《Flink官方文档》Batch Examples

原文链接

批处理示例

下面的程序展示了从简单的单词词频统计到图算法等不同的Flink应用。代码展示了Flink数据集API的使用方法。

下面案例和更多案例的完整源码可以参见Flink源码中的flink-examples-batch和 flink-examples-streaming模块。

运行实例

为了运行Flink的例子，我们假设你拥有已经启动的Flink实例。在导航栏中的“Quickstart” 和 “Setup”介绍了启动Flink的几种不同方法。

最简单的方法是运行脚本./bin/start-local.sh，执行后一个启动本地JobManager。

每个编译好的Flink源码包含了一个实例目录，其中包括了此页面每个例子的jar包。

执行如下命令，来运行WordCount例子

./bin/flink run ./examples/batch/WordCount.jar
其他的例子都可以用类似的方式执行

如果运行例子的时候没有带参数，默认使用缺省参数。如果希望使用真实数据来运行WordCount,需要将数据的路径传递进去

./bin/flink run ./examples/batch/WordCount.jar –input /path/to/some/text/data –output /path/to/result
注意非本地文件系统需要标明数据库前缀，比如HDFS：//

词频统计

单词词频统计是大数据处理系统“hello world”程序。它计算了文本中的词频。算法分成两步，第一部分，将文本分隔成不同单词。第二步，讲这些单词分组并计数。

ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();

DataSet<String> text = env.readTextFile("/path/to/file");

DataSet<Tuple2<String, Integer>> counts =
        // split up the lines in pairs (2-tuples) containing: (word,1)
        text.flatMap(new Tokenizer())
        // group by the tuple field "0" and sum up tuple field "1"
        .groupBy(0)
        .sum(1);

counts.writeAsCsv(outputPath, "\n", " ");

// User-defined functions
public static class Tokenizer implements FlatMapFunction<String, Tuple2<String, Integer>> {

    @Override
    public void flatMap(String value, Collector<Tuple2<String, Integer>> out) {
        // normalize and split the line
        String[] tokens = value.toLowerCase().split("\\W+");

        // emit the pairs
        for (String token : tokens) {
            if (token.length() > 0) {
                out.collect(new Tuple2<String, Integer>(token, 1));
            }   
        }
    }
}

词频统计例子实现了上述描述的算法，需要两个输入参数。–input –output 。测试数据可以替换为任何文本。

val env = ExecutionEnvironment.getExecutionEnvironment

// get input data
val text = env.readTextFile("/path/to/file")

val counts = text.flatMap { _.toLowerCase.split("\\W+") filter { _.nonEmpty } }
  .map { (_, 1) }
  .groupBy(0)
  .sum(1)

counts.writeAsCsv(outputPath, "\n", " ")

Page Rank

PageRank算法计算了图中页面的重要性，一个页面到另一页面的点形成了链接，这些链接定义成图。它是迭代式的算法，意味着相同的计算会被重复执行。在每次迭代中，每个页面对它的邻居贡献出相同的评分，并接受来自它的邻居的加权评分作为新的评分。PageRank算法因google搜索引擎众所周知，它被用来计算网页搜索查询结果的评分。

这个例子中，PageRank通过一批迭代和固定次数的迭代来完成。

ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();

// read the pages and initial ranks by parsing a CSV file
DataSet<Tuple2<Long, Double>> pagesWithRanks = env.readCsvFile(pagesInputPath)
                           .types(Long.class, Double.class)

// the links are encoded as an adjacency list: (page-id, Array(neighbor-ids))
DataSet<Tuple2<Long, Long[]>> pageLinkLists = getLinksDataSet(env);

// set iterative data set
IterativeDataSet<Tuple2<Long, Double>> iteration = pagesWithRanks.iterate(maxIterations);

DataSet<Tuple2<Long, Double>> newRanks = iteration
        // join pages with outgoing edges and distribute rank
        .join(pageLinkLists).where(0).equalTo(0).flatMap(new JoinVertexWithEdgesMatch())
        // collect and sum ranks
        .groupBy(0).sum(1)
        // apply dampening factor
        .map(new Dampener(DAMPENING_FACTOR, numPages));

DataSet<Tuple2<Long, Double>> finalPageRanks = iteration.closeWith(
        newRanks,
        newRanks.join(iteration).where(0).equalTo(0)
        // termination condition
        .filter(new EpsilonFilter()));

finalPageRanks.writeAsCsv(outputPath, "\n", " ");

