_HOME}}/share/hadoop/hdfs/lib/*<CPS>{{HADOOP_YARN_HOME}}/share/hadoop/yarn/*<CPS>{{HADOOP_YARN_HOME}}/share/hadoop/yarn/lib/*<CPS>./log4j.properties"
      },
      {
        "key": "DISTRIBUTEDSHELLSCRIPTLEN",
        "value": "6"
      },
      {
        "key": "DISTRIBUTEDSHELLSCRIPTLOCATION",
        "value": "hdfs://hdfs-namenode:9000/user/testuser/demo-app/shellCommands"
      }
    ]
  }
},
"unmanaged-AM":false,
"max-app-attempts":2,
"resource":
{
  "memory":1024,
  "vCores":1
},
"application-type":"YARN",
"keep-containers-across-application-attempts":false,
"log-aggregation-context":
{
  "log-include-pattern":"file1",
  "log-exclude-pattern":"file2",
  "rolled-log-include-pattern":"file3",
  "rolled-log-exclude-pattern":"file4",
  "log-aggregation-policy-class-name":"org.apache.hadoop.yarn.server.nodemanager.containermanager.logaggregation.AllContainerLogAggregationPolicy",
  "log-aggregation-policy-parameters":""
},
"attempt-failures-validity-interval":3600000,
"reservation-id":"reservation_1454114874_1",
"am-black-listing-requests":
{
  "am-black-listing-enabled":true,
  "disable-failure-threshold":0.01
}
}

我试图将我的论点翻译成 POST 请求的 JSON 正文,但这似乎是不可能的。有谁知道我是否可以从我提交的 JSON 有效载荷通过 REST 发送的 运行 应用程序进行逆向工程?或者我可以使用什么映射来获取客户端参数并将它们放在 JSON?

经过一番搜索,我设法仅通过 REST API 提交了申请。这不是一个有据可查的过程,所以我把它贴在这里。

注意:如果任何时候您希望将请求的内容与客户端发送的请求进行比较,请使用调试断点来检查客户端使用的应用程序上下文. 打开 class org.apache.hadoop.yarn.client.api.impl.YarnClientImpl 并转到方法 submitApplication(ApplicationSubmissionContext appContext).

首先,要用 REST API 请求替换 spark.deploy.yarn.Client,解决方案必须确保配置中提到的所有文件在 HDFS 上可用。 稍后,它需要编写并上传一个名为 __spark_conf__.zip 的额外文件。

步骤 1

查看 SparkConfClient 的第二个参数)中的文件:“AllJars”标签中提到的文件, “mainJarPath”,以及“FilesList”中提到的文件。

对于每个文件,检查它是否存在于 HDFS 中,如果不存在 - 从本地计算机上传。对于每个文件,从 HDFS 获取其 FileStatus。 聚合资源列表,它是包含这 6 个属性的每个文件的属性映射:

  • 大小 = getLen()
  • 时间戳 = getModificationTime()
  • 类型=文件
  • 能见度=PUBLIC

另外两个属性:密钥和资源。

  • 来自 allJars 列表的文件:键是 spark_libs/{{filename}},资源是文件名。
  • FilesList 中的文件:键是 "localEntry" 标签,资源是 "hdfsPath" 标签。
  • mainJarPath 中的文件:key 是“app.jar”,resource 是文件名。

步骤 2

正在创建 __spark_conf__.zip 文件。您可以直接在 hdfs 中创建它,在通常为 {{HDFS_base_folder}}/user/{{username}}/.sparkStaging/{{application_id}}/__spark_conf__.zip 的暂存路径中。 此存档文件包含两个文件和一个空目录:一个文件 __spark_hadoop_conf__.xml(重命名为 core-site.xml),另一个文件名为 __spark_conf__.properties,这是一个略微修改的版本 来自配置的 sparkConf 部分。

要创建 __spark_conf__.properties,您需要从“sparkConf”->"org$apache$spark$SparkConf$$settings" 读取 JSON 映射,并转换 [=200] 中的每一行=] 格式 "spark.safemine.addcontrol.driverMemory": "5120M"spark.safemine.addcontrol.driverMemory=5120M

在文件底部添加 6 行:

