这是一个简单的代码示例来说明我的问题:

case class Record( key: String, value: Int )

object Job extends App
{
  val env = StreamExecutionEnvironment.getExecutionEnvironment
  val data = env.fromElements( Record("01",1), Record("02",2), Record("03",3), Record("04",4), Record("05",5) )
  val step1 = data.filter( record => record.value % 3 != 0  ) // introduces some data loss
  val step2 = data.map( r => Record( r.key, r.value * 2 ) )
  val step3 = data.map( r => Record( r.key, r.value * 3 ) )
  val merged = step1.union( step2, step3 )
  val keyed = merged.keyBy(0)
  val windowed = keyed.countWindow( 3 )
  val summed = windowed.sum( 1 )
  summed.print()
  env.execute("test")
}

这将产生以下结果:
Record(01,6)
Record(02,12)
Record(04,24)
Record(05,30)

正如预期的那样,键“03”不会产生任何结果,因为计数窗口需要3个元素,并且流中仅存在2个元素。

我想要的是一种带有超时的计数窗口,以便在一定的超时后,如果未达到计数窗口期望的元素数量,则现有元素会产生部分结果。

有了这种行为,在我的示例中,当达到超时时,将生成Record(03,15)。

最佳答案

我已经遵循了David和NIrav的方法,这是结果。

1)使用自定义触发器:

在这里,我颠倒了最初的逻辑。我没有使用“时间窗口”,而是使用了一个与超时相对应的持续时间的“时间窗口”,然后是在处理完所有元素后触发的触发器。

case class Record( key: String, value: Int )

object Job extends App
{
  val env = StreamExecutionEnvironment.getExecutionEnvironment
  val data = env.fromElements( Record("01",1), Record("02",2), Record("03",3), Record("04",4), Record("05",5) )
  val step1 = data.filter( record => record.value % 3 != 0  ) // introduces some data loss
  val step2 = data.map( r => Record( r.key, r.value * 2 ) )
  val step3 = data.map( r => Record( r.key, r.value * 3 ) )
  val merged = step1.union( step2, step3 )
  val keyed = merged.keyBy(0)
  val windowed = keyed.timeWindow( Time.milliseconds( 50 ) )
  val triggered = windowed.trigger( new CountTriggerWithTimeout( 3, env.getStreamTimeCharacteristic ) )
  val summed = triggered.sum( 1 )
  summed.print()
  env.execute("test")
}

这是触发代码:
import org.apache.flink.annotation.PublicEvolving
import org.apache.flink.api.common.functions.ReduceFunction
import org.apache.flink.api.common.functions.RuntimeContext
import org.apache.flink.api.common.state.ReducingState
import org.apache.flink.api.common.state.ReducingStateDescriptor
import org.apache.flink.api.common.typeutils.base.LongSerializer
import org.apache.flink.streaming.api.TimeCharacteristic
import org.apache.flink.streaming.api.windowing.triggers._
import org.apache.flink.streaming.api.windowing.triggers.Trigger.TriggerContext
import org.apache.flink.streaming.api.windowing.windows.TimeWindow

/**
 * A trigger that fires when the count of elements in a pane reaches the given count or a
 * timeout is reached whatever happens first.
 */
class CountTriggerWithTimeout[W <: TimeWindow](maxCount: Long, timeCharacteristic: TimeCharacteristic) extends Trigger[Object,W]
{
  private val countState: ReducingStateDescriptor[java.lang.Long] = new ReducingStateDescriptor[java.lang.Long]( "count", new Sum(), LongSerializer.INSTANCE)

  override def onElement(element: Object, timestamp: Long, window: W, ctx: TriggerContext): TriggerResult =
  {
      val count: ReducingState[java.lang.Long] = ctx.getPartitionedState(countState)
      count.add( 1L )
      if ( count.get >= maxCount || timestamp >= window.getEnd ) TriggerResult.FIRE_AND_PURGE else TriggerResult.CONTINUE
  }

  override def onProcessingTime(time: Long, window: W, ctx: TriggerContext): TriggerResult =
  {
      if (timeCharacteristic == TimeCharacteristic.EventTime) TriggerResult.CONTINUE else
      {
          if ( time >= window.getEnd ) TriggerResult.CONTINUE else TriggerResult.FIRE_AND_PURGE
      }
  }

  override def onEventTime(time: Long, window: W, ctx: TriggerContext): TriggerResult =
  {
      if (timeCharacteristic == TimeCharacteristic.ProcessingTime) TriggerResult.CONTINUE else
      {
          if ( time >= window.getEnd ) TriggerResult.CONTINUE else TriggerResult.FIRE_AND_PURGE
      }
  }

  override def clear(window: W, ctx: TriggerContext): Unit =
  {
          ctx.getPartitionedState( countState ).clear
    }

    class Sum extends ReduceFunction[java.lang.Long]
  {
        def reduce(value1: java.lang.Long, value2: java.lang.Long): java.lang.Long = value1 + value2
  }
}

2)使用过程功能:
case class Record( key: String, value: Int )

object Job extends App
{
  val env = StreamExecutionEnvironment.getExecutionEnvironment
  env.setStreamTimeCharacteristic( TimeCharacteristic.IngestionTime )
  val data = env.fromElements( Record("01",1), Record("02",2), Record("03",3), Record("04",4), Record("05",5) )
  val step1 = data.filter( record => record.value % 3 != 0  ) // introduces some data loss
  val step2 = data.map( r => Record( r.key, r.value * 2 ) )
  val step3 = data.map( r => Record( r.key, r.value * 3 ) )
  val merged = step1.union( step2, step3 )
  val keyed = merged.keyBy(0)
  val processed = keyed.process( new TimeCountWindowProcessFunction( 3, 100 ) )
  processed.print()
  env.execute("test")
}

在所有逻辑(即窗口化,触发和求和)进入函数后:
import org.apache.flink.streaming.api.functions._
import org.apache.flink.util._
import org.apache.flink.api.common.state._

case class Status( count: Int, key: String, value: Long )

class TimeCountWindowProcessFunction( count: Long, windowSize: Long ) extends ProcessFunction[Record,Record]
{
    lazy val state: ValueState[Status] = getRuntimeContext
      .getState(new ValueStateDescriptor[Status]("state", classOf[Status]))

    override def processElement( input: Record, ctx: ProcessFunction[Record,Record]#Context, out: Collector[Record] ): Unit =
    {
        val updated: Status = Option( state.value ) match {
            case None => {
                ctx.timerService().registerEventTimeTimer( ctx.timestamp + windowSize )
                Status( 1, input.key, input.value )
            }
            case Some( current ) => Status( current.count + 1, input.key, input.value + current.value )
        }
        if ( updated.count == count )
        {
            out.collect( Record( input.key, updated.value ) )
            state.clear
        }
        else
        {
            state.update( updated )
        }
    }

    override def onTimer( timestamp: Long, ctx: ProcessFunction[Record,Record]#OnTimerContext, out: Collector[Record] ): Unit =
    {
        Option( state.value ) match {
            case None => // ignore
            case Some( status ) => {
                out.collect( Record( status.key, status.value ) )
                state.clear
            }
        }
    }
}

10-07 19:00
查看更多