• 【源码解析】Flink 是如何处理迟到数据


    相信会看到这篇文章的都对Flink的时间类型(事件时间、处理时间、摄入时间)和Watermark有些了解,当然不了解可以先看下官网的介绍:https://ci.apache.org/projects/flink/flink-docs-master/dev/event_time.html 

    这里就会有这样一个问题:FLink 是怎么基于事件时间和Watermark处理迟到数据的呢

    在回答这个问题之前,建议大家可以看下下面的Google 的三篇论文,关于流处理的模型:

    https://www.vldb.org/pvldb/vol8/p1792-Akidau.pdf 《The Dataflow Model: A Practical Approach to Balancing Correctness, Latency, and Cost in Massive-Scale, Unbounded, Out-of-Order Data Processing》

    high-level的现代数据处理概念指引:

    https://www.oreilly.com/ideas/the-world-beyond-batch-streaming-101

    https://www.oreilly.com/ideas/the-world-beyond-batch-streaming-102

    ---------------------------进入正题--------------------------------

    现在进入正题:FLink 是怎么基于事件时间和Watermark处理迟到数据的呢?

    这个问题可以分成两个部分:

      1. 基于事件时间创建Timestamp 和Watermark(后面会详细介绍)
    
      2. 处理迟到数据

    1. 基于事件时间创建Timestamp 和Watermark

     为了方便查看,这里使用 assignTimestampsAndWatermarks(assigner: AssignerWithPeriodicWatermarks[T]): DataStream[T]  重载方法基于每个事件生成水印代码如下:

    val input = env.addSource(source)
          .map(json => {
            // json : {"id" : 0, "createTime" : "2019-08-24 11:13:14.942", "amt" : "9.8"}
            val id = json.get("id").asText()
            val createTime = json.get("createTime").asText()
            val amt = json.get("amt").asText()
            LateDataEvent("key", id, createTime, amt)
          })
          // assign watermarks every event
          .assignTimestampsAndWatermarks(new AssignerWithPunctuatedWatermarks[LateDataEvent]() {
          // check extractTimestamp emitted watermark is non-null and large than previously
          override def checkAndGetNextWatermark(lastElement: LateDataEvent, extractedTimestamp: Long): Watermark = {
            new Watermark(extractedTimestamp)
          }
          // generate next watermark
          override def extractTimestamp(element: LateDataEvent, previousElementTimestamp: Long): Long = {
            val eventTime = sdf.parse(element.createTime).getTime
            eventTime
          }
        })

    ```

    扩展:数据在算子中是以StreamRecord 对象作为流转抽象结构如下:

    public final class StreamRecord<T> extends StreamElement {
    
        /** The actual value held by this record.  具体数据*/
        private T value;
    
        /** The timestamp of the record.  该数据对应的时间戳 */
        private long timestamp;
    
    }

    StreamElement 也是 Watermark 和 StreamStatus的父类,简单来说就是Flink 承载消息的基类(这里可以指定,Watermark 是和事件一个级别的抽象,而Timestamp 是Watermark和事件的成员变量,代表Watermark和事件的时间)

    ```

    assignTimestampsAndWatermarks 是基于事件的数据(extractTimestamp 方法中返回的Timestamp),替换StreamRecord 对象中的Timestamp和发出新的Watermark(如果当前事件的Timestamp 生成的Watermark大于上一次的Watermark)

    下面我们来debug这部分源码:

    首先在extractTimestamp  方法中添加断点查看Timestamp 和Watermark的生成:

    TimestampsAndPunctuatedWatermarksOperator.processElement(使用的类取决于assignTimestampsAndWatermarks 方法的参数) 中处理事件的Timestamp和对应的Watermark

    StreamRecord对象的创建在 StreamSourceContexts.processAndCollectWithTimestamp 中,使用的Timestamp 是数据在kafka的时间,在KafkaFetcher.emitRecord方法中从consumerRecord中获取