// User-defined functions

public static final class JoinVertexWithEdgesMatch
                    implements FlatJoinFunction<Tuple2<Long, Double>, Tuple2<Long, Long[]>,
                                            Tuple2<Long, Double>> {

    @Override
    public void join(<Tuple2<Long, Double> page, Tuple2<Long, Long[]> adj,
                        Collector<Tuple2<Long, Double>> out) {
        Long[] neighbors = adj.f1;
        double rank = page.f1;
        double rankToDistribute = rank / ((double) neigbors.length);

        for (int i = 0; i < neighbors.length; i++) {
            out.collect(new Tuple2<Long, Double>(neighbors[i], rankToDistribute));
        }
    }
}

public static final class Dampener implements MapFunction<Tuple2<Long,Double>, Tuple2<Long,Double>> {
    private final double dampening, randomJump;

    public Dampener(double dampening, double numVertices) {
        this.dampening = dampening;
        this.randomJump = (1 - dampening) / numVertices;
    }

    @Override
    public Tuple2<Long, Double> map(Tuple2<Long, Double> value) {
        value.f1 = (value.f1 * dampening) + randomJump;
        return value;
    }
}

public static final class EpsilonFilter
                implements FilterFunction<Tuple2<Tuple2<Long, Double>, Tuple2<Long, Double>>> {

    @Override
    public boolean filter(Tuple2<Tuple2<Long, Double>, Tuple2<Long, Double>> value) {
        return Math.abs(value.f0.f1 - value.f1.f1) > EPSILON;
    }
}

pagerank 程序实现了上面的例子。需要下面的运行参数–pages –links –output –numPages –iterations 。

scala

// User-defined types
case class Link(sourceId: Long, targetId: Long)
case class Page(pageId: Long, rank: Double)
case class AdjacencyList(sourceId: Long, targetIds: Array[Long])

// set up execution environment
val env = ExecutionEnvironment.getExecutionEnvironment

// read the pages and initial ranks by parsing a CSV file
val pages = env.readCsvFile[Page](pagesInputPath)

// the links are encoded as an adjacency list: (page-id, Array(neighbor-ids))
val links = env.readCsvFile[Link](linksInputPath)

// assign initial ranks to pages
val pagesWithRanks = pages.map(p => Page(p, 1.0 / numPages))

// build adjacency list from link input
val adjacencyLists = links
  // initialize lists
  .map(e => AdjacencyList(e.sourceId, Array(e.targetId)))
  // concatenate lists
  .groupBy("sourceId").reduce {
  (l1, l2) => AdjacencyList(l1.sourceId, l1.targetIds ++ l2.targetIds)
  }

// start iteration
val finalRanks = pagesWithRanks.iterateWithTermination(maxIterations) {
  currentRanks =>
    val newRanks = currentRanks
      // distribute ranks to target pages
      .join(adjacencyLists).where("pageId").equalTo("sourceId") {
        (page, adjacent, out: Collector[Page]) =>
        for (targetId <- adjacent.targetIds) {
          out.collect(Page(targetId, page.rank / adjacent.targetIds.length))
        }
      }
      // collect ranks and sum them up
      .groupBy("pageId").aggregate(SUM, "rank")
      // apply dampening factor
      .map { p =>
        Page(p.pageId, (p.rank * DAMPENING_FACTOR) + ((1 - DAMPENING_FACTOR) / numPages))
      }

    // terminate if no rank update was significant
    val termination = currentRanks.join(newRanks).where("pageId").equalTo("pageId") {
      (current, next, out: Collector[Int]) =>
        // check for significant update
        if (math.abs(current.rank - next.rank) > EPSILON) out.collect(1)
    }

    (newRanks, termination)
}

val result = finalRanks

// emit result
result.writeAsCsv(outputPath, "\n", " ")

输入文件必须是普通文本文件而且文件必须是遵循下列格式：

–Pages 用long型的ID表示，并以换行符分隔，如”1\n2\n12\n42\n63\n”体现了5个页面，id分别是1, 2, 12, 42, and 63。

–Links表示了多对pageId的组合，每对之间通过空格分隔，不同links用换行符分隔。”1 2\n2 12\n1 12\n42 63\n”表示了(1)->(2), (2)->(12), (1)->(12), and (42)->(63)四个有向链接。

