【原创】大数据基础之Spark（3）Spark Thrift实现原理及代码实现

spark 2.1.1

一 启动命令

启动spark thrift命令

$SPARK_HOME/sbin/start-thriftserver.sh

然后会执行

org.apache.spark.deploy.SparkSubmit --class org.apache.spark.sql.hive.thriftserver.HiveThriftServer2

二 启动过程及代码分析

hive thrift代码详见：https://www.cnblogs.com/barneywill/p/10185168.html

HiveThriftServer2是spark thrift核心类，继承自Hive的HiveServer2

org.apache.spark.sql.hive.thriftserver.HiveThriftServer2 extends org.apache.hive.service.server.HiveServer2

启动过程：

HiveThriftServer2.main

SparkSQLEnv.init (sparkConf sparkSession sparkContext sqlContext)

HiveThriftServer2.init

addService(ThriftBinaryCLIService)

HiveThriftServer2.start

ThriftBinaryCLIService.run

TServer.serve

类结构：【接口或父类->子类】

TServer->TThreadPoolServer

TProcessorFactory->SQLPlainProcessorFactory

TProcessor->TSetIpAddressProcessor

ThriftCLIService->ThriftBinaryCLIService

CLIService->SparkSQLCLIService (核心子类)

服务初始化过程：

CLIService.init

SparkSQLCLIService.init

addService(SparkSQLSessionManager)

initCompositeService

SparkSQLSessionManager.init

addService(SparkSQLOperationManager)

initCompositeService

SparkSQLOperationManager.init

三 DDL执行过程

ddl执行过程需要和hive metastore交互

从执行计划开始：

spark-sql> explain create table test_table(id string);
== Physical Plan ==
ExecutedCommand
+- CreateTableCommand CatalogTable(
Table: `test_table`
Created: Wed Dec 19 18:04:15 CST 2018
Last Access: Thu Jan 01 07:59:59 CST 1970
Type: MANAGED
Schema: [StructField(id,StringType,true)]
Provider: hive
Storage(InputFormat: org.apache.hadoop.mapred.TextInputFormat, OutputFormat: org.apache.hadoop.hive.ql.io.HiveIgnoreKeyTextOutputFormat)), false
Time taken: 0.28 seconds, Fetched 1 row(s)

从执行计划里可以找到具体的Command，这里是CreateTableCommand

org.apache.spark.sql.execution.command.tables

case class CreateTableCommand(table: CatalogTable, ifNotExists: Boolean) extends RunnableCommand {

  override def run(sparkSession: SparkSession): Seq[Row] = {
    sparkSession.sessionState.catalog.createTable(table, ifNotExists)
    Seq.empty[Row]
  }
}

这里可以看到是直接将请求分发给sparkSession.sessionState.catalog

org.apache.spark.sql.internal.SessionState

  /**
   * Internal catalog for managing table and database states.
   */
  lazy val catalog = new SessionCatalog(
    sparkSession.sharedState.externalCatalog,
    sparkSession.sharedState.globalTempViewManager,
    functionResourceLoader,
    functionRegistry,
    conf,
    newHadoopConf())

取的是sparkSession.sharedState.externalCatalog

org.apache.spark.sql.internal.SharedState

  /**
   * A catalog that interacts with external systems.
   */
  val externalCatalog: ExternalCatalog =
    SharedState.reflect[ExternalCatalog, SparkConf, Configuration](
      SharedState.externalCatalogClassName(sparkContext.conf),
      sparkContext.conf,
      sparkContext.hadoopConfiguration)
...
  private val HIVE_EXTERNAL_CATALOG_CLASS_NAME = "org.apache.spark.sql.hive.HiveExternalCatalog"

  private def externalCatalogClassName(conf: SparkConf): String = {
    conf.get(CATALOG_IMPLEMENTATION) match {
      case "hive" => HIVE_EXTERNAL_CATALOG_CLASS_NAME
      case "in-memory" => classOf[InMemoryCatalog].getCanonicalName
    }
  }

这里可以看到是通过externalCatalogClassName反射实例化的，代码里硬编码使用的是org.apache.spark.sql.hive.HiveExternalCatalog

org.apache.spark.sql.hive.HiveExternalCatalog

  /**
   * A Hive client used to interact with the metastore.
   */
  val client: HiveClient = {
    HiveUtils.newClientForMetadata(conf, hadoopConf)
  }

  private def withClient[T](body: => T): T = synchronized {
    try {
      body
    } catch {
      case NonFatal(exception) if isClientException(exception) =>
        val e = exception match {
          // Since we are using shim, the exceptions thrown by the underlying method of
          // Method.invoke() are wrapped by InvocationTargetException
          case i: InvocationTargetException => i.getCause
          case o => o
        }
        throw new AnalysisException(
          e.getClass.getCanonicalName + ": " + e.getMessage, cause = Some(e))
    }
  }

  override def createDatabase(
      dbDefinition: CatalogDatabase,
      ignoreIfExists: Boolean): Unit = withClient {
    client.createDatabase(dbDefinition, ignoreIfExists)
  }