    KafkaFetcher.emitRecord   发出从kafka中消费到的数据:

    protected void emitRecord(
            T record,
            KafkaTopicPartitionState<TopicPartition> partition,
            long offset,
            ConsumerRecord<?, ?> consumerRecord) throws Exception {
    
            emitRecordWithTimestamp(record, partition, offset, consumerRecord.timestamp());
        }

    StreamSourceContexts.processAndCollectWithTimestamp 创建StreamRecord 对象

    protected void processAndCollectWithTimestamp(T element, long timestamp) {
                output.collect(reuse.replace(element, timestamp)); // 放入真正的事件时间戳
            }

    下面我们来看 TimestampsAndPunctuatedWatermarksOperator.processElement 的源码 

    @Override
        public void processElement(StreamRecord<T> element) throws Exception {
            // 获取这条数据
            final T value = element.getValue();
            // userFunction 就是代码里面创建的匿名类 AssignerWithPunctuatedWatermarks
            // 调用extractTimestamp,获取新的Timestamp
            // element.hasTimestamp 有的话就用,没有就给默认值long类型 的最小值
            final long newTimestamp = userFunction.extractTimestamp(value,
                    element.hasTimestamp() ? element.getTimestamp() : Long.MIN_VALUE);
            // 使用新的Timestamp 替换StreamRecord 旧的Timestamp
            output.collect(element.replace(element.getValue(), newTimestamp));
            // 获取下一个Watermark,调用实现的 checkAndGetNextWatermark 方法
            final Watermark nextWatermark = userFunction.checkAndGetNextWatermark(value, newTimestamp);
            // 如果新的Watermark 大于上一个Watermark 就发出新的
            if (nextWatermark != null && nextWatermark.getTimestamp() > currentWatermark) {
                currentWatermark = nextWatermark.getTimestamp();
                output.emitWatermark(nextWatermark);
            }
        }

    至此Timestamp和Watermark的创建(或者说生成)就好了

    2. Flink 处理迟到数据

      为了演示这个功能,在上面的程序中添加了window算子迟到数据侧边输出的方法 sideOutputLateData,为了方便查看,这里再添加一次全部代码

    val source = new FlinkKafkaConsumer[ObjectNode]("late_data", new JsonNodeDeserializationSchema(), Common.getProp)
        // 侧边输出的tag
        val late = new OutputTag[LateDataEvent]("late")
    
        val input = env.addSource(source)
          .map(json => {
            // json : {"id" : 0, "createTime" : "2019-08-24 11:13:14.942", "amt" : "9.8"}
            val id = json.get("id").asText()
            val createTime = json.get("createTime").asText()
            val amt = json.get("amt").asText()
            LateDataEvent("key", id, createTime, amt)
          })
          // assign watermarks every event
          .assignTimestampsAndWatermarks(new AssignerWithPunctuatedWatermarks[LateDataEvent]() {
          // check extractTimestamp emitted watermark is non-null and large than previously
          override def checkAndGetNextWatermark(lastElement: LateDataEvent, extractedTimestamp: Long): Watermark = {
            new Watermark(extractedTimestamp)
          }
          // generate next watermark
          override def extractTimestamp(element: LateDataEvent, previousElementTimestamp: Long): Long = {
            val eventTime = sdf.parse(element.createTime).getTime
            eventTime
          }
        })
          // after keyBy will have window number of different key
          .keyBy("key")
          .window(TumblingEventTimeWindows.of(Time.minutes(1)))
          // get lateData
          .sideOutputLateData(late)
          .process(new ProcessWindowFunction[LateDataEvent, LateDataEvent, Tuple, TimeWindow] {
            // just for debug window process late data
            override def process(key: Tuple, context: Context, elements: Iterable[LateDataEvent], out: Collector[LateDataEvent]): Unit = {
              // print window start timestamp & end timestamp & current watermark time
              println("window:" + context.window.getStart + "-" + context.window.getEnd + ", currentWatermark : " + context.currentWatermark)
              val it = elements.toIterator
              while (it.hasNext) {
                val current = it.next()
                out.collect(current)
              }
            }
          })
        // print late data
        input.getSideOutput(late).print("late:")
        input.print("apply:")
        env.execute("LateDataProcess")

