谈谈Flink DataStream API中的三种双流join操作

前言

本文是基础中的基础，看官可以放心食用。

在数据库中的静态表上做OLAP分析时，两表join是非常常见的操作。同理，在流式处理作业中，有时也需要在两条流上做join以获得更丰富的信息。Flink DataStream API为用户提供了3个算子来实现双流join，分别是：

join()
coGroup()
intervalJoin()

本文举例说明它们的使用方法，顺便聊聊比较特殊的interval join的原理。

准备数据

从Kafka分别接入点击流和订单流，并转化为POJO。

DataStream<String> clickSourceStream = env
  .addSource(new FlinkKafkaConsumer011<>(
    "ods_analytics_access_log",
    new SimpleStringSchema(),
    kafkaProps
  ).setStartFromLatest());
DataStream<String> orderSourceStream = env
  .addSource(new FlinkKafkaConsumer011<>(
    "ods_ms_order_done",
    new SimpleStringSchema(),
    kafkaProps
  ).setStartFromLatest());

DataStream<AnalyticsAccessLogRecord> clickRecordStream = clickSourceStream
  .map(message -> JSON.parseObject(message, AnalyticsAccessLogRecord.class));
DataStream<OrderDoneLogRecord> orderRecordStream = orderSourceStream
  .map(message -> JSON.parseObject(message, OrderDoneLogRecord.class));

join()

join()算子提供的语义为"Window join"，即按照指定字段和（滚动/滑动/会话）窗口进行inner join，支持处理时间和事件时间两种时间特征。

以下示例以10秒滚动窗口，将两个流通过商品ID关联，取得订单流中的售价相关字段。

clickRecordStream
  .join(orderRecordStream)
  .where(record -> record.getMerchandiseId())
  .equalTo(record -> record.getMerchandiseId())
  .window(TumblingProcessingTimeWindows.of(Time.seconds(10)))
  .apply(new JoinFunction<AnalyticsAccessLogRecord, OrderDoneLogRecord, String>() {
    @Override
    public String join(AnalyticsAccessLogRecord accessRecord, OrderDoneLogRecord orderRecord) throws Exception {
      return StringUtils.join(Arrays.asList(
        accessRecord.getMerchandiseId(),
        orderRecord.getPrice(),
        orderRecord.getCouponMoney(),
        orderRecord.getRebateAmount()
      ), '\t');
    }
  })
  .print().setParallelism(1);

简单易用。

coGroup()

只有inner join肯定还不够，如何实现left/right outer join呢？答案就是利用coGroup()算子。它的调用方式类似于join()算子，也需要开窗，但是CoGroupFunction比JoinFunction更加灵活，可以按照用户指定的逻辑匹配左流和/或右流的数据并输出。

以下的例子就实现了点击流left join订单流的功能，是很朴素的nested loop join思想（二重循环）。

clickRecordStream
  .coGroup(orderRecordStream)
  .where(record -> record.getMerchandiseId())
  .equalTo(record -> record.getMerchandiseId())
  .window(TumblingProcessingTimeWindows.of(Time.seconds(10)))
  .apply(new CoGroupFunction<AnalyticsAccessLogRecord, OrderDoneLogRecord, Tuple2<String, Long>>() {
    @Override
    public void coGroup(Iterable<AnalyticsAccessLogRecord> accessRecords, Iterable<OrderDoneLogRecord> orderRecords, Collector<Tuple2<String, Long>> collector) throws Exception {
      for (AnalyticsAccessLogRecord accessRecord : accessRecords) {
        boolean isMatched = false;
        for (OrderDoneLogRecord orderRecord : orderRecords) {
          // 右流中有对应的记录
          collector.collect(new Tuple2<>(accessRecord.getMerchandiseName(), orderRecord.getPrice()));
          isMatched = true;
        }
        if (!isMatched) {
          // 右流中没有对应的记录
          collector.collect(new Tuple2<>(accessRecord.getMerchandiseName(), null));
        }
      }
    }
  })
  .print().setParallelism(1);

intervalJoin()

