在Spark Scala中使用自定义数据帧类时任务不可序列化

Question

我在Scala/Spark（1.5）和齐柏林飞艇上遇到了一个奇怪的问题：

如果我运行以下Scala/Spark代码，它将正常运行：

// TEST NO PROBLEM SERIALIZATION
val rdd = sc.parallelize(Seq(1, 2, 3))
val testList = List[String]("a", "b")

rdd.map{a => 
    val aa = testList(0)
    None}

但是，在声明了此处建议的自定义数据帧类型之后

//DATAFRAME EXTENSION
import org.apache.spark.sql.DataFrame

object ExtraDataFrameOperations {
  implicit class DFWithExtraOperations(df : DataFrame) {

    //drop several columns
    def drop(colToDrop:Seq[String]):DataFrame = {
        var df_temp = df
        colToDrop.foreach{ case (f: String) =>
            df_temp = df_temp.drop(f)//can be improved with Spark 2.0
        }
        df_temp
    }   
  }
}

使用它的例子如下：

//READ ALL THE FILES INTO different DF and save into map
import ExtraDataFrameOperations._
val filename = "myInput.csv"

val delimiter =  ","

val colToIgnore = Seq("c_9", "c_10")

val inputICFfolder = "hdfs:///group/project/TestSpark/"

val df = sqlContext.read
            .format("com.databricks.spark.csv")
            .option("header", "true") // Use first line of all files as header
            .option("inferSchema", "false") // Automatically infer data types? => no cause we need to merge all df, with potential null values => keep string only
            .option("delimiter", delimiter)
            .option("charset", "UTF-8")
            .load(inputICFfolder + filename)
            .drop(colToIgnore)//call the customize dataframe

这次运行成功。

现在如果我再次运行下面的代码（同上）

// TEST NO PROBLEM SERIALIZATION
val rdd = sc.parallelize(Seq(1, 2, 3))
val testList = List[String]("a", "b")
rdd.map{a => 
    val aa = testList(0)
    None}

我收到了错误信息：

rdd:org。阿帕奇。火花rdd。RDD[Int]=ParallelCollectionRDD[8]位于parallelize at:32 testList:List[String]=List（a，b）org。阿帕奇。火花SparkException:任务在组织中不可序列化。阿帕奇。火花util。ClosureCleaner美元。EnsureCleaner.scala:304）可在org。阿帕奇。火花util。ClosureCleaner美元。org$apache$spark$util$ClosureCleaner$$clean（ClosureCleaner.scala:294）位于org。阿帕奇。火花util。ClosureCleaner美元。clean（ClosureCleaner.scala:122）在org上。阿帕奇。火花SparkContext。clean（SparkContext.scala:2032）位于org。阿帕奇。火花rdd。RDD$$anonfun$map$1。应用（RDD.scala:314）。。。原因：java。伊奥。NotSerializableException:$iwC$$iwC$$iwC$$iwC$$iwC$$iwC$$iwC$$ExtraDataFrameOperations$序列化堆栈：-对象不可序列化（类：$iwC$$iwC$$iwC$$iwC$$iwC$$iwC$$iwC$$iwC$$iwC$$ExtraDataFrameOperations$，值：$iwC$$iwC$$iwC$$ExtraframeOperations$$6E70E）-字段（类：$iwC$$iwC$$iwC$$iwC$$iwC$$iwC$$iwC，名称：ExtraDataFrameOperations$模块，类型：类$iwC$$iwC$$iwC$$iwC$$iwC$$iwC$$iwC$$iwC$$iwC$$ExtraDataFrameOperations$）-object（类$iwC$$iwC$$iwC$$iwC$$iwC$$iwC$$iwC$$iwC$$iwC$$iwC$$$$iwC@4c6d0802)-字段（类：$iwC$$iwC$$iwC$$iwC$$iwC$$iwC$$iwC，名称：$iw，类型：类$iwC$$iwC$$iwC$$iwC$$iwC$$iwC$$iwC$$iwC）。。。

我不明白：

• 为什么在没有对数据帧执行任何操作时发生此错误
• 为什么“ExtraDataFrameOperations”在以前成功使用时不可序列化

更新：

试图与

@inline val testList = List[String]("a", "b")

没有帮助。

Answer 1

看起来spark试图序列化testList周围的所有范围。尝试内联数据@inline val testList=List[String]（“a”、“b”），或者使用不同的对象存储传递给驱动程序的函数/数据。

Answer 2

只需为我添加“扩展可序列化”这项工作

/**
   * A wrapper around ProducerRecord RDD that allows to save RDD to Kafka.
   *
   * KafkaProducer is shared within all threads in one executor.
   * Error handling strategy - remember "last" seen exception and rethrow it to allow task fail.
   */
 implicit class DatasetKafkaSink(ds: Dataset[ProducerRecord[String, GenericRecord]]) extends Serializable {

   class ExceptionRegisteringCallback extends Callback {
     private[this] val lastRegisteredException = new AtomicReference[Option[Exception]](None)

     override def onCompletion(metadata: RecordMetadata, exception: Exception): Unit = {
       Option(exception) match {
         case a @ Some(_) => lastRegisteredException.set(a) // (re)-register exception if send failed
         case _ => // do nothing if encountered successful send
       }
     }

     def rethrowException(): Unit = lastRegisteredException.getAndSet(None).foreach(e => throw e)
   }

   /**
     * Save to Kafka reusing KafkaProducer from singleton holder.
     * Returns back control only once all records were actually sent to Kafka, in case of error rethrows "last" seen
     * exception in the same thread to allow Spark task to fail
     */
   def saveToKafka(kafkaProducerConfigs: Map[String, AnyRef]): Unit = {
     ds.foreachPartition { records =>
       val callback = new ExceptionRegisteringCallback
       val producer = KafkaProducerHolder.getInstance(kafkaProducerConfigs)

       records.foreach(record => producer.send(record, callback))

       producer.flush()
       callback.rethrowException()
     }
   }
 }'