Flink 原理与实现:如何生成 StreamGraph
nwfj2581
8年前
<p>继上文 <a href="http://www.open-open.com/lib/view/open1462323115590.html">Flink 原理与实现:架构和拓扑概览</a> 中介绍了Flink的四层执行图模型,本文将主要介绍 Flink 是如何根据用户用Stream API编写的程序,构造出一个代表拓扑结构的StreamGraph的。</p> <p>注:本文比较偏源码分析,所有代码都是基于 flink-1.0.x 版本,建议在阅读本文前先对Stream API有个了解,详见 <a href="/misc/goto?guid=4959672280338512252" rel="nofollow,noindex">官方文档</a> 。</p> <p>StreamGraph 相关的代码主要在 org.apache.flink.streaming.api.graph 包中。构造StreamGraph的入口函数是 StreamGraphGenerator.generate(env, transformations) 。该函数会由触发程序执行的方法 StreamExecutionEnvironment.execute() 调用到。也就是说 StreamGraph 是在 Client 端构造的,这也意味着我们可以在本地通过调试观察 StreamGraph 的构造过程。</p> <h2>Transformation</h2> <p>StreamGraphGenerator.generate 的一个关键的参数是 List<StreamTransformation<?>> 。 StreamTransformation 代表了从一个或多个 DataStream 生成新 DataStream 的操作。 DataStream 的底层其实就是一个 StreamTransformation ,描述了这个 DataStream 是怎么来的。</p> <p>StreamTransformation的类图如下图所示:</p> <p><img src="https://simg.open-open.com/show/e4887e5656c77939e4459cff7515f44c.png"></p> <p>DataStream 上常见的 transformation 有 map、flatmap、filter等(见 <a href="/misc/goto?guid=4959672280422365277" rel="nofollow,noindex">DataStream Transformation</a> 了解更多)。这些transformation会构造出一棵 StreamTransformation 树,通过这棵树转换成 StreamGraph。比如 DataStream.map 源码如下,其中 SingleOutputStreamOperator 为DataStream的子类:</p> <pre> <code class="language-java">public <R> SingleOutputStreamOperator<R> map(MapFunction<T, R> mapper) { // 通过java reflection抽出mapper的返回值类型 TypeInformation<R> outType = TypeExtractor.getMapReturnTypes(clean(mapper), getType(), Utils.getCallLocationName(), true); // 返回一个新的DataStream,SteramMap 为 StreamOperator 的实现类 return transform("Map", outType, new StreamMap<>(clean(mapper))); } public <R> SingleOutputStreamOperator<R> transform(String operatorName, TypeInformation<R> outTypeInfo, OneInputStreamOperator<T, R> operator) { // read the output type of the input Transform to coax out errors about MissingTypeInfo transformation.getOutputType(); // 新的transformation会连接上当前DataStream中的transformation,从而构建成一棵树 OneInputTransformation<T, R> resultTransform = new OneInputTransformation<>( this.transformation, operatorName, operator, outTypeInfo, environment.getParallelism()); @SuppressWarnings({ "unchecked", "rawtypes" }) SingleOutputStreamOperator<R> returnStream = new SingleOutputStreamOperator(environment, resultTransform); // 所有的transformation都会存到 env 中,调用execute时遍历该list生成StreamGraph getExecutionEnvironment().addOperator(resultTransform); return returnStream; } </code></pre> <p>从上方代码可以了解到,map转换将用户自定义的函数 MapFunction 包装到 StreamMap 这个Operator中,再将 StreamMap 包装到 OneInputTransformation ,最后该transformation存到env中,当调用 env.execute 时,遍历其中的transformation集合构造出StreamGraph。其分层实现如下图所示:</p> <p><img src="https://simg.open-open.com/show/92f4cfb14798149881b746ec895ca5ed.png"></p> <p>另外,并不是每一个 StreamTransformation 都会转换成 runtime 层中物理操作。有一些只是逻辑概念,比如 union、split/select、partition等。如下图所示的转换树,在运行时会优化成下方的操作图。</p> <p><img src="https://simg.open-open.com/show/79371524f1ec00f9c90c18a932fa9002.png"></p> <p>union、split/select、partition中的信息会被写入到 Source –> Map 的边中。通过源码也可以发现, UnionTransformation , SplitTransformation , SelectTransformation , PartitionTransformation 由于不包含具体的操作所以都没有StreamOperator成员变量,而其他StreamTransformation的子类基本上都有。</p> <h2>StreamOperator</h2> <p>DataStream 上的每一个 Transformation 都对应了一个 StreamOperator,StreamOperator是运行时的具体实现,会决定UDF(User-Defined Funtion)的调用方式。下图所示为 StreamOperator 的类图(点击查看大图):</p> <p><img src="https://simg.open-open.com/show/8e38c6140860ac0afae095d56b5317d1.png"></p> <p>可以发现,所有实现类都继承了 AbstractStreamOperator 。另外除了 project 操作,其他所有可以执行UDF代码的实现类都继承自 AbstractUdfStreamOperator ,该类是封装了UDF的StreamOperator。UDF就是实现了 Function 接口的类,如 MapFunction , FilterFunction 。</p> <h2>生成 StreamGraph 的源码分析</h2> <p>我们通过在DataStream上做了一系列的转换(map、filter等)得到了StreamTransformation集合,然后通过 StreamGraphGenerator.generate 获得StreamGraph,该方法的源码如下:</p> <pre> <code class="language-java">// 构造 StreamGraph 入口函数 public static StreamGraph generate(StreamExecutionEnvironment env, List<StreamTransformation<?