Flink之聚合、WindowFunction

一、WindowFunction

Flink提供了四种类型Window Function，其中有ReduceFunction、AggregateFunction、FlodFunction和ProcessWindowFunction。其中ReduceFunction、AggregateFunction、FlodFunction根据计算原理，属于增量聚合函数，而ProcessWindowFunction属于全量聚合函数。增量聚合函数是基于中间状态计算结果的，窗口中只维护中间状态结果值，不需要缓存原始的数据，而全量窗口函数在窗口触发时对所有的原始数据进行汇总计算，因此相对性能会较差。

• ReduceFunction 增量聚合
• AggregateFunction 增量聚合
• FlodFunction 已废弃，推荐使用AggregateFunction
• ProcessWindowFunction 全量聚合

二、ReduceFunction

x以前的聚合结果，y下一个数据，输入输出都为同类型数据

.reduce(new ReduceFunction[Sensor] {
      override def reduce(x: Sensor, y: Sensor): Sensor = {
        Sensor(x.id, x.timestamp + 1, y.temperature + 1)
      }

三、AggregateFunction

聚合函数只传AggregateFunction参数，如果局和后海域需要的操作可以使用AggregateFunction，WindowFunction(如果需要获取上下文或操作State等可以使用RichWindowFuncton)

    val pvStream = dataStream.filter(_.behavior == "pv")
      .keyBy(_.itemId)
      .timeWindow(Time.hours(1), Time.minutes(5))
      //.aggregate(new CountAgg())
      .aggregate(new CountAgg(), new WindowResult())
      .keyBy(_.windowEnd)
      .process(new TopHotItems(3))

// IN输入, ACC中间累加值(可以定义为任意类型), OUT输出给window函数的
class CountAgg extends AggregateFunction[UserBehavior, Long, Long] {
  //初始值
  override def createAccumulator(): Long = 0L
  //对重分区数据merge处理
  override def merge(acc: Long, acc1: Long): Long = acc + acc1
  //返回值
  override def getResult(acc: Long): Long = acc
  //新来的一条数据如何处理
  override def add(in: UserBehavior, acc: Long): Long = acc + 1
}

class WindowResult extends WindowFunction[Long, ItemViewCount, Long, TimeWindow] {
  override def apply(key: Long, window: TimeWindow, input: Iterable[Long], out: Collector[ItemViewCount]): Unit = {
    out.collect(ItemViewCount(key, window.getEnd, input.iterator.next()))
  }
}

四、ProcessWindowFunction

示例代码，仅供参考

val pvStream = dataStream.filter(_.behavior == "pv")
      .keyBy(_.itemId)
      .timeWindow(Time.hours(1), Time.minutes(5))
      .process(new ProcessWindowFunction[UserBehavior, Long, Long, TimeWindow] {
        lazy val count: ValueState[Long] = getRuntimeContext.getState(new ValueStateDescriptor("count", classOf[Long]))
        @throws[Exception]
        override def process(key: Long, context: Context, elements: Iterable[UserBehavior], out: Collector[Long]): Unit = {
          val preCount = count.value()
          count.update(preCount + 1)
          out.collect(count.value())
        }
        override def close(): Unit = count.clear()
      })