DataStreamReader和DataStreamWriter等

流的讀取是從DataStreamReader和DataStreamWriter開始的。sql

DataStreamReader

DataStreamReader是生成流讀取者的入口所在,關鍵方法是load。這段代碼很關鍵,因此把所有代碼先貼出來,慢慢分析。apache

def load(): DataFrame = {
    
    val ds = DataSource.lookupDataSource(source, sparkSession.sqlContext.conf).

      getConstructor().newInstance()
   
    val v1DataSource = DataSource(

      sparkSession,

      userSpecifiedSchema = userSpecifiedSchema,

      className = source,

      options = extraOptions.toMap)

    val v1Relation = ds match {

      case _: StreamSourceProvider => Some(StreamingRelation(v1DataSource))

      case _ => None

    }

    ds match {

      case provider: TableProvider =>

        val sessionOptions = DataSourceV2Utils.extractSessionConfigs(

          source = provider, conf = sparkSession.sessionState.conf)

        val options = sessionOptions ++ extraOptions

        val dsOptions = new CaseInsensitiveStringMap(options.asJava)

        val table = userSpecifiedSchema match {

          case Some(schema) => provider.getTable(dsOptions, schema)

          case _ => provider.getTable(dsOptions)

        }

        import org.apache.spark.sql.execution.datasources.v2.DataSourceV2Implicits._

        table match {

          case _: SupportsRead if table.supportsAny(MICRO_BATCH_READ, CONTINUOUS_READ) =>

            Dataset.ofRows(

              sparkSession,

              StreamingRelationV2(

                provider, source, table, dsOptions, table.schema.toAttributes, v1Relation)(

                sparkSession))

          // fallback to v1

          // TODO (SPARK-27483): we should move this fallback logic to an analyzer rule.

          case _ => Dataset.ofRows(sparkSession, StreamingRelation(v1DataSource))

        }

      case _ =>

        // Code path for data source v1.

        Dataset.ofRows(sparkSession, StreamingRelation(v1DataSource))

    }

  }

有好多分支,重要的是區分開V1和V2。session

V1用的邏輯關係是StreamingRelation;而V2用的邏輯關係是StreamingRelationV2。這裏先看看他們對應的物理計劃是什麼?app

在SparkStrategies.scala文件中,定義了物理計劃:ide

/**

   * This strategy is just for explaining `Dataset/DataFrame` created by `spark.readStream`.

   * It won't affect the execution, because `StreamingRelation` will be replaced with

   * `StreamingExecutionRelation` in `StreamingQueryManager` and `StreamingExecutionRelation` will

   * be replaced with the real relation using the `Source` in `StreamExecution`.

   */

object StreamingRelationStrategy extends Strategy {

    def apply(plan: LogicalPlan): Seq[SparkPlan] = plan match {

      case s: StreamingRelation =>

        StreamingRelationExec(s.sourceName, s.output) :: Nil

      case s: StreamingExecutionRelation =>

        StreamingRelationExec(s.toString, s.output) :: Nil

      case s: StreamingRelationV2 =>

        StreamingRelationExec(s.sourceName, s.output) :: Nil

      case _ => Nil

    }

  }

物理計劃都是StreamingRelationExec,StreamingRelationExec的代碼其實啥都沒實現,因此最後其實看代碼註釋StreamingRelationExec也不是真正的物理計劃。函數

這裏先記得相關的類ContinuousExecution和MicroBatchExecution。一時找不到怎麼執行到具體的物理計劃ContinuousExecution和MicroBatchExecution的,咱們就試試反推把。先看看ContinuousExecution的代碼。oop

StreamExecution

StreamExecution是抽象類。其抽象方法runActivatedStream是執行具體的連續流讀取任務的,子類會重寫該函數。post

runStream方法封裝了runActivatedStream方法,額外加了些事件通知等處理機制,知道這一點就好了。this

StreamingQueryManager

這裏先嚐試看看StreamingQueryManager是幹什麼用的,看註釋應該是管理全部的StreamingQuery的。spa

 private def createQuery(...): StreamingQueryWrapper ={

   (sink, trigger) match {

      case (table: SupportsWrite, trigger: ContinuousTrigger) =>
       

        new StreamingQueryWrapper(new ContinuousExecution(

          sparkSession,

          userSpecifiedName.orNull,

          checkpointLocation,

          analyzedPlan,

          table,

          trigger,

          triggerClock,

          outputMode,

          extraOptions,

          deleteCheckpointOnStop))

      case _ =>

        if (operationCheckEnabled) {

          UnsupportedOperationChecker.checkForStreaming(analyzedPlan, outputMode)

