[SPARK-28340] Noisy exceptions when tasks are killed: "DiskBlockObjectWriter: Uncaught exception while reverting partial writes to file: java.nio.channels.ClosedByInterruptException"

• Type: Improvement
• Status: Resolved
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: 3.0.0
• Fix Version/s: 3.0.0
• Component/s: Spark Core
• Assignee: Colin Ma
• Reporter: Josh Rosen
• Created: 11/Jul/19 00:05
• Updated: 09/Sep/19 10:28
• Resolved: 09/Sep/19 10:28

Description

If a Spark task is killed while writing blocks to disk (due to intentional job kills, automated killing of redundant speculative tasks, etc) then Spark may log exceptions like

19/07/10 21:31:08 ERROR storage.DiskBlockObjectWriter: Uncaught exception while reverting partial writes to file /<FILENAME>
java.nio.channels.ClosedByInterruptException
	at java.nio.channels.spi.AbstractInterruptibleChannel.end(AbstractInterruptibleChannel.java:202)
	at sun.nio.ch.FileChannelImpl.truncate(FileChannelImpl.java:372)
	at org.apache.spark.storage.DiskBlockObjectWriter$$anonfun$revertPartialWritesAndClose$2.apply$mcV$sp(DiskBlockObjectWriter.scala:218)
	at org.apache.spark.util.Utils$.tryWithSafeFinally(Utils.scala:1369)
	at org.apache.spark.storage.DiskBlockObjectWriter.revertPartialWritesAndClose(DiskBlockObjectWriter.scala:214)
	at org.apache.spark.shuffle.sort.BypassMergeSortShuffleWriter.stop(BypassMergeSortShuffleWriter.java:237)
	at org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:105)
	at org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:55)
	at org.apache.spark.scheduler.Task.run(Task.scala:121)
	at org.apache.spark.executor.Executor$TaskRunner$$anonfun$10.apply(Executor.scala:402)
	at org.apache.spark.util.Utils$.tryWithSafeFinally(Utils.scala:1360)
	at org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:408)
	at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149)
	at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624)
	at java.lang.Thread.run(Thread.java:748)

If BypassMergeSortShuffleWriter is being used then a single cancelled task can result in hundreds of these stacktraces being logged.

Here are some StackOverflow questions asking about this:

• https://stackoverflow.com/questions/40027870/spark-jobserver-job-crash
• https://stackoverflow.com/questions/50646953/why-is-java-nio-channels-closedbyinterruptexceptio-called-when-caling-multiple
• https://stackoverflow.com/questions/41867053/java-nio-channels-closedbyinterruptexception-in-spark
• https://stackoverflow.com/questions/56845041/are-closedbyinterruptexception-exceptions-expected-when-spark-speculation-kills 

Can we prevent this exception from occurring? If not, can we treat this "expected exception" in a special manner to avoid log spam? My concern is that the presence of large numbers of spurious exceptions is confusing to users when they are inspecting Spark logs to diagnose other issues.

Issue Links

• relates to: SPARK-23816 FetchFailedException when killing speculative task
• links to: GitHub Pull Request #25674