[
https://issues.apache.org/jira/browse/SPARK-3633?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14263956#comment-14263956
]
Stephen Haberman commented on SPARK-3633:
-----------------------------------------
Sorry for the latest reply on this, but our skew was a join with a key of 15k
rows one side, and 165k rows on the other, which resulted in .join trying to
create a list with ~2.5 billion entries when joined together (this was before
the recent "use an iterator" fix to .join).
So, yes, this pounded the GC which then showed up as fetch failures to the rest
of our nodes. Apologies for the false negative. It is odd that somehow 0.9.2
powered through this, which is why I thought the job failing in 1.1/1.2 was a
regression. We were probably just getting lucky with having barely enough
memory/something (even though using the same machine sizes).
> Fetches failure observed after SPARK-2711
> -----------------------------------------
>
> Key: SPARK-3633
> URL: https://issues.apache.org/jira/browse/SPARK-3633
> Project: Spark
> Issue Type: Bug
> Components: Block Manager
> Affects Versions: 1.1.0
> Reporter: Nishkam Ravi
> Priority: Blocker
> Fix For: 1.1.1, 1.2.0
>
>
> Running a variant of PageRank on a 6-node cluster with a 30Gb input dataset.
> Recently upgraded to Spark 1.1. The workload fails with the following error
> message(s):
> {code}
> 14/09/19 12:10:38 WARN TaskSetManager: Lost task 51.0 in stage 2.1 (TID 552,
> c1705.halxg.cloudera.com): FetchFailed(BlockManagerId(1,
> c1706.halxg.cloudera.com, 49612, 0), shuffleId=3, mapId=75, reduceId=120)
> 14/09/19 12:10:38 INFO DAGScheduler: Resubmitting failed stages
> {code}
> In order to identify the problem, I carried out change set analysis. As I go
> back in time, the error message changes to:
> {code}
> 14/09/21 12:56:54 WARN TaskSetManager: Lost task 35.0 in stage 3.0 (TID 519,
> c1706.halxg.cloudera.com): java.io.FileNotFoundException:
> /var/lib/jenkins/workspace/tmp/spark-local-20140921123257-68ee/1c/temp_3a1ade13-b48a-437a-a466-673995304034
> (Too many open files)
> java.io.FileOutputStream.open(Native Method)
> java.io.FileOutputStream.<init>(FileOutputStream.java:221)
>
> org.apache.spark.storage.DiskBlockObjectWriter.open(BlockObjectWriter.scala:117)
>
> org.apache.spark.storage.DiskBlockObjectWriter.write(BlockObjectWriter.scala:185)
>
> org.apache.spark.util.collection.ExternalAppendOnlyMap.spill(ExternalAppendOnlyMap.scala:197)
>
> org.apache.spark.util.collection.ExternalAppendOnlyMap.insertAll(ExternalAppendOnlyMap.scala:145)
> org.apache.spark.Aggregator.combineValuesByKey(Aggregator.scala:58)
>
> org.apache.spark.shuffle.hash.HashShuffleWriter.write(HashShuffleWriter.scala:51)
>
> org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:68)
>
> org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:41)
> org.apache.spark.scheduler.Task.run(Task.scala:54)
> org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:199)
>
> java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1145)
>
> java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:615)
> java.lang.Thread.run(Thread.java:745)
> {code}
> All the way until Aug 4th. Turns out the problem changeset is 4fde28c.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
