【源码】Join 算子 Transforation 过程 - Flink菜鸟

又看下了 Join 算子 Transformation 的过程，发现有使用 union 和 coGroup，比较特殊，就仔细梳理一下

join demo 代码：两个 Stream join 只能在窗口中进行 join，join 在处理无界数据集的时候，必须指定窗口，让无界数据变成有界数据，Flink 状态缓存左右两条流的部分数据做 join 联接，在时间（或条数、指定窗口）上清除超过 join 窗口范围的数据，Flink 状态的大小才能保持在一个合理的范围内，而不是一直增大，直到超出大小失败。

val join = process.join(map)
      .where(str => str)
      .equalTo(str => str)
      .window(TumblingProcessingTimeWindows.of(Time.minutes(1)))
      .apply(new JoinFunction[String, String, String] {
        override def join(first: String, second: String): String = {
          first + ";" + second
        }
      })

先是 input1.join(input2) , 使用 A、B 两个流创建一个 JoinedStreams， input1、input2 分别是左右两个流

DataStream.scala

def join[T2](otherStream: DataStream[T2]): JoinedStreams[T, T2] = {
    new JoinedStreams(this, otherStream)
  }

where、equalTo、window 没什么内容，跳过

apply 方法是 join 算子的关键

先使用 JoinedStreams 的 input1、input2 创建一个 JavaJoinedStreams （多直白的名字，上篇已经说过了， Java 开发了 Flink 的方法，Scala Api 相当于一个壳，调用了 Java 的内容）

对 JavaJoinedStreams 的对象 join 调用 Join 对应的内容，如果没有指定 trigger、evictor、allowedLateness 就是 null

返回的结果做为参数调用 asScalaStream，将 Java 的 DataStream 转为 Scala 的 DataStream，供后续使用

JoinedStreams.scala

def apply[T: TypeInformation](function: JoinFunction[T1, T2, T]): DataStream[T] = {
          // 创建 JavaJoinedStreams
          val join = new JavaJoinedStreams[T1, T2](input1.javaStream, input2.javaStream)

          asScalaStream(join
            .where(keySelector1)
            .equalTo(keySelector2)
            .window(windowAssigner)
            .trigger(trigger)
            .evictor(evictor)
            .allowedLateness(allowedLateness)
            // apply join
            .apply(clean(function), implicitly[TypeInformation[T]]))
        }

再看 Java的 JoinedStreams 的 apply 放，将 JoinedStreams 转成 CoGroupedStreams 来处理 join 算子，input、where、equalTo 等直接平移过来，最后调用 CoGroupedStreams 的 apply 方法

JoinedStreams.java

public <T> DataStream<T> apply(JoinFunction<T1, T2, T> function, TypeInformation<T> resultType) {
      //clean the closure
      function = input1.getExecutionEnvironment().clean(function);
      // join 变 coGroup 了, input1 input2 还是 他们
      coGroupedWindowedStream = input1.coGroup(input2)
        .where(keySelector1)
        .equalTo(keySelector2)
        .window(windowAssigner)
        .trigger(trigger)
        .evictor(evictor)
        .allowedLateness(allowedLateness);
      // 调用 coGroupedWindowedStream 的 apply 处理
      return coGroupedWindowedStream
          .apply(new JoinCoGroupFunction<>(function), resultType);
    }

CoGroupedStreams 的 apply 方法看着就更有意思了，将 input1、input2 转成类型是 TaggedUnion<T1, T2> 的 DataStream，对两个新流调用 map(new Input1Tagger<T1, T2>())、map(new Input2Tagger<T1, T2>()) 方法，将两个流的类型转成一样，只是在输出数据是，只有自己这边有数据，另一边直接给 null

有将两个流的 keySelect 组合成 unionKeySelector

使用 union 后的流，创建 KeyedStream 传入 unionKeySelector，指定分区的 PartitionTransformation，并生成窗口

