Josh Rosen created SPARK-19685:
----------------------------------
Summary: PipedRDD tasks should not hang on interruption / errors
Key: SPARK-19685
URL: https://issues.apache.org/jira/browse/SPARK-19685
Project: Spark
Issue Type: Bug
Components: Spark Core
Affects Versions: 2.1.0, 2.0.0, 1.6.0
Reporter: Josh Rosen
While looking at WARN and ERROR-level logs from Spark executors, I spotted a problem where
PipedRDD tasks may continue running after being cancelled or after failing. Specifically,
I saw many cancelled tasks hanging in the following stacks:
{code}
java.io.BufferedOutputStream.flush(BufferedOutputStream.java:140)
java.io.FilterOutputStream.close(FilterOutputStream.java:158)
java.lang.UNIXProcess.destroy(UNIXProcess.java:445)
java.lang.UNIXProcess.destroy(UNIXProcess.java:478)
org.apache.spark.rdd.PipedRDD$$anon$1.propagateChildException(PipedRDD.scala:203)
org.apache.spark.rdd.PipedRDD$$anon$1.hasNext(PipedRDD.scala:183)
scala.collection.Iterator$$anon$12.hasNext(Iterator.scala:439)
scala.collection.Iterator$class.foreach(Iterator.scala:893)
scala.collection.AbstractIterator.foreach(Iterator.scala:1336)
scala.collection.TraversableOnce$class.foldLeft(TraversableOnce.scala:157)
scala.collection.AbstractIterator.foldLeft(Iterator.scala:1336)
scala.collection.TraversableOnce$class.fold(TraversableOnce.scala:212)
scala.collection.AbstractIterator.fold(Iterator.scala:1336)
org.apache.spark.rdd.RDD$$anonfun$fold$1$$anonfun$20.apply(RDD.scala:1086)
org.apache.spark.rdd.RDD$$anonfun$fold$1$$anonfun$20.apply(RDD.scala:1086)
org.apache.spark.SparkContext$$anonfun$33.apply(SparkContext.scala:1980)
org.apache.spark.SparkContext$$anonfun$33.apply(SparkContext.scala:1980)
org.apache.spark.scheduler.ResultTask.runTask(ResultTask.scala:87)
org.apache.spark.scheduler.Task.run(Task.scala:99)
org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:322)
java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
java.lang.Thread.run(Thread.java:745)
{code}
and
{code}
java.io.FileInputStream.readBytes(Native Method)
java.io.FileInputStream.read(FileInputStream.java:255)
java.io.BufferedInputStream.read1(BufferedInputStream.java:284)
java.io.BufferedInputStream.read(BufferedInputStream.java:345)
sun.nio.cs.StreamDecoder.readBytes(StreamDecoder.java:284)
sun.nio.cs.StreamDecoder.implRead(StreamDecoder.java:326)
sun.nio.cs.StreamDecoder.read(StreamDecoder.java:178)
java.io.InputStreamReader.read(InputStreamReader.java:184)
java.io.BufferedReader.fill(BufferedReader.java:161)
java.io.BufferedReader.readLine(BufferedReader.java:324)
java.io.BufferedReader.readLine(BufferedReader.java:389)
scala.io.BufferedSource$BufferedLineIterator.hasNext(BufferedSource.scala:72)
org.apache.spark.rdd.PipedRDD$$anon$1.hasNext(PipedRDD.scala:172)
scala.collection.Iterator$$anon$12.hasNext(Iterator.scala:439)
scala.collection.Iterator$class.foreach(Iterator.scala:893)
scala.collection.AbstractIterator.foreach(Iterator.scala:1336)
scala.collection.TraversableOnce$class.foldLeft(TraversableOnce.scala:157)
scala.collection.AbstractIterator.foldLeft(Iterator.scala:1336)
scala.collection.TraversableOnce$class.fold(TraversableOnce.scala:212)
scala.collection.AbstractIterator.fold(Iterator.scala:1336)
org.apache.spark.rdd.RDD$$anonfun$fold$1$$anonfun$20.apply(RDD.scala:1086)
org.apache.spark.rdd.RDD$$anonfun$fold$1$$anonfun$20.apply(RDD.scala:1086)
org.apache.spark.SparkContext$$anonfun$33.apply(SparkContext.scala:1980)
org.apache.spark.SparkContext$$anonfun$33.apply(SparkContext.scala:1980)
org.apache.spark.scheduler.ResultTask.runTask(ResultTask.scala:87)
org.apache.spark.scheduler.Task.run(Task.scala:99)
org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:322)
java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
java.lang.Thread.run(Thread.java:745)
{code}
I do not have a minimal reproduction of this issue yet, but I suspect that we can make one
by having PipedRDD call a process which hangs indefinitely without printing any output, then
cancel the Spark job with {{interruptOnCancel=true}}. If my hunch is right, we should witness
the PipedRDD tasks continuing to run either because the call to destroy the child process
is hanging or because we don't check whether the task has been interrupted.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org
|