为了这个简单实现至少需要每个页面至少有一个入链接和一个出链接。一个页面可以链接到他自己。

连通分支

连通分支算法识别会一个更大的图，这部分图通过被相同的组件ID链接的所有顶点连接。类似PageRank，连通组件是一个迭代算法。在每个步骤中，每个顶点都将其当前组件ID传给所有邻居。如果小于自己的组件ID，一个顶点从邻居接受组件ID。

此实现使用增量迭代：组件ID未变化的顶点不参与下一步骤。因为后来的迭代通常只处理一些离群顶点，这将产生更好的性能。

// read vertex and edge data
DataSet<Long> vertices = getVertexDataSet(env);
DataSet<Tuple2<Long, Long>> edges = getEdgeDataSet(env).flatMap(new UndirectEdge());

// assign the initial component IDs (equal to the vertex ID)
DataSet<Tuple2<Long, Long>> verticesWithInitialId = vertices.map(new DuplicateValue<Long>());

// open a delta iteration
DeltaIteration<Tuple2<Long, Long>, Tuple2<Long, Long>> iteration =
        verticesWithInitialId.iterateDelta(verticesWithInitialId, maxIterations, 0);

// apply the step logic:
DataSet<Tuple2<Long, Long>> changes = iteration.getWorkset()
        // join with the edges
        .join(edges).where(0).equalTo(0).with(new NeighborWithComponentIDJoin())
        // select the minimum neighbor component ID
        .groupBy(0).aggregate(Aggregations.MIN, 1)
        // update if the component ID of the candidate is smaller
        .join(iteration.getSolutionSet()).where(0).equalTo(0)
        .flatMap(new ComponentIdFilter());

// close the delta iteration (delta and new workset are identical)
DataSet<Tuple2<Long, Long>> result = iteration.closeWith(changes, changes);

// emit result
result.writeAsCsv(outputPath, "\n", " ");

// User-defined functions

public static final class DuplicateValue<T> implements MapFunction<T, Tuple2<T, T>> {

    @Override
    public Tuple2<T, T> map(T vertex) {
        return new Tuple2<T, T>(vertex, vertex);
    }
}

public static final class UndirectEdge
                    implements FlatMapFunction<Tuple2<Long, Long>, Tuple2<Long, Long>> {
    Tuple2<Long, Long> invertedEdge = new Tuple2<Long, Long>();

    @Override
    public void flatMap(Tuple2<Long, Long> edge, Collector<Tuple2<Long, Long>> out) {
        invertedEdge.f0 = edge.f1;
        invertedEdge.f1 = edge.f0;
        out.collect(edge);
        out.collect(invertedEdge);
    }
}

public static final class NeighborWithComponentIDJoin
                implements JoinFunction<Tuple2<Long, Long>, Tuple2<Long, Long>, Tuple2<Long, Long>> {

    @Override
    public Tuple2<Long, Long> join(Tuple2<Long, Long> vertexWithComponent, Tuple2<Long, Long> edge) {
        return new Tuple2<Long, Long>(edge.f1, vertexWithComponent.f1);
    }
}

public static final class ComponentIdFilter
                    implements FlatMapFunction<Tuple2<Tuple2<Long, Long>, Tuple2<Long, Long>>,
                                            Tuple2<Long, Long>> {

    @Override
    public void flatMap(Tuple2<Tuple2<Long, Long>, Tuple2<Long, Long>> value,
                        Collector<Tuple2<Long, Long>> out) {
        if (value.f0.f1 < value.f1.f1) {
            out.collect(value.f0);
        }
    }
}

scala

// set up execution environment
val env = ExecutionEnvironment.getExecutionEnvironment

// read vertex and edge data
// assign the initial components (equal to the vertex id)
val vertices = getVerticesDataSet(env).map { id => (id, id) }

// undirected edges by emitting for each input edge the input edges itself and an inverted
// version
val edges = getEdgesDataSet(env).flatMap { edge => Seq(edge, (edge._2, edge._1)) }

// open a delta iteration
val verticesWithComponents = vertices.iterateDelta(vertices, maxIterations, Array(0)) {
  (s, ws) =>

    // apply the step logic: join with the edges
    val allNeighbors = ws.join(edges).where(0).equalTo(0) { (vertex, edge) =>
      (edge._2, vertex._2)
    }

    // select the minimum neighbor
    val minNeighbors = allNeighbors.groupBy(0).min(1)