  • spark.yarn.cache.confArchive={{您将在 sparkStaging __spark_conf__.zip 中上传的位置}}
  • spark.yarn.cache.visibilities={{文件的所有可见性,逗号分隔 - 基本上 "PUBLIC,PUBLIC, ... ,PUBLIC"}}
  • spark.yarn.cache.timestamps={{文件的所有时间戳,以逗号分隔}}
  • spark.yarn.cache.types={{文件的所有类型,逗号分隔 - 基本上 "FILE,FILE, ... ,FILE"}}
  • spark.yarn.cache.filenames={{所有文件名和密钥,记录为资源#key 和逗号分隔}}
  • spark.yarn.cache.sizes={{文件的所有大小,以逗号分隔}}

确保按各自的顺序编译 5 行汇总。我使用了这段代码:

    String confArchive = "spark.yarn.cache.confArchive="+hdfs+"/user/"+userName+"/.sparkStaging/"+applicationId+"/__spark_conf__.zip";
    String filenames = "spark.yarn.cache.filenames=";
    String sizes = "spark.yarn.cache.sizes=";
    String timestamps = "spark.yarn.cache.timestamps=";
    String types = "spark.yarn.cache.types=";
    String visibilities = "spark.yarn.cache.visibilities=";
    for (Map<String,String> localResource:localResources) {
        filenames+=localResource.get("resource")+"#"+localResource.get("key")+",";
        sizes+=localResource.get("size")+",";
        timestamps+=localResource.get("timestamp")+",";
        types+=localResource.get("type")+",";
        visibilities+=localResource.get("visibility")+",";

    }
    properties+=confArchive+"\n";
    properties+=filenames.substring(0,filenames.length()-1)+"\n";
    properties+=sizes.substring(0,sizes.length()-1)+"\n";
    properties+=timestamps.substring(0,timestamps.length()-1)+"\n";
    properties+=types.substring(0,types.length()-1)+"\n";
    properties+=visibilities.substring(0,visibilities.length()-1)+"\n";

__spark_hadoop_conf__.xml文件是core-site.xml的简单重命名,用它们创建的文件夹名为__hadoop_conf__并留空。

你可以像这样直接将文件保存到hdfs:

private void generateSparkConfInHdfs(String applicationId, String userName, String sparkConfProperties, String sparkHadoopConf) throws IOException {
    String path = hdfs+"/user/"+userName+"/.sparkStaging/"+applicationId+"/__spark_conf__.zip";
    Path hdfsPath = new Path(path);
    ZipOutputStream os = new ZipOutputStream(getHdfs().create(hdfsPath));
    os.putNextEntry(new ZipEntry("__hadoop_conf__/"));
    os.putNextEntry(new ZipEntry("__spark_conf__.properties"));
    os.write(sparkConfProperties.getBytes(),0,sparkConfProperties.getBytes().length);
    os.putNextEntry(new ZipEntry("__spark_hadoop_conf__.xml"));
    os.write(sparkHadoopConf.getBytes(),0,sparkHadoopConf.getBytes().length);
    os.close();
}

创建完文件后,将其添加到具有以下规格的资源列表中:

  • 大小 = getLen()
  • 时间戳 = getModificationTime()
  • 类型=存档
  • 可见性 = 私有
  • 关键=__spark_conf__
  • 资源是暂存目录(通常{{HDFS_base_folder}}/user/{{username}}/.sparkStaging/{{application_id}}/__spark_conf__.zip)。

查看完整的资源列表并使用我们在 {{}} 占位符中收集的值为每个资源创建一个 XML/JSON:

    <entry>
        <key>{{key}}</key>
        <value>
            <resource>{{resource}}</resource>
            <size>{{size}}</size>
            <timestamp>{{timestamp}}</timestamp>
            <type>{{type}}</type>
            <visibility>{{visibility}}</visibility>
        </value>
    </entry>

累积的字符串将是您的 localResources XML 段,如下所示。

步骤 3

正在生成 Java 命令。您需要从 SparkConfig 中提取一些元素:

  • driverMemory - 来自 sparkConf
  • 中的相同属性
  • 额外Java选项 = 来自 spark.driver.extraJavaOptions 属性集合
  • mainClass - 来自 sparkConf
  • 中的相同属性
  • argstr - 收集除 --class 之外的所有 ClientArgs

包含元素的结果命令为:

