[
https://issues.apache.org/jira/browse/YARN-8751?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16605895#comment-16605895
]
Craig Condit commented on YARN-8751:
------------------------------------
{quote}[~jlowe] : So my vote is keep INVALID_CONTAINER_EXEC_PERMISSIONS and
INVALID_CONFIG_FILE fatal but the others should only fail the single container
launch rather than the whole NM process.
{quote}
Agreed. The remainder of the exit codes could be caused by any number of
things, such as disk failure, which you point out. Even if the problem were to
be caused by something more systemic, NM blacklisting should kick in pretty
quickly as tasks fail. +1 on making this non-fatal. Additionally, we may want
to consider updating the diagnostic message returned in the following {{else}}
clause to contain the exit code enum name as well as the number – this would
seem to make diagnosing problems much easier for both users and administrators.
> Container-executor permission check errors cause the NM to be marked unhealthy
> ------------------------------------------------------------------------------
>
> Key: YARN-8751
> URL: https://issues.apache.org/jira/browse/YARN-8751
> Project: Hadoop YARN
> Issue Type: Bug
> Reporter: Shane Kumpf
> Priority: Critical
>
> {{ContainerLaunch}} (and {{ContainerRelaunch}}) contains logic to mark a
> NodeManager as UNHEALTHY if a {{ConfigurationException}} is thrown by
> {{ContainerLaunch#launchContainer}} (or relaunchContainer). The exception
> occurs based on the exit code returned by container-executor, and 7 different
> exit codes cause the NM to be marked UNHEALTHY.
> {code:java}
> if (exitCode ==
> ExitCode.INVALID_CONTAINER_EXEC_PERMISSIONS.getExitCode() ||
> exitCode ==
> ExitCode.INVALID_CONFIG_FILE.getExitCode() ||
> exitCode ==
> ExitCode.COULD_NOT_CREATE_SCRIPT_COPY.getExitCode() ||
> exitCode ==
> ExitCode.COULD_NOT_CREATE_CREDENTIALS_FILE.getExitCode() ||
> exitCode ==
> ExitCode.COULD_NOT_CREATE_WORK_DIRECTORIES.getExitCode() ||
> exitCode ==
> ExitCode.COULD_NOT_CREATE_APP_LOG_DIRECTORIES.getExitCode() ||
> exitCode ==
> ExitCode.COULD_NOT_CREATE_TMP_DIRECTORIES.getExitCode()) {
> throw new ConfigurationException(
> "Linux Container Executor reached unrecoverable exception", e);{code}
> I can understand why these are treated as fatal with the existing process
> container model. However, with privileged Docker containers this may be too
> harsh, as Privileged Docker containers don't guarantee the user's identity
> will be propagated into the container, so these mismatches can occur. Outside
> of privileged containers, an application may inadvertently change the
> permissions on one of these directories, triggering this condition.
> In our case, a container changed the "appcache/<appid>/<containerid>"
> directory permissions to 774. Some time later, the process in the container
> died and the Retry Policy kicked in to RELAUNCH the container. When the
> RELAUNCH occurred, container-executor checked the permissions of the
> "appcache/<appid>/<containerid>" directory (the existing workdir is retained
> for RELAUNCH) and returned exit code 35. Exit code 35 is
> COULD_NOT_CREATE_WORK_DIRECTORIES, which is a fatal error. This killed all
> containers running on that node, when really only this container would have
> been impacted.
> {code:java}
> 2018-08-31 21:07:22,365 INFO nodemanager.ContainerExecutor
> (ContainerExecutor.java:logOutput(541)) - Exception from container-launch.
