[YARN-8382] cgroup file leak in NM

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 3.2.0, 3.1.1, 3.0.4, 2.10.1
• Component/s: nodemanager

Environment

we write an container with a shutdownHook which has a piece of code like  "while(true) sleep(100)" . when yarn.nodemanager.linux-container-executor.cgroups.delete-timeout-ms < yarn.nodemanager.sleep-delay-before-sigkill.ms , cgourp file leak happens; when yarn.nodemanager.linux-container-executor.cgroups.delete-timeout-ms > ** yarn.nodemanager.sleep-delay-before-sigkill.ms, cgroup file is deleted successfully**

• Hadoop Flags: Reviewed
• Assignee: Hu Ziqian
• Reporter: Hu Ziqian
• Created: 31/May/18 03:18
• Updated: 27/Apr/20 21:19
• Resolved: 04/Jun/18 18:17

Description

As Jiandan said in YARN-6562, NM may delete  Cgroup container file timeout with logs like below:

org.apache.hadoop.yarn.server.nodemanager.util.CgroupsLCEResourcesHandler: Unable to delete cgroup at: /cgroup/cpu/hadoop-yarn/container_xxx, tried to delete for 1000ms

 

we found one situation is that when we set yarn.nodemanager.sleep-delay-before-sigkill.ms bigger than yarn.nodemanager.linux-container-executor.cgroups.delete-timeout-ms, the cgroup file leak happens . 

 

One container process tree looks like follow graph:

bash(16097)───java(16099)─┬─{java}(16100) 

                                                  ├─{java}(16101) 

                       ├─{java}(16102)

 

when NM kills a container, NM sends kill -15 -pid to kill container process group. Bash process will exit when it received sigterm, but java process may do some job (shutdownHook etc.), and doesn't exit unit receive sigkill. And when bash process exits, CgroupsLCEResourcesHandler begin to try to delete cgroup files. So when yarn.nodemanager.linux-container-executor.cgroups.delete-timeout-ms arrived, the java processes may still running and cgourp/tasks still not empty and cause a cgroup file leak.

 

we add a condition that yarn.nodemanager.linux-container-executor.cgroups.delete-timeout-ms must bigger than yarn.nodemanager.sleep-delay-before-sigkill.ms to solve this problem.