[转载] Spark Structed Streaming执行过程

case class ProcessingTimeExecutor(processingTime: ProcessingTime, clock: Clock = new SystemClock())
  extends TriggerExecutor with Logging {

  private val intervalMs = processingTime.intervalMs
  require(intervalMs >= 0)

  override def execute(triggerHandler: () => Boolean): Unit = {
    while (true) {
      val triggerTimeMs = clock.getTimeMillis
      val nextTriggerTimeMs = nextBatchTime(triggerTimeMs)
      val terminated = !triggerHandler()
      if (intervalMs > 0) {
        val batchElapsedTimeMs = clock.getTimeMillis - triggerTimeMs
        if (batchElapsedTimeMs > intervalMs) {
          notifyBatchFallingBehind(batchElapsedTimeMs)
        }
        if (terminated) {
          return
        }
        clock.waitTillTime(nextTriggerTimeMs)
      } else {
        if (terminated) {
          return
        }
      }
    }
  }

override lazy val logicalPlan: LogicalPlan = {
    assert(queryExecutionThread eq Thread.currentThread,
      "logicalPlan must be initialized in QueryExecutionThread " +
        s"but the current thread was ${Thread.currentThread}")
    var nextSourceId = 0L
    val toExecutionRelationMap = MutableMap[StreamingRelation, StreamingExecutionRelation]()
    val v2ToExecutionRelationMap = MutableMap[StreamingRelationV2, StreamingExecutionRelation]()
    // We transform each distinct streaming relation into a StreamingExecutionRelation, keeping a
    // map as we go to ensure each identical relation gets the same StreamingExecutionRelation
    // object. For each microbatch, the StreamingExecutionRelation will be replaced with a logical
    // plan for the data within that batch.
    // Note that we have to use the previous `output` as attributes in StreamingExecutionRelation,
    // since the existing logical plan has already used those attributes. The per-microbatch
    // transformation is responsible for replacing attributes with their final values.
    val _logicalPlan = analyzedPlan.transform {
      case streamingRelation@StreamingRelation(dataSource, _, output) =>
        toExecutionRelationMap.getOrElseUpdate(streamingRelation, {
          // Materialize source to avoid creating it in every batch
          val metadataPath = s"$resolvedCheckpointRoot/sources/$nextSourceId"
          val source = dataSource.createSource(metadataPath)
          nextSourceId += 1
          StreamingExecutionRelation(source, output)(sparkSession)
        })
      case s@StreamingRelationV2(source: MicroBatchReadSupport, _, options, output, _) =>
        v2ToExecutionRelationMap.getOrElseUpdate(s, {
          // Materialize source to avoid creating it in every batch
          val metadataPath = s"$resolvedCheckpointRoot/sources/$nextSourceId"
          val reader = source.createMicroBatchReader(
            Optional.empty(), // user specified schema
            metadataPath,
            new DataSourceOptions(options.asJava))
          nextSourceId += 1
          StreamingExecutionRelation(reader, output)(sparkSession)
        })
      case s@StreamingRelationV2(_, sourceName, _, output, v1Relation) =>
        v2ToExecutionRelationMap.getOrElseUpdate(s, {
          // Materialize source to avoid creating it in every batch
          val metadataPath = s"$resolvedCheckpointRoot/sources/$nextSourceId"
          if (v1Relation.isEmpty) {
            throw new UnsupportedOperationException(
              s"Data source $sourceName does not support microbatch processing.")
          }
          val source = v1Relation.get.dataSource.createSource(metadataPath)
          nextSourceId += 1
          StreamingExecutionRelation(source, output)(sparkSession)
        })
    }
    sources = _logicalPlan.collect { case s: StreamingExecutionRelation => s.source }
    uniqueSources = sources.distinct
    _logicalPlan
  }

newData = reportTimeTaken("getBatch") {
      availableOffsets.flatMap {
        case (source: Source, available)
          if committedOffsets.get(source).map(_ != available).getOrElse(true) =>
          val current = committedOffsets.get(source)
          //这部分逻辑基于传入的起始offset范围（包含了每个partition的offset范围）形成一个kafka的DataFrame
          val batch = source.getBatch(current, available)

val newBatchesPlan = logicalPlan transform {
      case StreamingExecutionRelation(source, output) =>
        newData.get(source).map { dataPlan =>
          assert(output.size == dataPlan.output.size,
            s"Invalid batch: ${Utils.truncatedString(output, ",")} != " +
              s"${Utils.truncatedString(dataPlan.output, ",")}")
          replacements ++= output.zip(dataPlan.output)
          dataPlan
        }.getOrElse {
          LocalRelation(output, isStreaming = true)
        }
    }

reportTimeTaken("queryPlanning") {
      lastExecution = new IncrementalExecution(
        sparkSessionToRunBatch,
        triggerLogicalPlan,
        outputMode,
        checkpointFile("state"),
        runId,
        currentBatchId,
        offsetSeqMetadata)
      lastExecution.executedPlan // Force the lazy generation of execution plan
    }

    val nextBatch =
      new Dataset(sparkSessionToRunBatch, lastExecution, RowEncoder(lastExecution.analyzed.schema))

接下来基于不同的sink进行处理，其中SQLExecution.withNewExecutionId主要是为了跟踪jobs的信息

reportTimeTaken("addBatch") {
      SQLExecution.withNewExecutionId(sparkSessionToRunBatch, lastExecution) {
        sink match {
          case s: Sink =>
            if (s.isInstanceOf[MemorySinkExtend]) {
              s.addBatch(currentBatchId, nextBatch, batchIdOffsetMap.get(currentBatchId).getOrElse((None, None)))
            } else {
              s.addBatch(currentBatchId, nextBatch, (None, None))
            }
          case _: StreamWriteSupport =>
            // This doesn't accumulate any data - it just forces execution of the microbatch writer.
            nextBatch.collect()
        }
      }
    }

if (hasNewData) {
      var batchWatermarkMs = offsetSeqMetadata.batchWatermarkMs
      // Update the eventTime watermarks if we find any in the plan.
      if (lastExecution != null) {
        lastExecution.executedPlan.collect {
          case e: EventTimeWatermarkExec => e
        }.zipWithIndex.foreach {
          case (e, index) if e.eventTimeStats.value.count > 0 =>
            logDebug(s"Observed event time stats $index: ${e.eventTimeStats.value}")
            val newWatermarkMs = e.eventTimeStats.value.max - e.delayMs
            val prevWatermarkMs = watermarkMsMap.get(index)
            if (prevWatermarkMs.isEmpty || newWatermarkMs > prevWatermarkMs.get) {
              watermarkMsMap.put(index, newWatermarkMs)
            }

lazy val watermarkExpression: Option[Expression] = {
    WatermarkSupport.watermarkExpression(
      child.output.find(_.metadata.contains(EventTimeWatermark.delayKey)),
      eventTimeWatermark)
  }

  /** Predicate based on keys that matches data older than the watermark */
  lazy val watermarkPredicateForKeys: Option[Predicate] = watermarkExpression.flatMap { e =>
    if (keyExpressions.exists(_.metadata.contains(EventTimeWatermark.delayKey))) {
      Some(newPredicate(e, keyExpressions))
    } else {
      None
    }
  }

  /** Predicate based on the child output that matches data older than the watermark. */
  lazy val watermarkPredicateForData: Option[Predicate] =
    watermarkExpression.map(newPredicate(_, child.output))

ContinusExecution

该执行逻辑与上面类似，只是这部分在保存offset信息是异步方式，流中的数据一直在处理。