join()和coGroup()都是基于窗口做关联的。但是在某些情况下，两条流的数据步调未必一致。例如，订单流的数据有可能在点击流的购买动作发生之后很久才被写入，如果用窗口来圈定，很容易join不上。所以Flink又提供了"Interval join"的语义，按照指定字段以及右流相对左流偏移的时间区间进行关联，即：

right.timestamp ∈ [left.timestamp + lowerBound; left.timestamp + upperBound]

interval join也是inner join，虽然不需要开窗，但是需要用户指定偏移区间的上下界，并且只支持事件时间。

示例代码如下。注意在运行之前，需要分别在两个流上应用assignTimestampsAndWatermarks()方法获取事件时间戳和水印。

clickRecordStream
  .keyBy(record -> record.getMerchandiseId())
  .intervalJoin(orderRecordStream.keyBy(record -> record.getMerchandiseId()))
  .between(Time.seconds(-30), Time.seconds(30))
  .process(new ProcessJoinFunction<AnalyticsAccessLogRecord, OrderDoneLogRecord, String>() {
    @Override
    public void processElement(AnalyticsAccessLogRecord accessRecord, OrderDoneLogRecord orderRecord, Context context, Collector<String> collector) throws Exception {
      collector.collect(StringUtils.join(Arrays.asList(
        accessRecord.getMerchandiseId(),
        orderRecord.getPrice(),
        orderRecord.getCouponMoney(),
        orderRecord.getRebateAmount()
      ), '\t'));
    }
  })
  .print().setParallelism(1);

由上可见，interval join与window join不同，是两个KeyedStream之上的操作，并且需要调用between()方法指定偏移区间的上下界。如果想令上下界是开区间，可以调用upperBoundExclusive()/lowerBoundExclusive()方法。

interval join的实现原理

以下是KeyedStream.process(ProcessJoinFunction)方法调用的重载方法的逻辑。

public <OUT> SingleOutputStreamOperator<OUT> process(
        ProcessJoinFunction<IN1, IN2, OUT> processJoinFunction,
        TypeInformation<OUT> outputType) {
    Preconditions.checkNotNull(processJoinFunction);
    Preconditions.checkNotNull(outputType);
    final ProcessJoinFunction<IN1, IN2, OUT> cleanedUdf = left.getExecutionEnvironment().clean(processJoinFunction);
    final IntervalJoinOperator<KEY, IN1, IN2, OUT> operator =
        new IntervalJoinOperator<>(
            lowerBound,
            upperBound,
            lowerBoundInclusive,
            upperBoundInclusive,
            left.getType().createSerializer(left.getExecutionConfig()),
            right.getType().createSerializer(right.getExecutionConfig()),
            cleanedUdf
        );
    return left
        .connect(right)
        .keyBy(keySelector1, keySelector2)
        .transform("Interval Join", outputType, operator);
}

可见是先对两条流执行connect()和keyBy()操作，然后利用IntervalJoinOperator算子进行转换。在IntervalJoinOperator中，会利用两个MapState分别缓存左流和右流的数据。

private transient MapState<Long, List<BufferEntry<T1>>> leftBuffer;
private transient MapState<Long, List<BufferEntry<T2>>> rightBuffer;

@Override
public void initializeState(StateInitializationContext context) throws Exception {
    super.initializeState(context);
    this.leftBuffer = context.getKeyedStateStore().getMapState(new MapStateDescriptor<>(
        LEFT_BUFFER,
        LongSerializer.INSTANCE,
        new ListSerializer<>(new BufferEntrySerializer<>(leftTypeSerializer))
    ));
    this.rightBuffer = context.getKeyedStateStore().getMapState(new MapStateDescriptor<>(
        RIGHT_BUFFER,
        LongSerializer.INSTANCE,
        new ListSerializer<>(new BufferEntrySerializer<>(rightTypeSerializer))
    ));
}

其中Long表示事件时间戳，List<BufferEntry<T>>表示该时刻到来的数据记录。

当左流和右流有数据到达时，会分别调用processElement1()和processElement2()方法，它们都调用了processElement()方法，代码如下。

@Override
public void processElement1(StreamRecord<T1> record) throws Exception {
    processElement(record, leftBuffer, rightBuffer, lowerBound, upperBound, true);
}

@Override
public void processElement2(StreamRecord<T2> record) throws Exception {
    processElement(record, rightBuffer, leftBuffer, -upperBound, -lowerBound, false);
}