    代码逻辑很简单,主要是为了加入window算子,process算子是为了方便debug到window算子中

    下面开始debug源码:

    在process 方法中添加断点:

    这次直接从window算子接收上游发过来的数据开始看起:

    StreamInputProcessor.processInput方法负责将接收到的事件(数据、Watermark、StreamStatus、LatencyMarker),反序列化为 StreamElement(上文已经说得了,是事件抽象的基类),判断具体是那种消息,分别进行处理

    public boolean processInput() throws Exception {
            
            while (true) {
                if (currentRecordDeserializer != null) {
                    DeserializationResult result = currentRecordDeserializer.getNextRecord(deserializationDelegate);
    
                    if (result.isBufferConsumed()) {
                        currentRecordDeserializer.getCurrentBuffer().recycleBuffer();
                        currentRecordDeserializer = null;
                    }
    
                    if (result.isFullRecord()) {
                        StreamElement recordOrMark = deserializationDelegate.getInstance();
    
                        if (recordOrMark.isWatermark()) {
                            // handle watermark
                            statusWatermarkValve.inputWatermark(recordOrMark.asWatermark(), currentChannel);
                            continue;
                        } else if (recordOrMark.isStreamStatus()) {
                            // handle stream status
                            statusWatermarkValve.inputStreamStatus(recordOrMark.asStreamStatus(), currentChannel);
                            continue;
                        } else if (recordOrMark.isLatencyMarker()) {
                            // handle latency marker
                            synchronized (lock) {
                                streamOperator.processLatencyMarker(recordOrMark.asLatencyMarker());
                            }
                            continue;
                        } else {
                            // now we can do the actual processing
                            StreamRecord<IN> record = recordOrMark.asRecord();
                            synchronized (lock) {
                                numRecordsIn.inc();
                                streamOperator.setKeyContextElement1(record);
                                streamOperator.processElement(record);
                            }
                            return true;
                        }
                    }
                }
    
            }
        

    注:代码比较长,挑选了跟这次主题相关的部分

    Watermark:

    数据:

     这里我们主要看数据的处理逻辑:

    // now we can do the actual processing
    StreamRecord<IN> record = recordOrMark.asRecord();
    synchronized (lock) {
        // metric 的Counter,统计有多少条数据进来
        numRecordsIn.inc();  
        // 选择当前的key(类似与数据分区,每个key一个,里面存储自己的states)
        streamOperator.setKeyContextElement1(record);
        // 真正在进到WindowOperator 中处理数据了
        streamOperator.processElement(record);
    }
    就到了 WindowOperator.processElement 方法(主要判断逻辑都在这里)
    // 判断windowAssigner 是不是MergingWindowAssigner 
    if (windowAssigner instanceof MergingWindowAssigner) 

    区分开会话窗口和滑动、跳动窗口的处理逻辑,会话窗口的各个key的窗口是不对齐的

    直接到 else部分:

    } else {
        for (W window: elementWindows) {
    
            // drop if the window is already late 判断窗口数据是否迟到
            // 是,就直接跳过这条数据,重新处理下一条数据
            if (isWindowLate(window)) {
                continue;
            }

    PS: 写了这么久,终于到迟到数据处理的地方了 -_-

    下面看下 isWindowLate 部分的处理逻辑:

    /**
     * Returns {@code true} if the watermark is after the end timestamp plus the allowed lateness
     * of the given window.
     */
    protected boolean isWindowLate(W window) {
        // 只有事件时间下,并且 窗口元素的最大时间 + 允许迟到时间 <= 当前Watermark 的时候为true(即当前窗口元素迟到了)
        return (windowAssigner.isEventTime() && (cleanupTime(window) <= internalTimerService.currentWatermark()));
    }        
    
    /**
     * Returns the cleanup time for a window, which is
     * {@code window.maxTimestamp + allowedLateness}. In
     * case this leads to a value greater than {@link Long#MAX_VALUE}
     * then a cleanup time of {@link Long#MAX_VALUE} is
     * returned.
     * 返回窗口的cleanup 时间, 窗口的最大时间 + 允许延迟的时间
     * @param window the window whose cleanup time we are computing.
     */
    private long cleanupTime(W window) {
        if (windowAssigner.isEventTime()) {
            long cleanupTime = window.maxTimestamp() + allowedLateness;
            return cleanupTime >= window.maxTimestamp() ? cleanupTime : Long.MAX_VALUE;
        } else {
            return window.maxTimestamp();
        }
    }

