Question

Spark streaming textFileStream和fileStream可以监控目录并处理Dstream RDD中的新文件。

如何在特定时间间隔内获取DStream RDD正在处理的文件名？

Answer 1

fileStream生成UnionRDD个NewHadoopRDD个。 NewHadoopRDD创建的关于sc.newAPIHadoopFile的好处是他们的name被设置为路径。

以下是您可以使用该知识做的例子：

def namedTextFileStream(ssc: StreamingContext, directory: String): DStream[String] =
  ssc.fileStream[LongWritable, Text, TextInputFormat](directory)
    .transform( rdd =>
      new UnionRDD(rdd.context,
        rdd.dependencies.map( dep =>
          dep.rdd.asInstanceOf[RDD[(LongWritable, Text)]].map(_._2.toString).setName(dep.rdd.name)
        )
      )
    )

def transformByFile[U: ClassTag](unionrdd: RDD[String],
                                 transformFunc: String => RDD[String] => RDD[U]): RDD[U] = {
  new UnionRDD(unionrdd.context,
    unionrdd.dependencies.map{ dep =>
      if (dep.rdd.isEmpty) None
      else {
        val filename = dep.rdd.name
        Some(
          transformFunc(filename)(dep.rdd.asInstanceOf[RDD[String]])
            .setName(filename)
        )
      }
    }.flatten
  )
}

def main(args: Array[String]) = {
  val conf = new SparkConf()
    .setAppName("Process by file")
    .setMaster("local[2]")

  val ssc = new StreamingContext(conf, Seconds(30))

  val dstream = namesTextFileStream(ssc, "/some/directory")

  def byFileTransformer(filename: String)(rdd: RDD[String]): RDD[(String, String)] =
    rdd.map(line => (filename, line))

  val transformed = dstream.
    transform(rdd => transformByFile(rdd, byFileTransformer))

  // Do some stuff with transformed

  ssc.start()
  ssc.awaitTermination()
}

Answer 2

对于那些需要一些Java代码而不是Scala的人：

JavaPairInputDStream<LongWritable, Text> textFileStream = 
        jsc.fileStream(
            inputPath, 
            LongWritable.class, 
            Text.class,
            TextInputFormat.class, 
            FileInputDStream::defaultFilter,
            false
        );
JavaDStream<Tuple2<String, String>> namedTextFileStream = textFileStream.transform((pairRdd, time) -> {
        UnionRDD<Tuple2<LongWritable, Text>> rdd = (UnionRDD<Tuple2<LongWritable, Text>>) pairRdd.rdd();
        List<RDD<Tuple2<LongWritable, Text>>> deps = JavaConverters.seqAsJavaListConverter(rdd.rdds()).asJava();
        List<RDD<Tuple2<String, String>>> collectedRdds = deps.stream().map( depRdd -> {
            if (depRdd.isEmpty()) {
                return null;
            }
            JavaRDD<Tuple2<LongWritable, Text>> depJavaRdd = depRdd.toJavaRDD();
            String filename = depRdd.name();
            JavaPairRDD<String, String> newDep = JavaPairRDD.fromJavaRDD(depJavaRdd).mapToPair(t -> new Tuple2<String, String>(filename, t._2().toString())).setName(filename);
            return newDep.rdd();
        }).filter(t -> t != null).collect(Collectors.toList());
        Seq<RDD<Tuple2<String, String>>> rddSeq = JavaConverters.asScalaBufferConverter(collectedRdds).asScala().toIndexedSeq();
        ClassTag<Tuple2<String, String>> classTag = scala.reflect.ClassTag$.MODULE$.apply(Tuple2.class);
        return new UnionRDD<Tuple2<String, String>>(rdd.sparkContext(), rddSeq, classTag).toJavaRDD();
});

Answer 3

或者，通过修改FileInputDStream，而不是将文件内容加载到RDD中，而只是从文件名创建RDD。

如果您实际上不想将数据本身读入RDD，或者不希望将文件名传递给外部命令作为步骤之一，则可以提高性能。

只需更改filesToRDD（..），以使其对文件名进行RDD，而不是将数据加载到RDD中。

请参阅：https://github.com/HASTE-project/bin-packing-paper/blob/master/spark/spark-scala-cellprofiler/src/main/scala/FileInputDStream2.scala#L278

Spark流式传输DStream RDD以获取文件名

3 个答案: