[jira] [Commented] (YARN-499) On container failure, include last n lines of logs in diagnostics

["Sandy Ryza (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/YARN-499?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13626058#comment-13626058 ] 

Sandy Ryza commented on YARN-499:
---------------------------------

Thanks for the feedback, Vinod.

The issue I am aiming to solve is the last one you mention of the AM crashing before registering with the RM.  A few JIRAs have been filed around this problem with little progress, so I wanted to put forth a concrete proposal.  I also wanted to bring back the stuff that gets printed out on the console for failing tasks, and under my proposal, thought that this would fall under the same issue of surfacing failed container logs.  If we decide on an approach that only handles the former, I'll file a separate JIRA for the latter.

bq. The stuff that gets printed on console is the client pulling logs directly.
You're right, I misunderstood the code and thought it was coming through the task diagnostics.  In light of that, I agree that my proposal isn't the right path.  Is there a reason we have avoided pulling the logs directly in YARN as well?  If not, should we do this for both the AM and task containers?

                
> On container failure, include last n lines of logs in diagnostics
> -----------------------------------------------------------------
>
>                 Key: YARN-499
>                 URL: https://issues.apache.org/jira/browse/YARN-499
>             Project: Hadoop YARN
>          Issue Type: Sub-task
>          Components: nodemanager
>    Affects Versions: 2.0.3-alpha
>            Reporter: Sandy Ryza
>            Assignee: Sandy Ryza
>         Attachments: YARN-499.patch
>
>
> When a container fails, the only way to diagnose it is to look at the logs.  ContainerStatuses include a diagnostic string that is reported back to the resource manager by the node manager.
> Currently in MR2 I believe whatever is sent to the task's standard out is added to the diagnostics string, but for MR standard out is redirected to a file called stdout.  In MR1, this string was populated with the last few lines of the task's stdout file, and got printed to the console, allowing for easy debugging.
> Handling this would help to soothe the infuriating problem of an AM dying for a mysterious reason before setting a tracking URL (MAPREDUCE-3688).
> This could be done in one of two ways.
> * Use tee to send MR's standard out to both the stdout file and standard out.  This requires modifying ShellCmdExecutor to roll what it reads in, as we wouldn't want to be storing the entire task log in NM memory.
> * Read the task's log files.  This would require standardizing or making the container log files configurable.  Right now the log files are determined in userland and all that is YARN is aware of the log directory.
> Does this present any issues I'm not considering?  If so it this might only be needed for AMs? 

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira