关于aggegateByKey的任务无法序列化

Question

环境：火花1.60。 我用scala。 我可以通过sbt编译程序，但是当我提交程序时，遇到了错误。 我的完整错误如下：

238 17/01/21 18:32:24 INFO net.NetworkTopology: Adding a new node: /YH11070029/10.39.0.213:50010
17/01/21 18:32:24 INFO storage.BlockManagerMasterEndpoint: Registering block  manager 10.39.0.44:41961 with 2.7 GB RAM, BlockManagerId(349, 10.39.0.44, 41961)
17/01/21 18:32:24 INFO storage.BlockManagerMasterEndpoint: Registering block manager 10.39.2.178:48591 with 2.7 GB RAM, BlockManagerId(518, 10.39.2.178,  48591)
Exception in thread "main" org.apache.spark.SparkException: Task not     serializable
    at org.apache.spark.util.ClosureCleaner$.ensureSerializable(ClosureCleaner.scala:304)
    at org.apache.spark.util.ClosureCleaner$.org$apache$spark$util$ClosureCleaner$$clean(ClosureCleaner.scala:294)
    at org.apache.spark.util.ClosureCleaner$.clean(ClosureCleaner.scala:122)
    at org.apache.spark.SparkContext.clean(SparkContext.scala:2055)
    at org.apache.spark.rdd.PairRDDFunctions$$anonfun$combineByKeyWithClassTag$1.apply(PairRDDFunctions.scala:93)
    at org.apache.spark.rdd.PairRDDFunctions$$anonfun$combineByKeyWithClassTag$1.apply(PairRDDFunctions.scala:82)
    at org.apache.spark.rdd.RDDOperationScope$.withScope(RDDOperationScope.scala:150)
    at org.apache.spark.rdd.RDDOperationScope$.withScope(RDDOperationScope.scala:111)
    at org.apache.spark.rdd.RDD.withScope(RDD.scala:316)
    at org.apache.spark.rdd.PairRDDFunctions.combineByKeyWithClassTag(PairRDDFunctions.scala:82)
    at org.apache.spark.rdd.PairRDDFunctions$$anonfun$aggregateByKey$1.apply(PairRDDFunctions.scala:177)
    at org.apache.spark.rdd.PairRDDFunctions$$anonfun$aggregateByKey$1.apply(PairRDDFunctions.scala:166)
    at org.apache.spark.rdd.RDDOperationScope$.withScope(RDDOperationScope.scala:150)
    at org.apache.spark.rdd.RDDOperationScope$.withScope(RDDOperationScope.scala:111)
    at org.apache.spark.rdd.RDD.withScope(RDD.scala:316)
    at org.apache.spark.rdd.PairRDDFunctions.aggregateByKey(PairRDDFunctions.scala:166)
    at org.apache.spark.rdd.PairRDDFunctions$$anonfun$aggregateByKey$3.apply(PairRDDFunctions.scala:206)
    at org.apache.spark.rdd.PairRDDFunctions$$anonfun$aggregateByKey$3.apply(PairRDDFunctions.scala:206)
    at org.apache.spark.rdd.RDDOperationScope$.withScope(RDDOperationScope.scala:150)
    at org.apache.spark.rdd.RDDOperationScope$.withScope(RDDOperationScope.scala:111)
    at org.apache.spark.rdd.RDD.withScope(RDD.scala:316)
    at org.apache.spark.rdd.PairRDDFunctions.aggregateByKey(PairRDDFunctions.scala:205)
    at com.sina.adalgo.feature.ETL$$anonfun$13.apply(ETL.scala:190)
    at com.sina.adalgo.feature.ETL$$anonfun$13.apply(ETL.scala:102)
    at scala.collection.TraversableLike$$anonfun$map$1.apply(TraversableLike.scala:244)
    at scala.collection.TraversableLike$$anonfun$map$1.apply(TraversableLike.scala:244)
    at scala.collection.IndexedSeqOptimized$class.foreach(IndexedSeqOptimized.scala:33)

代码的目的是统计分类特征的出现频率。 主要代码如下：

object ETL extends Serializable {
           ... ...


val cateList = featureData.map{v =>
    case (psid: String, label: String, cate_features: ParArray[String], media_features: String) =>
        val pair_feature = cate_features.zipWithIndex.map(x => (x._2, x._1))
        pair_feature
}.flatMap(_.toList)

def seqop(m: HashMap[String, Int] , s: String) : HashMap[String, Int]={
    var x = m.getOrElse(s, 0)
    x += 1
    m += s -> x
    m   
}   

def combop(m: HashMap[String, Int], n: HashMap[String, Int]) : HashMap[String, Int]={
    for (k <- n) {
        var x = m.getOrElse(k._1, 0)
        x += k._2
        m += k._1 -> x
    }   
    m   
}   

val hash = HashMap[String, Int]()
val feaFreq = cateList.aggregateByKey(hash)(seqop, combop)// (i, HashMap[String, Int]) i corresponded with categorical feature

该对象已继承Serializable。 为什么？ 你能帮我吗？

Answer 1

对我来说，当我们使用闭包作为聚合函数时，通常会在Spark中发生此问题，该函数会非预期地关闭一些不需要的对象，并且/或者有时只是在我们的Spark驱动程序代码的主类内部的函数。

我怀疑这里可能是这种情况，因为您的堆栈跟踪涉及org.apache.spark.util.ClosureCleaner作为顶级元凶。

这是有问题的，因为在这种情况下，当Spark尝试将功能转发给工作人员以便他们可以进行实际的聚合时，它最终会比您实际想要的更多地进行序列化：函数itelf及其周围的类。

另请参阅Erik Erlandson的这篇文章，其中很好地解释了一些关于闭包序列化的边界案例，以及关于闭包的Spark 1.6注释 。

一个快速的解决方法可能是将您在aggregateByKey使用的函数的定义移至一个独立的对象，该对象与代码的其余部分完全独立。