[
https://issues.apache.org/jira/browse/SPARK-3032?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14144243#comment-14144243
]
Matei Zaharia commented on SPARK-3032:
--------------------------------------
I'm not completely sure that this is because hashCode provides a partial
ordering, because I believe TimSort is supposed to work on partial orderings as
well. I believe the problem is an integer over flow when we subtract
key1.hashCode - key2.hashCode. Can you try replacing the line that returns h1 -
h2 in keyComparator with returning Integer.compare(h1, h2)? This will properly
deal with overflow.
Returning h1 - h2 is definitely wrong: for example suppose that h1 =
Int.MaxValue and h2 = Int.MinValue, then h1 - h2 = -1.
Please add a unit test for this case as well.
> Potential bug when running sort-based shuffle with sorting using TimSort
> ------------------------------------------------------------------------
>
> Key: SPARK-3032
> URL: https://issues.apache.org/jira/browse/SPARK-3032
> Project: Spark
> Issue Type: Bug
> Components: Shuffle
> Affects Versions: 1.1.0
> Reporter: Saisai Shao
> Priority: Critical
>
> When using SparkPerf's aggregate-by-key workload to test sort-based shuffle,
> data type for key and value is (String, String), always meet this issue:
> {noformat}
> java.lang.IllegalArgumentException: Comparison method violates its general
> contract!
> at
> org.apache.spark.util.collection.Sorter$SortState.mergeLo(Sorter.java:755)
> at
> org.apache.spark.util.collection.Sorter$SortState.mergeAt(Sorter.java:493)
> at
> org.apache.spark.util.collection.Sorter$SortState.mergeCollapse(Sorter.java:420)
> at
> org.apache.spark.util.collection.Sorter$SortState.access$200(Sorter.java:294)
> at org.apache.spark.util.collection.Sorter.sort(Sorter.java:128)
> at
> org.apache.spark.util.collection.SizeTrackingPairBuffer.destructiveSortedIterator(SizeTrackingPairBuffer.scala:83)
> at
> org.apache.spark.util.collection.ExternalSorter.spillToMergeableFile(ExternalSorter.scala:323)
> at
> org.apache.spark.util.collection.ExternalSorter.spill(ExternalSorter.scala:271)
> at
> org.apache.spark.util.collection.ExternalSorter.maybeSpill(ExternalSorter.scala:249)
> at
> org.apache.spark.util.collection.ExternalSorter.insertAll(ExternalSorter.scala:220)
> at
> org.apache.spark.shuffle.sort.SortShuffleWriter.write(SortShuffleWriter.scala:85)
> at
> org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:68)
> at
> org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:41)
> at org.apache.spark.scheduler.Task.run(Task.scala:54)
> at
> org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:199)
> at
> java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1110)
> at
> java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:603)
> at java.lang.Thread.run(Thread.java:722)
> {noformat}
> Seems the current partitionKeyComparator which use hashcode of String as key
> comparator break some sorting contracts.
> Also I tested using data type Int as key, this is OK to pass the test, since
> hashcode of Int is its self. So I think potentially partitionDiff + hashcode
> of String may break the sorting contracts.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