String command = "$JAVA_HOME/bin/java -server -Xmx"+driverMemory+" -Djava.io.tmpdir=$PWD/tmp "+extraJavaOptions+" -Dspark.yarn.app.container.log.dir=&lt;LOG_DIR&gt; "
            + "org.apache.spark.deploy.yarn.ApplicationMaster --class "+mainClass+" "+argstr+" "
            + "--properties-file $PWD/__spark_conf__/__spark_conf__.properties 1&gt; &lt;LOG_DIR&gt;/stdout 2&gt; &lt;LOG_DIR&gt;/stderr";

步骤 4

编译请求XML。

注意:我的实现需要AM容器上的标签,所以添加了am-container-node-label-expression。这并不适用于所有情况。

从 sparkConf 到 REST 请求的映射是(在 XML 中显示,JSON 实现也受支持):

<application-submission-context>
    <application-id>"+applicationId+"</application-id> 
    <application-name>"+appName+"</application-name>
    <queue>default</queue>
    <priority>0</priority>
    <am-container-spec>
       <local-resources>+localResources+</local-resources>
       <environment>
          <entry>
             <key>SPARK_YARN_STAGING_DIR</key>
             <value>"+hdfs+"/user/"+userName+"/.sparkStaging/"+applicationId+"</value>
          </entry>
          <entry>
             <key>CLASSPATH</key>
             <value>$PWD:$PWD/__spark_conf__:$PWD/__spark_libs__/*:/spark-non-hdfs-storage/spark-assembly-2.3.0-hadoop2.7/*:%HADOOP_CONF_DIR%:%HADOOP_COMMON_HOME%/share/hadoop/common/*:%HADOOP_COMMON_HOME%/share/hadoop/common/lib/*:%HADOOP_HDFS_HOME%/share/hadoop/hdfs/*:%HADOOP_HDFS_HOME%/share/hadoop/hdfs/lib/*:%HADOOP_YARN_HOME%/share/hadoop/yarn/*:%HADOOP_YARN_HOME%/share/hadoop/yarn/lib/*:%HADOOP_MAPRED_HOME%/share/hadoop/mapreduce/*:%HADOOP_MAPRED_HOME%/share/hadoop/mapreduce/lib/*:$PWD/__spark_conf__/__hadoop_conf__</value>
          </entry>
          <entry>
             <key>SPARK_USER</key>
             <value>"+userName+"</value>
          </entry>
       </environment>
       <commands>
          <command>"+command+"</command>
       </commands>
    </am-container-spec>
    <unmanaged-AM>false</unmanaged-AM>
    <max-app-attempts>1</max-app-attempts>
    <resource>
      <memory>5632</memory>
      <vCores>1</vCores>
    </resource>
    <application-type>SPARK</application-type>
    <keep-containers-across-application-attempts>false</keep-containers-across-application-attempts>
    <application-tags>
      <tag>"+sparkYarnTag+"</tag>
    </application-tags>
    <am-container-node-label-expression>appMngr</am-container-node-label-expression>
    <log-aggregation-context/>
    <attempt-failures-validity-interval>1</attempt-failures-validity-interval>
    <reservation-id/>
</application-submission-context> 

第 5 步:

正在通过 REST http PUT 提交应用程序:

private void submitApplication (String body, String userName) throws SMSparkManagerException {
    HttpClient client = HttpClientBuilder.create().build();
    HttpPost request = new HttpPost(uri+"?user.name="+userName);
    try {
        request.setEntity(new StringEntity(body, ContentType.APPLICATION_XML));
        HttpResponse response = client.execute(request);
        if (response.getStatusLine().getStatusCode()!=202) {
            throw new SMSparkManagerException("The application could not be submitted to Yarn, response http code "+response.getStatusLine().getStatusCode());
        }
    } catch (UnsupportedEncodingException e) {
        logger.error("The application Could not be submitted due to UnsupportedEncodingException in the provided body: "+body, e );
        throw new SMSparkManagerException("Error in submitting application to yarn");
    } catch (ClientProtocolException e) {
        logger.error("The application Could not be submitted due to ClientProtocolException", e);
        throw new SMSparkManagerException("Error in submitting application to yarn");
    } catch (IOException e) {
        logger.error("The application Could not be submitted due to IOException", e);
        throw new SMSparkManagerException("Error in submitting application to yarn");
    }
}