Flink统计当日的UV、PV

　　Flink 统计当日的UV、PV

　　测试环境：

　　　　flink 1.7.2

　　1、数据流程

　　　　

　　　　a.模拟数据生成，发送到kafka（json 格式）　　

　　　　b.flink 读取数据，count

　　　　c. 输出数据到kafka（为了方便查看，输出了一份到控制台）

　　2、模拟数据生成器

　　　　数据格式如下 ： {"id" : 1, "createTime" : "2019-05-24 10:36:43.707"}

　　　　id 为数据生成的序号（累加），时间为数据时间（默认为数据生成时间）

　　模拟数据生成器代码如下：

　　

/**
  * test data maker
  */

object CurrentDayMaker {


  var minute : Int = 1
  val calendar: Calendar = Calendar.getInstance()

  /**
    * 一天时间比较长，不方便观察，将时间改为当前时间，
    * 每次累加10分钟，这样一天只需要144次循环，也就是144秒
    * @return
    */
  def getCreateTime(): String = {
//    minute = minute + 1
    calendar.add(Calendar.MINUTE, 10)
    sdf.format(calendar.getTime)
  }
  val sdf = new SimpleDateFormat("yyyy-MM-dd HH:mm:ss.SSS")

  def main(args: Array[String]): Unit = {
    val producer = new KafkaProducer[String, String](Common.getProp)
　　// 初始化开始时间为当前时间
    calendar.setTime(new Date())
    println(sdf.format(calendar.getTime))
    var i =0;
    while (true) {

//      val map = Map("id"-> i, "createTime"-> sdf.format(System.currentTimeMillis()))
      val map = Map("id"-> i, "createTime"-> getCreateTime())
      val jsonObject: JSONObject = new JSONObject(map)
      println(jsonObject.toString())
　　　　// topic current_day
      val msg = new ProducerRecord[String, String]("current_day", jsonObject.toString())
      producer.send(msg)
      producer.flush()
　　　　// 控制数据频率
      Thread.sleep(1000)
      i = i + 1
    }
  }

}

　　生成数据如下：　　

{"id" : 0, "createTime" : "2019-05-24 18:02:26.292"}
{"id" : 1, "createTime" : "2019-05-24 18:12:26.292"}
{"id" : 2, "createTime" : "2019-05-24 18:22:26.292"}
{"id" : 3, "createTime" : "2019-05-24 18:32:26.292"}
{"id" : 4, "createTime" : "2019-05-24 18:42:26.292"}

　　3、flink 程序　

　　

package com.venn.stream.api.dayWindow

import java.io.File
import java.text.SimpleDateFormat

import com.venn.common.Common
import com.venn.source.TumblingEventTimeWindows
import org.apache.flink.api.common.functions.ReduceFunction
import org.apache.flink.api.common.serialization.SimpleStringSchema
import org.apache.flink.api.scala._
import org.apache.flink.contrib.streaming.state.RocksDBStateBackend
import org.apache.flink.formats.json.JsonNodeDeserializationSchema
import org.apache.flink.shaded.jackson2.com.fasterxml.jackson.databind.node.ObjectNode
import org.apache.flink.streaming.api.TimeCharacteristic
import org.apache.flink.streaming.api.functions.timestamps.BoundedOutOfOrdernessTimestampExtractor
import org.apache.flink.streaming.api.scala.StreamExecutionEnvironment
import org.apache.flink.streaming.api.windowing.time.Time
import org.apache.flink.streaming.api.windowing.triggers.{ContinuousEventTimeTrigger}
import org.apache.flink.streaming.connectors.kafka.{FlinkKafkaConsumer, FlinkKafkaProducer}