>> transformations) { return new StreamGraphGenerator(env).generateInternal(transformations); } // 自底向上(sink->source)对转换树的每个transformation进行转换。 private StreamGraph generateInternal(List<StreamTransformation<?>> transformations) { for (StreamTransformation<?> transformation: transformations) { transform(transformation); } return streamGraph; } // 对具体的一个transformation进行转换,转换成 StreamGraph 中的 StreamNode 和 StreamEdge // 返回值为该transform的id集合,通常大小为1个(除FeedbackTransformation) private Collection<Integer> transform(StreamTransformation<?> transform) { // 跳过已经转换过的transformation if (alreadyTransformed.containsKey(transform)) { return alreadyTransformed.get(transform); } LOG.debug("Transforming " + transform); // 为了触发 MissingTypeInfo 的异常 transform.getOutputType(); Collection<Integer> transformedIds; if (transform instanceof OneInputTransformation<?, ?>) { transformedIds = transformOnInputTransform((OneInputTransformation<?, ?>) transform); } else if (transform instanceof TwoInputTransformation<?, ?, ?>) { transformedIds = transformTwoInputTransform((TwoInputTransformation<?, ?, ?>) transform); } else if (transform instanceof SourceTransformation<?>) { transformedIds = transformSource((SourceTransformation<?>) transform); } else if (transform instanceof SinkTransformation<?>) { transformedIds = transformSink((SinkTransformation<?>) transform); } else if (transform instanceof UnionTransformation<?>) { transformedIds = transformUnion((UnionTransformation<?>) transform); } else if (transform instanceof SplitTransformation<?>) { transformedIds = transformSplit((SplitTransformation<?>) transform); } else if (transform instanceof SelectTransformation<?>) { transformedIds = transformSelect((SelectTransformation<?>) transform); } else if (transform instanceof FeedbackTransformation<?>) { transformedIds = transformFeedback((FeedbackTransformation<?>) transform); } else if (transform instanceof CoFeedbackTransformation<?>) { transformedIds = transformCoFeedback((CoFeedbackTransformation<?>) transform); } else if (transform instanceof PartitionTransformation<?>) { transformedIds = transformPartition((PartitionTransformation<?>) transform); } else { throw new IllegalStateException("Unknown transformation: " + transform); } // need this check because the iterate transformation adds itself before // transforming the feedback edges if (!alreadyTransformed.containsKey(transform)) { alreadyTransformed.put(transform, transformedIds); } if (transform.getBufferTimeout() > 0) { streamGraph.setBufferTimeout(transform.getId(), transform.getBufferTimeout()); } if (transform.getUid() != null) { streamGraph.setTransformationId(transform.getId(), transform.getUid()); } return transformedIds; } </code></pre> <p>最终都会调用 transformXXX 来对具体的StreamTransformation进行转换。我们可以看下 transformOnInputTransform(transform) 的实现:</p> <pre> <code class="language-java">private <IN, OUT> Collection<Integer> transformOnInputTransform(OneInputTransformation<IN, OUT> transform) { // 递归对该transform的直接上游transform进行转换,获得直接上游id集合 Collection<Integer> inputIds = transform(transform.getInput()); // 递归调用可能已经处理过该transform了 if (alreadyTransformed.containsKey(transform)) { return alreadyTransformed.get(transform); } String slotSharingGroup = determineSlotSharingGroup(transform.getSlotSharingGroup(), inputIds); // 添加 StreamNode streamGraph.addOperator(transform.getId(), slotSharingGroup, transform.getOperator(), transform.getInputType(), transform.getOutputType(), transform.getName()); if (transform.getStateKeySelector() != null) { TypeSerializer<?> keySerializer = transform.getStateKeyType().createSerializer(env.getConfig()); streamGraph.setOneInputStateKey(transform.getId(), transform.getStateKeySelector(), keySerializer); } streamGraph.setParallelism(transform.getId(), transform.getParallelism()); // 添加 StreamEdge for (Integer inputId: inputIds) { streamGraph.addEdge(inputId, transform.getId(), 0); } return Collections.singleton(transform.getId()); } </code></pre> <p>该函数首先会对该transform的上游transform进行递归转换,确保上游的都已经完成了转化。