[
https://issues.apache.org/jira/browse/SPARK-26689?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16752961#comment-16752961
]
liupengcheng commented on SPARK-26689:
--------------------------------------
[~tgraves] In production environment, yarn.nodemanager.local-dirs is always
configured as multiple directories which are mounted on different disks. so I
think since we use this parameter in spark, we should also make use of this
feature, and should not expect job failure when encountering only a single disk
error.
This PR I put up can also reduce the FetchFailure and even Job failure caused
by FetchFailed if blacklist not enabled or node not blacklisted(task may be
repeated scheduled to the unhealthy node)
> Disk broken causing broadcast failure
> -------------------------------------
>
> Key: SPARK-26689
> URL: https://issues.apache.org/jira/browse/SPARK-26689
> Project: Spark
> Issue Type: Bug
> Components: Spark Core
> Affects Versions: 2.1.0, 2.4.0
> Environment: Spark on Yarn
> Mutliple Disk
> Reporter: liupengcheng
> Priority: Major
>
> We encoutered an application failure in our production cluster which caused
> by the bad disk problems. It will incur application failure.
> {code:java}
> Job aborted due to stage failure: Task serialization failed:
> java.io.IOException: Failed to create local dir in
> /home/work/hdd5/yarn/c3prc-hadoop/nodemanager/usercache/h_user_profile/appcache/application_1463372393999_144979/blockmgr-1f96b724-3e16-4c09-8601-1a2e3b758185/3b.
> org.apache.spark.storage.DiskBlockManager.getFile(DiskBlockManager.scala:73)
> org.apache.spark.storage.DiskStore.contains(DiskStore.scala:173)
> org.apache.spark.storage.BlockManager.org$apache$spark$storage$BlockManager$$getCurrentBlockStatus(BlockManager.scala:391)
> org.apache.spark.storage.BlockManager.doPut(BlockManager.scala:801)
> org.apache.spark.storage.BlockManager.putIterator(BlockManager.scala:629)
> org.apache.spark.storage.BlockManager.putSingle(BlockManager.scala:987)
> org.apache.spark.broadcast.TorrentBroadcast.writeBlocks(TorrentBroadcast.scala:99)
> org.apache.spark.broadcast.TorrentBroadcast.<init>(TorrentBroadcast.scala:85)
> org.apache.spark.broadcast.TorrentBroadcastFactory.newBroadcast(TorrentBroadcastFactory.scala:34)
> org.apache.spark.broadcast.BroadcastManager.newBroadcast(BroadcastManager.scala:63)
> org.apache.spark.SparkContext.broadcast(SparkContext.scala:1332)
> org.apache.spark.scheduler.DAGScheduler.org$apache$spark$scheduler$DAGScheduler$$submitMissingTasks(DAGScheduler.scala:863)
> org.apache.spark.scheduler.DAGScheduler$$anonfun$handleTaskCompletion$14$$anonfun$apply$1.apply$mcVI$sp(DAGScheduler.scala:1090)
> org.apache.spark.scheduler.DAGScheduler$$anonfun$handleTaskCompletion$14$$anonfun$apply$1.apply(DAGScheduler.scala:1086)
> org.apache.spark.scheduler.DAGScheduler$$anonfun$handleTaskCompletion$14$$anonfun$apply$1.apply(DAGScheduler.scala:1086)
> scala.Option.foreach(Option.scala:236)
> org.apache.spark.scheduler.DAGScheduler$$anonfun$handleTaskCompletion$14.apply(DAGScheduler.scala:1086)
> org.apache.spark.scheduler.DAGScheduler$$anonfun$handleTaskCompletion$14.apply(DAGScheduler.scala:1085)
> scala.collection.mutable.ResizableArray$class.foreach(ResizableArray.scala:59)
> scala.collection.mutable.ArrayBuffer.foreach(ArrayBuffer.scala:47)
> org.apache.spark.scheduler.DAGScheduler.handleTaskCompletion(DAGScheduler.scala:1085)
> org.apache.spark.scheduler.DAGSchedulerEventProcessLoop.doOnReceive(DAGScheduler.scala:1528)
> org.apache.spark.scheduler.DAGSchedulerEventProcessLoop.onReceive(DAGScheduler.scala:1493)
> org.apache.spark.scheduler.DAGSchedulerEventProcessLoop.onReceive(DAGScheduler.scala:1482)
> org.apache.spark.util.EventLoop$$anon$1.run(EventLoop.scala:48)
> {code}
> We have multiple disk on our cluster nodes, however, it still fails. I think
> it's because spark does not handle bad disk in `DiskBlockManager` currently.
> Actually, we can handle bad disk in multiple disk environment to avoid
> application failure.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