/**
  * Created by venn on 19-5-23.
  *
  * use TumblingEventTimeWindows count current day pv
  * for test, update day window to minute window
  *
  * .windowAll(TumblingEventTimeWindows.of(Time.minutes(1), Time.seconds(0)))
  * TumblingEventTimeWindows can ensure count o minute event,
  * and time start at 0 second (like : 00:00:00 to 00:00:59)
  *
  */
object CurrentDayPvCount {

  def main(args: Array[String]): Unit = {
    println(1558886400000L - (1558886400000L - 8 + 86400000) % 86400000)
    // environment
    val env: StreamExecutionEnvironment = StreamExecutionEnvironment.getExecutionEnvironment
    env.setStreamTimeCharacteristic(TimeCharacteristic.EventTime)
    env.setParallelism(1)
    if ("\\".equals(File.pathSeparator)) {
      val rock = new RocksDBStateBackend(Common.CHECK_POINT_DATA_DIR)
      env.setStateBackend(rock)
      // checkpoint interval
      env.enableCheckpointing(10000)
    }

    val topic = "current_day"
    val sdf = new SimpleDateFormat("yyyy-MM-dd HH:mm:ss.SSS")
    val kafkaSource = new FlinkKafkaConsumer[ObjectNode](topic, new JsonNodeDeserializationSchema(), Common.getProp)
    val sink = new FlinkKafkaProducer[String](topic + "_out", new SimpleStringSchema(), Common.getProp)
    sink.setWriteTimestampToKafka(true)

    val stream = env.addSource(kafkaSource)
      .map(node => {
        Event(node.get("id").asText(), node.get("createTime").asText())
      })
      //            .assignAscendingTimestamps(event => sdf.parse(event.createTime).getTime)
      .assignTimestampsAndWatermarks(new BoundedOutOfOrdernessTimestampExtractor[Event](Time.seconds(60)) {
      override def extractTimestamp(element: Event): Long = {
        sdf.parse(element.createTime).getTime
      }
    })
      // window is one minute, start at 0 second
      //.windowAll(TumblingEventTimeWindows.of(Time.minutes(1), Time.seconds(0)))
      // window is one hour, start at 0 second 注意事件时间，需要事件触发，在窗口结束的时候可能没有数据，有数据的时候，已经是下一个窗口了
      //      .windowAll(TumblingEventTimeWindows.of(Time.hours(1), Time.seconds(0)))
      // window is one day, start at 0 second, todo there have a bug(FLINK-11326), can't use negative number, 1.8 修复
      //      .windowAll(TumblingEventTimeWindows.of(Time.days(1)))
      .windowAll(TumblingEventTimeWindows.of(Time.days(1), Time.hours(-8)))
      // every event one minute
      //      .trigger(ContinuousEventTimeTrigger.of(Time.seconds(3800)))
      // every process one minute
      //      .trigger(ContinuousProcessingTimeTrigger.of(Time.seconds(10)))
      // every event, export current value,
      //      .trigger(CountTrigger.of(1))
      .reduce(new ReduceFunction[Event] {
      override def reduce(event1: Event, event2: Event): Event = {

        // 将结果中，id的最小值和最大值输出
        new Event(event1.id, event2.id, event1.count + event2.count)
      }
    })
    // format output even, connect min max id, add current timestamp
    //      .map(event => Event(event.id + "-" + event.createTime, sdf.format(System.currentTimeMillis()), event.count))
    stream.print("result : ")
    // execute job
    env.execute("CurrentDayCount")
  }

}

case class Event(id: String, createTime: String, count: Int = 1) {}

 

　　4、运行结果

　　测试数据如下：　　　　

{"id" : 0, "createTime" : "2019-05-24 20:29:49.102"}
{"id" : 1, "createTime" : "2019-05-24 20:39:49.102"}
...
{"id" : 20, "createTime" : "2019-05-24 23:49:49.102"}
{"id" : 21, "createTime" : "2019-05-24 23:59:49.102"}
{"id" : 22, "createTime" : "2019-05-25 00:09:49.102"}
{"id" : 23, "createTime" : "2019-05-25 00:19:49.102"}
...
{"id" : 163, "createTime" : "2019-05-25 23:39:49.102"}
{"id" : 164, "createTime" : "2019-05-25 23:49:49.102"}
{"id" : 165, "createTime" : "2019-05-25 23:59:49.102"}
{"id" : 166, "createTime" : "2019-05-26 00:09:49.102"}
...
{"id" : 308, "createTime" : "2019-05-26 23:49:49.102"}
{"id" : 309, "createTime" : "2019-05-26 23:59:49.102"}
{"id" : 310, "createTime" : "2019-05-27 00:09:49.102"}