然后通过transform构造出StreamNode,最后与上游的transform进行连接,构造出StreamNode。</p> <p>最后再来看下对逻辑转换(partition、union等)的处理,如下是 transformPartition 函数的源码:</p> <pre> <code class="language-java">private <T> Collection<Integer> transformPartition(PartitionTransformation<T> partition) { StreamTransformation<T> input = partition.getInput(); List<Integer> resultIds = new ArrayList<>(); // 直接上游的id Collection<Integer> transformedIds = transform(input); for (Integer transformedId: transformedIds) { // 生成一个新的虚拟id int virtualId = StreamTransformation.getNewNodeId(); // 添加一个虚拟分区节点,不会生成 StreamNode streamGraph.addVirtualPartitionNode(transformedId, virtualId, partition.getPartitioner()); resultIds.add(virtualId); } return resultIds; } </code></pre> <p>对partition的转换没有生成具体的StreamNode和StreamEdge,而是添加一个虚节点。当partition的下游transform(如map)添加edge时(调用 StreamGraph.addEdge ),会把partition信息写入到edge中。如 StreamGraph.addEdgeInternal 所示:</p> <pre> <code class="language-java">public void addEdge(Integer upStreamVertexID, Integer downStreamVertexID, int typeNumber) { addEdgeInternal(upStreamVertexID, downStreamVertexID, typeNumber, null, new ArrayList<String>()); } private void addEdgeInternal(Integer upStreamVertexID, Integer downStreamVertexID, int typeNumber, StreamPartitioner<?> partitioner, List<String> outputNames) { // 当上游是select时,递归调用,并传入select信息 if (virtualSelectNodes.containsKey(upStreamVertexID)) { int virtualId = upStreamVertexID; // select上游的节点id upStreamVertexID = virtualSelectNodes.get(virtualId).f0; if (outputNames.isEmpty()) { // selections that happen downstream override earlier selections outputNames = virtualSelectNodes.get(virtualId).f1; } addEdgeInternal(upStreamVertexID, downStreamVertexID, typeNumber, partitioner, outputNames); } // 当上游是partition时,递归调用,并传入partitioner信息 else if (virtuaPartitionNodes.containsKey(upStreamVertexID)) { int virtualId = upStreamVertexID; // partition上游的节点id upStreamVertexID = virtuaPartitionNodes.get(virtualId).f0; if (partitioner == null) { partitioner = virtuaPartitionNodes.get(virtualId).f1; } addEdgeInternal(upStreamVertexID, downStreamVertexID, typeNumber, partitioner, outputNames); } else { // 真正构建StreamEdge StreamNode upstreamNode = getStreamNode(upStreamVertexID); StreamNode downstreamNode = getStreamNode(downStreamVertexID); // 未指定partitioner的话,会为其选择 forward 或 rebalance 分区。 if (partitioner == null && upstreamNode.getParallelism() == downstreamNode.getParallelism()) { partitioner = new ForwardPartitioner<Object>(); } else if (partitioner == null) { partitioner = new RebalancePartitioner<Object>(); } // 健康检查, forward 分区必须要上下游的并发度一致 if (partitioner instanceof ForwardPartitioner) { if (upstreamNode.getParallelism() != downstreamNode.getParallelism()) { throw new UnsupportedOperationException("Forward partitioning does not allow " + "change of parallelism. Upstream operation: " + upstreamNode + " parallelism: " + upstreamNode.getParallelism() + ", downstream operation: " + downstreamNode + " parallelism: " + downstreamNode.getParallelism() + " You must use another partitioning strategy, such as broadcast, rebalance, shuffle or global."); } } // 创建 StreamEdge StreamEdge edge = new StreamEdge(upstreamNode, downstreamNode, typeNumber, outputNames, partitioner); // 将该 StreamEdge 添加到上游的输出,下游的输入 getStreamNode(edge.getSourceId()).addOutEdge(edge); getStreamNode(edge.getTargetId()).addInEdge(edge); } } </code></pre> <h2>总结</h2> <p>本文主要介绍了 Stream API 中 Transformation 和 Operator 的概念,以及如何根据Stream API编写的程序,构造出一个代表拓扑结构的StreamGraph的。本文的源码分析涉及到较多代码,如果有兴趣建议结合完整源码进行学习。下一篇文章将介绍 StreamGraph 如何转换成 JobGraph 的,其中设计到了图优化的技巧。</p> <p>来自: <a href="/misc/goto?guid=4959672280506114231" rel="nofollow">http://wuchong.me/blog/2016/05/04/flink-internal-how-to-build-streamgraph</a></p>