    看一条正常到达的数据

    {"id" : 891, "createTime" : "2019-08-24 17:51:44.152", "amt" : "5.6"}

    891 这条数据的事件时间是:2019-08-24 17:51:44.152 ,1 分钟的整分窗口,这条数据对应的窗口就是: [2019-08-24 17:51:00.000, 2019-08-24 17:52:000) ,对应的时间戳是 : [1566640260000, 1566640320000) ,当前的Watermark 是 : 1566640294102,窗口数据的最大时间戳大于 当前的Watermark, 不是迟到数据,不跳过。

     现在在来看一条迟到的数据

    {"id" : 892, "createTime" : "2019-08-24 17:51:54.152", "amt" : "3.6"}

    892 这条数据的事件时间是:2019-08-24 17:51:54.152 ,1 分钟的整分窗口,这条数据对应的窗口就是: [2019-08-24 17:51:00.000, 2019-08-24 17:52:000) ,对应的时间戳是 : [1566640260000, 1566640320000) ,当前的Watermark 是 : 1566652224102 ,窗口数据的最大时间戳小于 当前的Watermark, 数据是迟到数据,跳过。

     上面就是窗口对迟到数据的处理源码dubug了,到这里就已经讲完Flink 处理迟到数据的两个部分:

      1. 基于事件时间创建Timestamp 和Watermark(后面会详细介绍)
    
      2. 窗口处理迟到数据

    注: 这里加上“窗口”,明确是window 算子做的这些事情

    下面在来看下窗口迟到输出的SideOutput ,源码在:WindowOperator.processElement 方法的最后一段:

    // side output input event if 事件时间
    // element not handled by any window 没有window处理过这条数据,上面isSkippedElement 默认值为true,如果上面判断为迟到数据,isSkippedElement就会为false
    // late arriving tag has been set
    // windowAssigner is event time and current timestamp + allowed lateness no less than element timestamp
    if (isSkippedElement && isElementLate(element)) {
        // 设置了 lateDataOutputTag 即window 算子后面的  .sideOutputLateData(late) 
        if (lateDataOutputTag != null){
            sideOutput(element);
        } else {
            this.numLateRecordsDropped.inc();
        }
    }
    
    /**
     * Decide if a record is currently late, based on current watermark and allowed lateness.
     * 事件时间,并且 元素的时间戳 + 允许延迟的时间 <= 当前watermark 是为true
     * @param element The element to check
     * @return The element for which should be considered when sideoutputs
     */
    protected boolean isElementLate(StreamRecord<IN> element){
        return (windowAssigner.isEventTime()) &&
            (element.getTimestamp() + allowedLateness <= internalTimerService.currentWatermark());
    }
    
    /**
     * Write skipped late arriving element to SideOutput.
     * 
     * @param element skipped late arriving element to side output
     */
    protected void sideOutput(StreamRecord<IN> element){
        output.collect(lateDataOutputTag, element);
    }

    搞定

    欢迎关注Flink菜鸟公众号,会不定期更新Flink(开发技术)相关的推文

  • 相关阅读:
    VB程序破解常用函数
    去VB程序NAG窗口方法-4C法
    error LNK2005: _DllMain@12 已经在 XXXX.obj 中定义
    汇编中的test和cmp指令
    OD保存修改后的数据到EXE
    C++ 异常捕获 try 和 __try的区别
    CListCtrl选中行
    WindowsAPI解析IAT地址
    Usaco 4.3.1 Buy Low, Buy Lower 逢低吸纳详细解题报告
    全国青少年信息学奥林匹克分区联赛(N)竞赛大纲
  • 原文地址:https://www.cnblogs.com/Springmoon-venn/p/11403740.html
Copyright © 2020-2023  润新知