0 - 21 是 24号

22 -  165 是 25 号

166 - 309 是 26 号

输出结果（程序中reduce 方法，将窗口中第一条和最后一条数据的id，都放到 Event中 ）如下：

　　

与测试数据对应

5、说明

　　很多人会错误的以为，窗口时间的开始时间会是程序启动（初始化）的时间。事实上，窗口（以TumblingEventTimeWindows为例）的定义有两个重载的方法：包含两个参数，窗口的长度和窗口的offset（默认为0）　

源码：org.apache.flink.streaming.api.windowing.assigners.TumblingEventTimeWindows ：　

@PublicEvolving
public class TumblingEventTimeWindows extends WindowAssigner<Object, TimeWindow> {
    private static final long serialVersionUID = 1L;

    private final long size;

    private final long offset;

    protected TumblingEventTimeWindows(long size, long offset) {
        if (Math.abs(offset) >= size) {
            throw new IllegalArgumentException("TumblingEventTimeWindows parameters must satisfy abs(offset) < size");
        }

        this.size = size;
        this.offset = offset;
    }

    @Override
    public Collection<TimeWindow> assignWindows(Object element, long timestamp, WindowAssignerContext context) {
        if (timestamp > Long.MIN_VALUE) {
            // Long.MIN_VALUE is currently assigned when no timestamp is present
            long start = TimeWindow.getWindowStartWithOffset(timestamp, offset, size);
            System.out.println("start : " + start + ", end : " + (start+size));
            String startStr =new SimpleDateFormat("yyyy-MM-dd HH:mm:ss.SSS").format(start);
            String endStar =new SimpleDateFormat("yyyy-MM-dd HH:mm:ss.SSS").format(start + size);
            System.out.println("window start: " + new SimpleDateFormat("yyyy-MM-dd HH:mm:ss.SSS").format(start));
            System.out.println("window end: " + new SimpleDateFormat("yyyy-MM-dd HH:mm:ss.SSS").format(start + size));
            return Collections.singletonList(new TimeWindow(start, start + size));
        } else {
            throw new RuntimeException("Record has Long.MIN_VALUE timestamp (= no timestamp marker). " +
                    "Is the time characteristic set to 'ProcessingTime', or did you forget to call " +
                    "'DataStream.assignTimestampsAndWatermarks(...)'?");
        }
    }/**
     * Creates a new {@code TumblingEventTimeWindows} {@link WindowAssigner} that assigns
     * elements to time windows based on the element timestamp.
     *
     * @param size The size of the generated windows.
     * @return The time policy.
     */
    public static TumblingEventTimeWindows of(Time size) {
        return new TumblingEventTimeWindows(size.toMilliseconds(), 0);
    }

    /**
     * Creates a new {@code TumblingEventTimeWindows} {@link WindowAssigner} that assigns
     * elements to time windows based on the element timestamp and offset.
     *
     * <p>For example, if you want window a stream by hour,but window begins at the 15th minutes
     * of each hour, you can use {@code of(Time.hours(1),Time.minutes(15))},then you will get
     * time windows start at 0:15:00,1:15:00,2:15:00,etc.
     *
     * <p>Rather than that,if you are living in somewhere which is not using UTC±00:00 time,
     * such as China which is using UTC+08:00,and you want a time window with size of one day,
     * and window begins at every 00:00:00 of local time,you may use {@code of(Time.days(1),Time.hours(-8))}.
     * The parameter of offset is {@code Time.hours(-8))} since UTC+08:00 is 8 hours earlier than UTC time.
     *
     * @param size The size of the generated windows.
     * @param offset The offset which window start would be shifted by.
     * @return The time policy.
     */
    public static TumblingEventTimeWindows of(Time size, Time offset) {
        return new TumblingEventTimeWindows(size.toMilliseconds(), offset.toMilliseconds());
    }
}