    // update if the component of the candidate is smaller
    val updatedComponents = minNeighbors.join(s).where(0).equalTo(0) {
      (newVertex, oldVertex, out: Collector[(Long, Long)]) =>
        if (newVertex._2 < oldVertex._2) out.collect(newVertex)
    }

    // delta and new workset are identical
    (updatedComponents, updatedComponents)
}


verticesWithComponents.writeAsCsv(outputPath, "\n", " ")

该连通分支程序实现了上述例子。它需要运行下列参数：–vertices –edges –output –iterations 。
输入文件是纯文本文件，必须格式化如下：

–Vertices 以IDS表示的顶点，由换行字符分隔。例如“1\n2\n12\n42\n63\n”给出了五个订单（1）、（2）、（12）、（42）和（63）。

–Edges 边通过以空格分隔的两个顶点ID表示。不同边是由换行符分隔。例如“1 2\n2 12\n1 12\n42 63\n”表示了四个无方向链接（1）-（2）、（2）-（12）、（1）-（12）和（42）-（63）。

关系型查询

关系型查询示例假定会使用两张表，一张订单表，另一张是TPC-H决策支持基准测试表。TPC-H是数据库行业标准基准测试。如何生成输入数据请参见下面的说明。

该示例实现以下sql查询。
SELECT l_orderkey, o_shippriority, sum(l_extendedprice) as revenue
FROM orders, lineitem
WHERE l_orderkey = o_orderkey
AND o_orderstatus = "F"
AND YEAR(o_orderdate) > 1993
AND o_orderpriority LIKE "5%"
GROUP BY l_orderkey, o_shippriority;
Flink程序中按照如下的方式进行sql查询

// get orders data set: (orderkey, orderstatus, orderdate, orderpriority, shippriority)
DataSet<Tuple5<Integer, String, String, String, Integer>> orders = getOrdersDataSet(env);
// get lineitem data set: (orderkey, extendedprice)
DataSet<Tuple2<Integer, Double>> lineitems = getLineitemDataSet(env);

// orders filtered by year: (orderkey, custkey)
DataSet<Tuple2<Integer, Integer>> ordersFilteredByYear =
        // filter orders
        orders.filter(
            new FilterFunction<Tuple5<Integer, String, String, String, Integer>>() {
                @Override
                public boolean filter(Tuple5<Integer, String, String, String, Integer> t) {
                    // status filter
                    if(!t.f1.equals(STATUS_FILTER)) {
                        return false;
                    // year filter
                    } else if(Integer.parseInt(t.f2.substring(0, 4)) <= YEAR_FILTER) {
                        return false;
                    // order priority filter
                    } else if(!t.f3.startsWith(OPRIO_FILTER)) {
                        return false;
                    }
                    return true;
                }
            })
        // project fields out that are no longer required
        .project(0,4).types(Integer.class, Integer.class);

// join orders with lineitems: (orderkey, shippriority, extendedprice)
DataSet<Tuple3<Integer, Integer, Double>> lineitemsOfOrders =
        ordersFilteredByYear.joinWithHuge(lineitems)
                            .where(0).equalTo(0)
                            .projectFirst(0,1).projectSecond(1)
                            .types(Integer.class, Integer.class, Double.class);

// extendedprice sums: (orderkey, shippriority, sum(extendedprice))
DataSet<Tuple3<Integer, Integer, Double>> priceSums =
        // group by order and sum extendedprice
        lineitemsOfOrders.groupBy(0,1).aggregate(Aggregations.SUM, 2);

// emit result
priceSums.writeAsCsv(outputPath);

缺少scala例子（译者注）

关系查询程序实现了上述查询。它需要以下参数运行–orders –lineitem –output 。
order和lineitem文件可以使用TPC-H基准测试套件的数据生成工具（DBGEN）生成。采取以下步骤生成需提供给flink程序输入的任意大小的数据文件。

1、下载并解压DBGEN

2、复制makefile.suite并更名为Makefile，编辑修改如下：

DATABASE = DB2
MACHINE  = LINUX
WORKLOAD = TPCH
CC       = gcc

1、使用make命令构建DBGEN

2、使用DBGEN生成lineitem和orders表。-s命令传入1，将会一个生成约1 GB的大小的数据集。

./dbgen -T o -s 1


原创文章，转载请注明： 转载自并发编程网 – ifeve.com本文链接地址: 《Flink官方文档》Batch Examples

  • Trackback 关闭
  • 评论 (0)

  1. 暂无评论

return top