[
https://issues.apache.org/jira/browse/SPARK-19462?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15854534#comment-15854534
]
Ian commented on SPARK-19462:
-----------------------------
I appears that the state mutating of newPartitioning of
org.apache.spark.sql.execution.Exchange is preventing the recalculation when
dynamic shuffle partitioning is enabled.
> when spark.sql.adaptive.enabled is enabled, DF is not resilient to
> node/container failure
> -----------------------------------------------------------------------------------------
>
> Key: SPARK-19462
> URL: https://issues.apache.org/jira/browse/SPARK-19462
> Project: Spark
> Issue Type: Bug
> Components: SQL
> Affects Versions: 1.6.3
> Reporter: Ian
>
> property spark.sql.adaptive.enabled needs to be set "true" for the issue to
> be reproduced.
> reproducible steps using spark-shell
> 0. we use yarn as cluster manager, spark-shell runs in client mode
> 1. launch spark-shell
> 2.
> {code}
> val df1 = sc.parallelize( 1 to 1000, 2).toDF("number")
> df1.registerTempTable("test")
> val data1 = sqlContext.sql("SELECT * FROM test WHERE number > 50")
> data1.collect
> val data2 = sqlContext.sql("SELECT number, count(*) cnt FROM test GROUP BY
> number")
> data2.collect
> // everything is fine up to this point
> // manually kill both the AM and all the NMs of the spark-shell app
> // re-run data1.collect, the result is returned successfully
> data1.collect
> // but data2.collect will fail
> data2.collect
> // stacktrace
> Caused by: java.lang.RuntimeException: Exchange not implemented for
> UnknownPartitioning(1)
> at scala.sys.package$.error(package.scala:27)
> at
> org.apache.spark.sql.execution.Exchange.org$apache$spark$sql$execution$Exchange$$getPartitionKeyExtractor$1(Exchange.scala:198)
> at
> org.apache.spark.sql.execution.Exchange$$anonfun$3.apply(Exchange.scala:208)
> at
> org.apache.spark.sql.execution.Exchange$$anonfun$3.apply(Exchange.scala:207)
> at
> org.apache.spark.rdd.RDD$$anonfun$mapPartitionsInternal$1$$anonfun$apply$21.apply(RDD.scala:728)
> at
> org.apache.spark.rdd.RDD$$anonfun$mapPartitionsInternal$1$$anonfun$apply$21.apply(RDD.scala:728)
> at org.apache.spark.rdd.MapPartitionsRDD.compute(MapPartitionsRDD.scala:38)
> at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:306)
> at org.apache.spark.rdd.RDD.iterator(RDD.scala:270)
> at
> org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:73)
> at
> org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:41)
> at org.apache.spark.scheduler.Task.run(Task.scala:89)
> at org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:227)
> at
> java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
> at
> java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
> at java.lang.Thread.run(Thread.java:745)
> {code}
> The difference between data1 and data2 is whether ShuffledRowRDD is present
> in lineage.
> When the RDD lineage contains ShuffledRowRDD, the above mentioned behavior
> can be observed when node failures or container loss happens.
> {code}
> scala> data2.rdd.toDebugString
> res6: String =
> (1) MapPartitionsRDD[20] at rdd at <console>:26 []
> | MapPartitionsRDD[19] at rdd at <console>:26 []
> | ShuffledRowRDD[8] at collect at <console>:26 []
> +-(2) MapPartitionsRDD[7] at collect at <console>:26 []
> | MapPartitionsRDD[6] at collect at <console>:26 []
> | MapPartitionsRDD[5] at collect at <console>:26 []
> | MapPartitionsRDD[1] at intRddToDataFrameHolder at <console>:25 []
> | ParallelCollectionRDD[0] at parallelize at <console>:25 []
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