@SuppressWarnings("unchecked")
private <THIS, OTHER> void processElement(
        final StreamRecord<THIS> record,
        final MapState<Long, List<IntervalJoinOperator.BufferEntry<THIS>>> ourBuffer,
        final MapState<Long, List<IntervalJoinOperator.BufferEntry<OTHER>>> otherBuffer,
        final long relativeLowerBound,
        final long relativeUpperBound,
        final boolean isLeft) throws Exception {
    final THIS ourValue = record.getValue();
    final long ourTimestamp = record.getTimestamp();
    if (ourTimestamp == Long.MIN_VALUE) {
        throw new FlinkException("Long.MIN_VALUE timestamp: Elements used in " +
                "interval stream joins need to have timestamps meaningful timestamps.");
    }
    if (isLate(ourTimestamp)) {
        return;
    }
    addToBuffer(ourBuffer, ourValue, ourTimestamp);
    for (Map.Entry<Long, List<BufferEntry<OTHER>>> bucket: otherBuffer.entries()) {
        final long timestamp  = bucket.getKey();
        if (timestamp < ourTimestamp + relativeLowerBound ||
                timestamp > ourTimestamp + relativeUpperBound) {
            continue;
        }
        for (BufferEntry<OTHER> entry: bucket.getValue()) {
            if (isLeft) {
                collect((T1) ourValue, (T2) entry.element, ourTimestamp, timestamp);
            } else {
                collect((T1) entry.element, (T2) ourValue, timestamp, ourTimestamp);
            }
        }
    }
    long cleanupTime = (relativeUpperBound > 0L) ? ourTimestamp + relativeUpperBound : ourTimestamp;
    if (isLeft) {
        internalTimerService.registerEventTimeTimer(CLEANUP_NAMESPACE_LEFT, cleanupTime);
    } else {
        internalTimerService.registerEventTimeTimer(CLEANUP_NAMESPACE_RIGHT, cleanupTime);
    }
}

这段代码的思路是：

取得当前流StreamRecord的时间戳，调用isLate()方法判断它是否是迟到数据（即时间戳小于当前水印值），如是则丢弃。
调用addToBuffer()方法，将时间戳和数据一起插入当前流对应的MapState。
遍历另外一个流的MapState，如果数据满足前述的时间区间条件，则调用collect()方法将该条数据投递给用户定义的ProcessJoinFunction进行处理。collect()方法的代码如下，注意结果对应的时间戳是左右流时间戳里较大的那个。

private void collect(T1 left, T2 right, long leftTimestamp, long rightTimestamp) throws Exception {
    final long resultTimestamp = Math.max(leftTimestamp, rightTimestamp);
    collector.setAbsoluteTimestamp(resultTimestamp);
    context.updateTimestamps(leftTimestamp, rightTimestamp, resultTimestamp);
    userFunction.processElement(left, right, context, collector);
}

调用TimerService.registerEventTimeTimer()注册时间戳为timestamp + relativeUpperBound的定时器，该定时器负责在水印超过区间的上界时执行状态的清理逻辑，防止数据堆积。注意左右流的定时器所属的namespace是不同的，具体逻辑则位于onEventTime()方法中。

@Override
public void onEventTime(InternalTimer<K, String> timer) throws Exception {
    long timerTimestamp = timer.getTimestamp();
    String namespace = timer.getNamespace();
    logger.trace("onEventTime @ {}", timerTimestamp);
    switch (namespace) {
        case CLEANUP_NAMESPACE_LEFT: {
            long timestamp = (upperBound <= 0L) ? timerTimestamp : timerTimestamp - upperBound;
            logger.trace("Removing from left buffer @ {}", timestamp);
            leftBuffer.remove(timestamp);
            break;
        }
        case CLEANUP_NAMESPACE_RIGHT: {
            long timestamp = (lowerBound <= 0L) ? timerTimestamp + lowerBound : timerTimestamp;
            logger.trace("Removing from right buffer @ {}", timestamp);
            rightBuffer.remove(timestamp);
            break;
        }
        default:
            throw new RuntimeException("Invalid namespace " + namespace);
    }
}

The End

民那晚安晚安。