最后调用 windowedStream.apply 方法

CoGroupedStreams.java

public <T> DataStream<T> apply(CoGroupFunction<T1, T2, T> function, TypeInformation<T> resultType) {
      //clean the closure
      function = input1.getExecutionEnvironment().clean(function);

      // 定义 union 的 UnionTypeInfo, 两种类型组合
      UnionTypeInfo<T1, T2> unionType = new UnionTypeInfo<>(input1.getType(), input2.getType());
      // 定义 union 的 KeySelector，两个 keySelector
      UnionKeySelector<T1, T2, KEY> unionKeySelector = new UnionKeySelector<>(keySelector1, keySelector2);

      // input1 创建 DataStream<TaggedUnion<T1, T2>> 指定返回类型是 unionType
      DataStream<TaggedUnion<T1, T2>> taggedInput1 = input1
          .map(new Input1Tagger<T1, T2>())
          .setParallelism(input1.getParallelism())
          .returns(unionType);
      // input2 创建 DataStream<TaggedUnion<T1, T2>> 指定返回类型是 unionType
      DataStream<TaggedUnion<T1, T2>> taggedInput2 = input2
          .map(new Input2Tagger<T1, T2>())
          .setParallelism(input2.getParallelism())
          .returns(unionType);
      // join 两个流，上面已经将两个流的类型转为一样了： DataStream<TaggedUnion<T1, T2>>
      DataStream<TaggedUnion<T1, T2>> unionStream = taggedInput1.union(taggedInput2);

      // we explicitly create the keyed stream to manually pass the key type information in
      // 使用 union 的 Stream 创建 KeyedStream，同时指定 分区的 PartitionTransformation
      // 调用window 生成 windowStream
      windowedStream =
          new KeyedStream<TaggedUnion<T1, T2>, KEY>(unionStream, unionKeySelector, keyType)
          .window(windowAssigner);

      if (trigger != null) {
        windowedStream.trigger(trigger);
      }
      if (evictor != null) {
        windowedStream.evictor(evictor);
      }
      if (allowedLateness != null) {
        windowedStream.allowedLateness(allowedLateness);
      }
      // 调用 windowedStream apply 方法，参数是个 CoGroupWindowFunction
      return windowedStream.apply(new CoGroupWindowFunction<T1, T2, T, KEY, W>(function), resultType);

Input1Tagger/Input2Tagger 的 map 方法

private static class Input1Tagger<T1, T2> implements MapFunction<T1, TaggedUnion<T1, T2>> {
    private static final long serialVersionUID = 1L;

    @Override
    public TaggedUnion<T1, T2> map(T1 value) throws Exception {
      return TaggedUnion.one(value);
    }
  }

  private static class Input2Tagger<T1, T2> implements MapFunction<T2, TaggedUnion<T1, T2>> {
    private static final long serialVersionUID = 1L;

    @Override
    public TaggedUnion<T1, T2> map(T2 value) throws Exception {
      return TaggedUnion.two(value);
    }
  }

TaggedUnion 的 one/two 方法

public static <T1, T2> TaggedUnion<T1, T2> one(T1 one) {
      // one 方法 右边数据为 null
      return new TaggedUnion<>(one, null);
    }

    public static <T1, T2> TaggedUnion<T1, T2> two(T2 two) {
      // two 方法 左边数据为 null
      return new TaggedUnion<>(null, two);
    }

KeyedStream

public KeyedStream(DataStream<T> dataStream, KeySelector<T, KEY> keySelector, TypeInformation<KEY> keyType) {
    this(
      dataStream,
      new PartitionTransformation<>(
        dataStream.getTransformation(),
        new KeyGroupStreamPartitioner<>(keySelector, StreamGraphGenerator.DEFAULT_LOWER_BOUND_MAX_PARALLELISM)),
      keySelector,
      keyType);
  }

windowStream.apply 方法

public <R> SingleOutputStreamOperator<R> apply(WindowFunction<T, R, K, W> function, TypeInformation<R> resultType) {
    function = input.getExecutionEnvironment().clean(function);
    return apply(new InternalIterableWindowFunction<>(function), resultType, function);
  }

