【k8s系列2】spark on k8s(kubernetes) DynamicResourceAllocation(DRA)

 随着大数据时代的到来，以及kubernetes的愈发火热，好多公司已经把spark应用从yarn迁移到k8s，当然也踩了不少的坑，    
 现在我们来分析一下spark on k8s的DynamicResourceAllocation这个坑

注意：该文基于spark 3.0.0分析

spark on yarn 中的DynamicResourceAllocation

spark on yarn对于DynamicResourceAllocation分配来说，从spark 1.2版本就已经开始支持了.
对于spark熟悉的人都知道，如果我们要开启DynamicResourceAllocation，就得有ExternalShuffleService服务，
对于yarn来说ExternalShuffleService是作为辅助服务开启的，具体配置如下：

<property>
   <name>yarn.nodemanager.aux-services</name>
   <value>spark_shuffle</value>
</property>

<property>
   <name>yarn.nodemanager.aux-services.spark_shuffle.class</name>
   <value>org.apache.spark.network.yarn.YarnShuffleService</value>
</property>

<property>
   <name>spark.shuffle.service.port</name>
   <value>7337</value>
</property>

重启nodeManager，这样在每个nodeManager节点就会启动一个YarnShuffleService，之后在spark应用中设置spark.dynamicAllocation.enabled 为true,这样就能达到运行时资源动态分配的效果

我们直接从CoarseGrainedExecutorBackend中SparkEnv创建开始说，每一个executor的启动，必然会经过CoarseGrainedExecutorBackend main方法,而main中就涉及到SparkEnv的创建

 val env = SparkEnv.createExecutorEnv(driverConf, arguments.executorId, arguments.bindAddress,
       arguments.hostname, arguments.cores, cfg.ioEncryptionKey, isLocal = false)

而sparkEnv的创建就涉及到BlockManager的创建。沿着代码往下走，最终

val blockTransferService =
     new NettyBlockTransferService(conf, securityManager, bindAddress, advertiseAddress,
       blockManagerPort, numUsableCores, blockManagerMaster.driverEndpoint)
val blockManager = new BlockManager(
     executorId,
     rpcEnv,
     blockManagerMaster,
     serializerManager,
     conf,
     memoryManager,
     mapOutputTracker,
     shuffleManager,
     blockTransferService,
     securityManager,
     externalShuffleClient)

在blockManager的initialize方法中,就会进行registerWithExternalShuffleServer

 // Register Executors' configuration with the local shuffle service, if one should exist.
   if (externalShuffleServiceEnabled && !blockManagerId.isDriver) {
     registerWithExternalShuffleServer()
   }

如果我们开启了ExternalShuffleService，对于yarn就是YarnShuffleService，就会把当前的ExecutorShuffleInfo注册到host为shuffleServerId.host, port为shuffleServerId.port的ExternalShuffleService中,ExecutorShuffleInfo的信息如下：

val shuffleConfig = new ExecutorShuffleInfo(
     diskBlockManager.localDirsString,
     diskBlockManager.subDirsPerLocalDir,
     shuffleManager.getClass.getName)

这里我重点分析一下registerWithExternalShuffleServer的方法中的以下片段

// Synchronous and will throw an exception if we cannot connect.
       blockStoreClient.asInstanceOf[ExternalBlockStoreClient].registerWithShuffleServer(
         shuffleServerId.host, shuffleServerId.port, shuffleServerId.executorId, shuffleConfig)            

该代码中shuffleServerId来自于：

shuffleServerId = if (externalShuffleServiceEnabled) {
     logInfo(s"external shuffle service port = $externalShuffleServicePort")
     BlockManagerId(executorId, blockTransferService.hostName, externalShuffleServicePort)
   } else {
     blockManagerId
   }

而blockTransferService.hostName 是我们在SparkEnv中创建的时候由advertiseAddress传过来的,
最终由CoarseGrainedExecutorBackend 主类参数hostname过来的，那到底怎么传过来的呢？
参照ExecutorRunnable的prepareCommand方法，

val commands = prefixEnv ++
     Seq(Environment.JAVA_HOME.$$() + "/bin/java", "-server") ++
     javaOpts ++
     Seq("org.apache.spark.executor.YarnCoarseGrainedExecutorBackend",
       "--driver-url", masterAddress,
       "--executor-id", executorId,
       "--hostname", hostname,
       "--cores", executorCores.toString,
       "--app-id", appId,
       "--resourceProfileId", resourceProfileId.toString) ++

而这个hostname的值最终由YarnAllocator的方法runAllocatedContainers

val executorHostname = container.getNodeId.getHost

传递过来的，也就是说我们最终获取到了yarn节点，也就是nodeManager的host
这样每个启动的executor，就向executor所在的nodeManager的YarnShuffleService注册了ExecutorShuffleInfo信息，这样对于开启了动态资源分配的
ExternalBlockStoreClient 来说fetchBlocksg过程就和未开启动态资源分配的NettyBlockTransferService大同小异了

spark on k8s(kubernetes) 中的DynamicResourceAllocation

参考之前的文章,我们知道在entrypoint中我们在启动executor的时候，我们传递了hostname参数

executor)
    shift 1
    CMD=(
      ${JAVA_HOME}/bin/java
      "${SPARK_EXECUTOR_JAVA_OPTS[@]}"
      -Xms$SPARK_EXECUTOR_MEMORY
      -Xmx$SPARK_EXECUTOR_MEMORY
      -cp "$SPARK_CLASSPATH:$SPARK_DIST_CLASSPATH"
      org.apache.spark.executor.CoarseGrainedExecutorBackend
      --driver-url $SPARK_DRIVER_URL
      --executor-id $SPARK_EXECUTOR_ID
      --cores $SPARK_EXECUTOR_CORES
      --app-id $SPARK_APPLICATION_ID
      --hostname $SPARK_EXECUTOR_POD_IP
    )

而SPARK_EXECUTOR_POD_IP是运行中的POD IP,参考BasicExecutorFeatureStep类片段：

Seq(new EnvVarBuilder()
          .withName(ENV_EXECUTOR_POD_IP)
          .withValueFrom(new EnvVarSourceBuilder()
            .withNewFieldRef("v1", "status.podIP")
            .build())
          .build())

这样按照以上流程的分析，
executor也不能向k8s节点ExternalShuffleService服务注册，因为我们注册的节点是POD IP,而不是节点IP，
当然spark社区早就提出了未开启external shuffle service的动态资源分配,且已经合并到master分支.
具体配置，可以参照如下:

spark.dynamicAllocation.enabled  true 
spark.dynamicAllocation.shuffleTracking.enabled  true
spark.dynamicAllocation.minExecutors  1
spark.dynamicAllocation.maxExecutors  4
spark.dynamicAllocation.executorIdleTimeout     60s

本文由博客群发一文多发等运营工具平台 OpenWrite 发布