这个类里执行任何ddl方法都会执行withClient，而withClient有synchronized，执行过程是直接把请求分发给client，下面看client是什么

org.apache.spark.sql.hive.client.IsolatedClientLoader

  /** The isolated client interface to Hive. */
  private[hive] def createClient(): HiveClient = {
    if (!isolationOn) {
      return new HiveClientImpl(version, sparkConf, hadoopConf, config, baseClassLoader, this)
    }
    // Pre-reflective instantiation setup.
    logDebug("Initializing the logger to avoid disaster...")
    val origLoader = Thread.currentThread().getContextClassLoader
    Thread.currentThread.setContextClassLoader(classLoader)

    try {
      classLoader
        .loadClass(classOf[HiveClientImpl].getName)
        .getConstructors.head
        .newInstance(version, sparkConf, hadoopConf, config, classLoader, this)
        .asInstanceOf[HiveClient]
    } catch {

可见client直接用的是org.apache.spark.sql.hive.client.HiveClientImpl

org.apache.spark.sql.hive.client.HiveClientImpl

  def withHiveState[A](f: => A): A = retryLocked {
    val original = Thread.currentThread().getContextClassLoader
    // Set the thread local metastore client to the client associated with this HiveClientImpl.
    Hive.set(client)
    // The classloader in clientLoader could be changed after addJar, always use the latest
    // classloader
    state.getConf.setClassLoader(clientLoader.classLoader)
    // setCurrentSessionState will use the classLoader associated
    // with the HiveConf in `state` to override the context class loader of the current
    // thread.
    shim.setCurrentSessionState(state)
    val ret = try f finally {
      Thread.currentThread().setContextClassLoader(original)
      HiveCatalogMetrics.incrementHiveClientCalls(1)
    }
    ret
  }
  private def retryLocked[A](f: => A): A = clientLoader.synchronized {
...

  override def createDatabase(
      database: CatalogDatabase,
      ignoreIfExists: Boolean): Unit = withHiveState {
    client.createDatabase(
      new HiveDatabase(
        database.name,
        database.description,
        database.locationUri,
        Option(database.properties).map(_.asJava).orNull),
        ignoreIfExists)
  }

这个类执行任何ddl方法都会执行withHiveState，withHiveState会执行retryLocked，retryLocked上有synchronized；而且这里也是直接将请求分发给client，这里的client是hive的类org.apache.hadoop.hive.ql.metadata.Hive

四 DML执行过程

dml执行过程最后会执行到spark.sql

sql执行过程：

CLIService.executeStatement （返回OperationHandle）

SessionManager.getSession

SessionManager.openSession

SparkSQLSessionManager.openSession

SparkSQLOperationManager.sessionToContexts.set （openSession时：session和sqlContext建立映射）

HiveSession.executeStatement

HiveSessionImpl.executeStatementInternal

OperationManager.newExecuteStatementOperation

SparkSQLOperationManager.newExecuteStatementOperation

SparkSQLOperationManager.sessionToContexts.get （通过session取到sqlContext）

ExecuteStatementOperation.run

SparkExecuteStatementOperation.run

SparkExecuteStatementOperation.execute

SQLContext.sql （熟悉的spark sql）

可见从SparkSQLCLIService初始化开始，逐个将各个类的实现类改为spark的子类比如：

org.apache.spark.sql.hive.thriftserver.SparkSQLSessionManager extends org.apache.hive.service.cli.session.SessionManager
org.apache.spark.sql.hive.thriftserver.server.SparkSQLOperationManager extends org.apache.hive.service.cli.operation.OperationManager
org.apache.spark.sql.hive.thriftserver.SparkExecuteStatementOperation extends org.apache.hive.service.cli.operation.ExecuteStatementOperation

从而实现底层实现的替换；

hive的HiveServer2为什么这么容易的被扩展，详见spark代码的sql/hive-thriftserver，这里应该是将hive1.2代码做了很多修改，以后升级就不那么容易了；
至于spark为什么要花这么大力气扩展HiveServer2而不是重新实现，可能是为了保持接口一致，这样有利于原来使用hive thrift的用户平滑的迁移到spark thrift，因为唯一的改动就是切换url，实际上，相同sql下的spark thrift和hive thrift表现还是有很多不同的。