再 apply，看到这里，又看到了熟悉的样子，先 operator，再 Transformation

private <R> SingleOutputStreamOperator<R> apply(InternalWindowFunction<Iterable<T>, R, K, W> function, TypeInformation<R> resultType, Function originalFunction) {

    // operatorName
    final String opName = generateOperatorName(windowAssigner, trigger, evictor, originalFunction, null);
    // keySelector 就是之前的 UnionKeySelector
    KeySelector<T, K> keySel = input.getKeySelector();

    WindowOperator<K, T, Iterable<T>, R, W> operator;

    if (evictor != null) {
      @SuppressWarnings({"unchecked", "rawtypes"})
      TypeSerializer<StreamRecord<T>> streamRecordSerializer =
          (TypeSerializer<StreamRecord<T>>) new StreamElementSerializer(input.getType().createSerializer(getExecutionEnvironment().getConfig()));

      ListStateDescriptor<StreamRecord<T>> stateDesc =
          new ListStateDescriptor<>("window-contents", streamRecordSerializer);

      operator =
        new EvictingWindowOperator<>(windowAssigner,
          windowAssigner.getWindowSerializer(getExecutionEnvironment().getConfig()),
          keySel,
          input.getKeyType().createSerializer(getExecutionEnvironment().getConfig()),
          stateDesc,
          function,
          trigger,
          evictor,
          allowedLateness,
          lateDataOutputTag);

    } else {
      ListStateDescriptor<T> stateDesc = new ListStateDescriptor<>("window-contents",
        input.getType().createSerializer(getExecutionEnvironment().getConfig()));
      // 创建 window 的 WindowOperator
      operator =
        new WindowOperator<>(windowAssigner,
          windowAssigner.getWindowSerializer(getExecutionEnvironment().getConfig()),
          keySel,
          input.getKeyType().createSerializer(getExecutionEnvironment().getConfig()),
          stateDesc,
          function,
          trigger,
          allowedLateness,
          lateDataOutputTag);
    }
    // 调用 transform 方法 生成 Transformation
    return input.transform(opName, resultType, operator);
  }

join 算子 doTransform 的过程，先生成了一个 OneInputTransformation, 再用 OneInputTransformation 生成了一个 SingleOutputStreamOperator 返回，所以最后是个 SingleOutputStreamOperator 的 DataStream

@PublicEvolving
  public <R> SingleOutputStreamOperator<R> transform(
      String operatorName,
      TypeInformation<R> outTypeInfo,
      OneInputStreamOperatorFactory<T, R> operatorFactory) {

    return doTransform(operatorName, outTypeInfo, operatorFactory);
  }

  protected <R> SingleOutputStreamOperator<R> doTransform(
      String operatorName,
      TypeInformation<R> outTypeInfo,
      StreamOperatorFactory<R> operatorFactory) {

    // read the output type of the input Transform to coax out errors about MissingTypeInfo
    transformation.getOutputType();

    OneInputTransformation<T, R> resultTransform = new OneInputTransformation<>(
        this.transformation,
        operatorName,
        operatorFactory,
        outTypeInfo,
        environment.getParallelism());

    @SuppressWarnings({"unchecked", "rawtypes"})
    SingleOutputStreamOperator<R> returnStream = new SingleOutputStreamOperator(environment, resultTransform);

    getExecutionEnvironment().addOperator(resultTransform);

    return returnStream;
  }

到这里简单总结下中间转换的过程：

* 1 先是将 join 转成 Scala JoinedStreams，再到 Java JoinedStreams
* 2 Java JoinedStreams 转成 CoGroupedStreams
* 3 CoGroupedStreams 转成 union （UnionTransformation ）
* 4 再转成 KeyedStream (PartitionTransformation)
* 5 再转成 SingleOutputStreamOperator （OneInputTransformation）

看起来略复杂，但是总体来说，还是跟之前的差不多

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

posted on 2020-11-05 15:15 Flink菜鸟阅读(452) 评论(0) 编辑收藏举报

刷新页面返回顶部