Blocker Description
We've been combatting an orphaned process issue on AMPLab Jenkins since October and I finally was able to dig in and figure out what's going on.
After some sleuthing and working around OS limits and JDK bugs, I was able to get the full launch commands for the hanging orphaned processes. It looks like they're all running spark-submit:
org.apache.spark.deploy.SparkSubmit --master local-cluster[1,1,1024] --conf spark.driver.extraClassPath=/home/jenkins/workspace/Spark-Master-SBT/AMPLAB_JENKINS_BUILD_PROFILE/hadoop2.0/label/spark-test/core/target/scala-2.10/test-classes:/home/jenkins/workspace/Spark-Master-SBT/AMPLAB_JENKINS_BUILD_PROFILE/hadoop2.0/label/spark-test/core/target/scala-2.10/classes:/home/jenkins/workspace/Spark-Master-SBT/AMPLAB_JENKINS_BUILD_PROFILE/hadoop2.0/label/spark-test/launcher/target/scala-2.10/classes:/home/jenkins/workspace/Spark-Master-SBT/AMPLAB_JENKINS_BUILD_PROFILE/hadoop2.0/label/spark-test/network/common/target/scala-2.10/classes:/home/jenkins/workspace/Spark-Master-SBT/AMPLAB_JENKINS_BUILD_PROFILE/hadoop2.0/label/spark-test/network/shuffle/target/scala-2.10/classes:/home/jenkins/workspace/Spark-Master-SBT/AMPLAB_JENKINS_BUILD_PROFILE/hadoop2.0/label/spark-test/unsafe/target/scala-2.10/classes:/home/jenkins/workspace/Spark-Master-SBT/AMPLAB_JENKINS_BUILD_PROFILE/hadoop2.0/label/spark-test/tags/target/scala-2.10/ -Xms1g -Xmx1g -Dtest.appender=console -XX:MaxPermSize=256m
Based on the output of some Ganglia graphs, I was able to figure out that these leaks started around October 9.
This roughly lines up with when https://github.com/apache/spark/pull/7052 was merged, which added LauncherBackendSuite. The launch arguments used in this suite seem to line up with the arguments that I observe in the hanging processes' jps output: https://github.com/apache/spark/blame/1bc41125ee6306e627be212969854f639969c440/core/src/test/scala/org/apache/spark/launcher/LauncherBackendSuite.scala#L46
Interestingly, Jenkins doesn't show test timing or output for this suite! I think that what might be happening is that we have a mixed Scala/Java package, so maybe the two test runner XML files aren't being merged properly: https://amplab.cs.berkeley.edu/jenkins/job/Spark-1.5-SBT/746/AMPLAB_JENKINS_BUILD_PROFILE=hadoop1.0,label=spark-test/testReport/org.apache.spark.launcher/
Whenever I try running this suite locally, it looks like it ends up creating a zombie SparkSubmit process! I think that what's happening is that the launcher's handle.kill() call ends up destroying the bash spark-submit subprocess such that its child process (a JVM) leaks.
I think that we'll have to do something similar to what we do in PySpark when launching a child JVM from a Python / Bash process: connect it to a socket or stream such that it can detect its parent's death and clean up after itself appropriately.
/cc shaneknapp and vanzin.
Attachments
Attachments
Issue Links
- links to