> 2018-08-31 21:07:22,365 INFO nodemanager.ContainerExecutor
> (ContainerExecutor.java:logOutput(541)) - Container id:
> container_e15_1535130383425_0085_01_000005
> 2018-08-31 21:07:22,365 INFO nodemanager.ContainerExecutor
> (ContainerExecutor.java:logOutput(541)) - Exit code: 35
> 2018-08-31 21:07:22,365 INFO nodemanager.ContainerExecutor
> (ContainerExecutor.java:logOutput(541)) - Exception message: Relaunch
> container failed
> 2018-08-31 21:07:22,365 INFO nodemanager.ContainerExecutor
> (ContainerExecutor.java:logOutput(541)) - Shell error output: Could not
> create container dirsCould not create local files and directories 5 6
> 2018-08-31 21:07:22,365 INFO nodemanager.ContainerExecutor
> (ContainerExecutor.java:logOutput(541)) -
> 2018-08-31 21:07:22,365 INFO nodemanager.ContainerExecutor
> (ContainerExecutor.java:logOutput(541)) - Shell output: main : command
> provided 4
> 2018-08-31 21:07:22,365 INFO nodemanager.ContainerExecutor
> (ContainerExecutor.java:logOutput(541)) - main : run as user is user
> 2018-08-31 21:07:22,365 INFO nodemanager.ContainerExecutor
> (ContainerExecutor.java:logOutput(541)) - main : requested yarn user is yarn
> 2018-08-31 21:07:22,365 INFO nodemanager.ContainerExecutor
> (ContainerExecutor.java:logOutput(541)) - Creating script paths...
> 2018-08-31 21:07:22,365 INFO nodemanager.ContainerExecutor
> (ContainerExecutor.java:logOutput(541)) - Creating local dirs...
> 2018-08-31 21:07:22,365 INFO nodemanager.ContainerExecutor
> (ContainerExecutor.java:logOutput(541)) - Path
> /grid/0/hadoop/yarn/local/usercache/user/appcache/application_1535130383425_0085/container_e15_1535130383425_0085_01_000005
> has permission 774 but needs per
> mission 750.
> 2018-08-31 21:07:22,365 INFO nodemanager.ContainerExecutor
> (ContainerExecutor.java:logOutput(541)) - Wrote the exit code 35 to (null)
> 2018-08-31 21:07:22,386 ERROR launcher.ContainerRelaunch
> (ContainerRelaunch.java:call(129)) - Failed to launch container due to
> configuration error.
> org.apache.hadoop.yarn.exceptions.ConfigurationException: Linux Container
> Executor reached unrecoverable exception
> at
> org.apache.hadoop.yarn.server.nodemanager.LinuxContainerExecutor.handleExitCode(LinuxContainerExecutor.java:633)
> at
> org.apache.hadoop.yarn.server.nodemanager.LinuxContainerExecutor.handleLaunchForLaunchType(LinuxContainerExecutor.java:573)
> at
> org.apache.hadoop.yarn.server.nodemanager.LinuxContainerExecutor.relaunchContainer(LinuxContainerExecutor.java:486)
> at
> org.apache.hadoop.yarn.server.nodemanager.containermanager.launcher.ContainerLaunch.relaunchContainer(ContainerLaunch.java:504)
> at
> org.apache.hadoop.yarn.server.nodemanager.containermanager.launcher.ContainerRelaunch.call(ContainerRelaunch.java:111)
> at
> org.apache.hadoop.yarn.server.nodemanager.containermanager.launcher.ContainerRelaunch.call(ContainerRelaunch.java:47)
> at java.util.concurrent.FutureTask.run(FutureTask.java:266)
> at
> java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
> at
> java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
> at java.lang.Thread.run(Thread.java:745)
> Caused by:
> org.apache.hadoop.yarn.server.nodemanager.containermanager.runtime.ContainerExecutionException:
> Relaunch container failed
> at
> org.apache.hadoop.yarn.server.nodemanager.containermanager.linux.runtime.DockerLinuxContainerRuntime.relaunchContainer(DockerLinuxContainerRuntime.java:987)
> at
> org.apache.hadoop.yarn.server.nodemanager.containermanager.linux.runtime.DelegatingLinuxContainerRuntime.relaunchContainer(DelegatingLinuxContainerRuntime.java:150)
> at
> org.apache.hadoop.yarn.server.nodemanager.LinuxContainerExecutor.handleLaunchForLaunchType(LinuxContainerExecutor.java:562)
> ... 8 more
> {code}
> The root of the issue could be considered the fact that we can't guarantee
> which user is running in the container, and should eliminate writable mounts
> in this scenario. However, marking the NM unhealthy in all these cases does
> seem overkill.
> Opening this to discuss how we want to address this issue. [~jlowe]
> [~ebadger] [~Jim_Brennan] [~eyang] [~billie.rinaldi] [~ccondit-target] let me
> know your thoughts.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
