为什么 DataFrame.stat.approxQuantile 失败，n 个任务的序列化结果大小 (1030.8 MB) 大于 spark.driver.maxResultSize

Question

val postsQuantiles = posts.stat.approxQuantile("_score", Array(0.25, 0.75), 0.0)因以下错误而失败。我显然可以设置spark.driver.maxResultSize克服这个错误，但我很好奇为什么这会向驱动程序收集数据？

[Stage 3:==================>                                      (7 + 15) / 22]19/06/01 20:46:30 ERROR TaskSetManager: Total size of serialized results of 18 tasks (1030.8 MB) is bigger than spark.driver.maxResultSize (1024.0 MB)
19/06/01 20:46:30 ERROR TaskSetManager: Total size of serialized results of 19 tasks (1087.7 MB) is bigger than spark.driver.maxResultSize (1024.0 MB)
19/06/01 20:46:30 ERROR TaskSetManager: Total size of serialized results of 20 tasks (1145.6 MB) is bigger than spark.driver.maxResultSize (1024.0 MB)
19/06/01 20:46:30 ERROR TaskSetManager: Total size of serialized results of 21 tasks (1203.5 MB) is bigger than spark.driver.maxResultSize (1024.0 MB)
19/06/01 20:46:30 ERROR TaskSetManager: Total size of serialized results of 22 tasks (1261.4 MB) is bigger than spark.driver.maxResultSize (1024.0 MB)
[Stage 3:====================================>                    (14 + 8) / 22]org.apache.spark.SparkException: Job aborted due to stage failure: Total size of serialized results of 18 tasks (1030.8 MB) is bigger than spark.driver.maxResultSize (1024.0 MB)
  at org.apache.spark.scheduler.DAGScheduler.org$apache$spark$scheduler$DAGScheduler$$failJobAndIndependentStages(DAGScheduler.scala:1599)
  at org.apache.spark.scheduler.DAGScheduler$$anonfun$abortStage$1.apply(DAGScheduler.scala:1587)
  at org.apache.spark.scheduler.DAGScheduler$$anonfun$abortStage$1.apply(DAGScheduler.scala:1586)
  at scala.collection.mutable.ResizableArray$class.foreach(ResizableArray.scala:59)
  at scala.collection.mutable.ArrayBuffer.foreach(ArrayBuffer.scala:48)
  at org.apache.spark.scheduler.DAGScheduler.abortStage(DAGScheduler.scala:1586)
  at org.apache.spark.scheduler.DAGScheduler$$anonfun$handleTaskSetFailed$1.apply(DAGScheduler.scala:831)
  at org.apache.spark.scheduler.DAGScheduler$$anonfun$handleTaskSetFailed$1.apply(DAGScheduler.scala:831)
  at scala.Option.foreach(Option.scala:257)
  at org.apache.spark.scheduler.DAGScheduler.handleTaskSetFailed(DAGScheduler.scala:831)
  at org.apache.spark.scheduler.DAGSchedulerEventProcessLoop.doOnReceive(DAGScheduler.scala:1820)
  at org.apache.spark.scheduler.DAGSchedulerEventProcessLoop.onReceive(DAGScheduler.scala:1769)
  at org.apache.spark.scheduler.DAGSchedulerEventProcessLoop.onReceive(DAGScheduler.scala:1758)
  at org.apache.spark.util.EventLoop$$anon$1.run(EventLoop.scala:48)
  at org.apache.spark.scheduler.DAGScheduler.runJob(DAGScheduler.scala:642)
  at org.apache.spark.SparkContext.runJob(SparkContext.scala:2034)
  at org.apache.spark.SparkContext.runJob(SparkContext.scala:2131)
  at org.apache.spark.rdd.RDD$$anonfun$fold$1.apply(RDD.scala:1092)
  at org.apache.spark.rdd.RDDOperationScope$.withScope(RDDOperationScope.scala:151)
  at org.apache.spark.rdd.RDDOperationScope$.withScope(RDDOperationScope.scala:112)
  at org.apache.spark.rdd.RDD.withScope(RDD.scala:363)
  at org.apache.spark.rdd.RDD.fold(RDD.scala:1086)
  at org.apache.spark.rdd.RDD$$anonfun$treeAggregate$1.apply(RDD.scala:1155)
  at org.apache.spark.rdd.RDDOperationScope$.withScope(RDDOperationScope.scala:151)
  at org.apache.spark.rdd.RDDOperationScope$.withScope(RDDOperationScope.scala:112)
  at org.apache.spark.rdd.RDD.withScope(RDD.scala:363)
  at org.apache.spark.rdd.RDD.treeAggregate(RDD.scala:1131)
  at org.apache.spark.sql.execution.stat.StatFunctions$.multipleApproxQuantiles(StatFunctions.scala:102)
  at org.apache.spark.sql.DataFrameStatFunctions.approxQuantile(DataFrameStatFunctions.scala:100)
  at org.apache.spark.sql.DataFrameStatFunctions.approxQuantile(DataFrameStatFunctions.scala:75)
  ... 56 elided

Answer 1

该approxQuantile方法遵循 Greenwald-Khanna 算法来计算近似分位数（基于他们的文档https://spark.apache.org/docs/latest/api/scala/index.html#org.apache.spark.sql.DataFrameStatFunctions）。它允许您选择相对误差项。

在他们警告您的文档中，选择相对误差0.0（如您所见）可能非常昂贵，这正是您所看到的。该算法比直接分位数更适合近似分位数。之所以要提取这么多数据，是因为要计算直接分位数，它至少需要将列中的所有数据提取到驱动程序中。

您可以从已发表的论文中阅读有关该算法的更多信息：http : //infolab.stanford.edu/~datar/courses/cs361a/papers/quantiles.pdf

为了克服这个问题，我建议使用一些适当的小值的相对误差项，这给你“足够接近”的信心。