        }

        new StreamingQueryWrapper(new MicroBatchExecution(

          sparkSession,

          userSpecifiedName.orNull,

          checkpointLocation,

          analyzedPlan,

          sink,

          trigger,

          triggerClock,

          outputMode,

          extraOptions,

          deleteCheckpointOnStop))

    }

}

對於連續流,返回一個:

  new StreamingQueryWrapper(new ContinuousExecution))

StreamingQueryWrapper的做用,就是將StreamingQuery封裝成可序列化的,別的和StreamingQuery沒什麼區別。這裏對於連續流就是包裝了ContinuousExecution。

ContinuousExecution

ContinuousExecution看名稱應該是對應連續流的物理執行計劃的,繼承自StreamExecution(抽象類)。看看主要代碼其實就是重寫了runActivatedStream方法。

 override protected def runActivatedStream(sparkSessionForStream: SparkSession): Unit = {

    val stateUpdate = new UnaryOperator[State] {

      override def apply(s: State) = s match {

        // If we ended the query to reconfigure, reset the state to active.

        case RECONFIGURING => ACTIVE

        case _ => s

      }

    }

    do {

      runContinuous(sparkSessionForStream)

    } while (state.updateAndGet(stateUpdate) == ACTIVE)

    stopSources()

  }

真正的執行邏輯代碼在私有方法runContinuous中,這裏就不詳細展開了,知道了主要流程就能夠了。

下面就是要看看ContinuousExecution究竟是在哪裏被從邏輯計劃轉換到物理計劃的。

搜索全文,找到了StreamingQueryManager.scala這個文件。對了,就是從上面的StreamingQueryManager找到這個ContinuousExecution。

DataStreamWriter

DataStreamWriter是真正觸發流計算開始啓動執行的地方。

start()方法獲得要給StreamingQuery,方法裏的關鍵代碼片斷:

 df.sparkSession.sessionState.streamingQueryManager.startQuery(

        extraOptions.get("queryName"),

        extraOptions.get("checkpointLocation"),

        df,

        extraOptions.toMap,

        sink,

        outputMode,

        useTempCheckpointLocation = source == "console" || source == "noop",

        recoverFromCheckpointLocation = true,

        trigger = trigger)

跟蹤進去到了StreamingQueryManager,看它的startQuery方法。

startQuery方法分爲幾步:

  1. 調用createQuery方法返回StreamingQuery。

val query = createQuery(

      userSpecifiedName,

      userSpecifiedCheckpointLocation,

      df,

      extraOptions,

      sink,

      outputMode,

      useTempCheckpointLocation,

      recoverFromCheckpointLocation,

      trigger,

      triggerClock)

query就是StreamingQueryWrapper,就是相似這樣的代碼:

new StreamingQueryWrapper(new ContinuousExecution))

二、啓動上一步的query 

try {     

      query.streamingQuery.start()

    } catch {     

    }

這裏的代碼直接調用到StreamingQuery的父類StreamExecution的start方法。代碼定義:

def start(): Unit = {

    logInfo(s"Starting $prettyIdString. Use $resolvedCheckpointRoot to store the query checkpoint.")

    queryExecutionThread.setDaemon(true)

    queryExecutionThread.start()

    startLatch.await()  // Wait until thread started and QueryStart event has been posted

  }

queryExecutionThread線程的定義又是這樣的:

val queryExecutionThread: QueryExecutionThread =

    new QueryExecutionThread(s"stream execution thread for $prettyIdString") {

      override def run(): Unit = {

        sparkSession.sparkContext.setCallSite(callSite)

        runStream()

      }

    }

最後在線程中啓動runStream這個私有方法。

三、返回query

最後返回query,注意這裏的query在上面的代碼中已經start運行了。

相關文章
相關標籤/搜索