每条数据都会触发： assignWindows 方法

计算函数如下：

public static long getWindowStartWithOffset(long timestamp, long offset, long windowSize) {
        return timestamp - (timestamp - offset + windowSize) % windowSize;
    }

dubug 如下：

 

　　6、特别说明

　　　　FLink 1.6.3/1.7.1/1.7.2 在 TumblingEventTimeWindows 构造器上有个bug：offset 不能小于0， 但是of 方法中又说明，可以使用： of(Time.days(1),Time.hours(-8)) 表示在中国的 0 点开始的一天窗口。

 JIRA ： FLINK-11326 ，jira 上注明1.8.0 修复。(我本来准备提个bug的，有人先下手了)

 

这个bug 可以通过自己创建一个相同包的相同类，将对应代码修改即可。

flink 1.7.2 源码：

protected TumblingEventTimeWindows(long size, long offset) {
        if (offset < 0 || offset >= size) {
            throw new IllegalArgumentException("TumblingEventTimeWindows parameters must satisfy 0 <= offset < size");
        }

        this.size = size;
        this.offset = offset;
    }

最新版源码：

protected TumblingEventTimeWindows(long size, long offset) {
        if (Math.abs(offset) >= size) {
            throw new IllegalArgumentException("TumblingEventTimeWindows parameters must satisfy abs(offset) < size");
        }

        this.size = size;
        this.offset = offset;
    }

修改：

　　　　

7、上面的案例主要讲Flink 的窗口，pv、uv核心代码如下：

.keyBy(0)
      .window(TumblingProcessingTimeWindows.of(Time.days(1), Time.hours(-8)))
      .trigger(ContinuousProcessingTimeTrigger.of(Time.seconds(10)))
      .evictor(TimeEvictor.of(Time.seconds(0), true))
      .process(new ProcessWindowFunction[(String, String), (String, String, Long), Tuple, TimeWindow] {
        /*
        这是使用state是因为，窗口默认只会在创建结束的时候触发一次计算，然后数据结果，
        如果长时间的窗口，比如：一天的窗口，要是等到一天结束在输出结果，那还不如跑批。
        所有大窗口会添加trigger，以一定的频率输出中间结果。
        加evictor 是因为，每次trigger，触发计算是，窗口中的所有数据都会参与，所以数据会触发很多次，比较浪费，加evictor 驱逐已经计算过的数据，就不会重复计算了
        驱逐了已经计算过的数据，导致窗口数据不完全，所以需要state 存储我们需要的中间结果
         */
        var wordState: MapState[String, String] = _
        var pvCount: ValueState[Long] = _

        override def open(parameters: Configuration): Unit = {
          // new MapStateDescriptor[String, String]("word", classOf[String], classOf[String])
          wordState = getRuntimeContext.getMapState(new MapStateDescriptor[String, String]("word", classOf[String], classOf[String]))
          pvCount = getRuntimeContext.getState[Long](new ValueStateDescriptor[Long]("pvCount", classOf[Long]))
        }

        override def process(key: Tuple, context: Context, elements: Iterable[(String, String)], out: Collector[(String, String, Long)]): Unit = {


          var pv = 0;
          val elementsIterator = elements.iterator
          // 遍历窗口数据，获取唯一word
          while (elementsIterator.hasNext) {
            pv += 1
            val word = elementsIterator.next()._2
            wordState.put(word, null)
          }
          // add current
          pvCount.update(pvCount.value() + pv)
          var count: Long = 0
          val wordIterator = wordState.keys().iterator()
          while (wordIterator.hasNext) {
            wordIterator.next()
            count += 1
          }
          // uv
          out.collect((key.getField(0), "uv", count))
          out.collect(key.getField(0), "pv", pv)

        }
      })

完整代码见：  https://github.com/springMoon/flink-rookie/blob/master/src/main/scala/com/venn/demo/WordCountDistinct.scala