hadoop-hdfs-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Alejandro Abdelnur <t...@cloudera.com>
Subject Re: Job end notification does not always work (Hadoop 2.x)
Date Mon, 24 Jun 2013 16:11:46 GMT
if we ought to do this in a yarn service it 
should be the RM or the HS. the RM is, IMO, the natural fit. the HS, would be a good choice
if we are concerned about the extra work this would cause in the RM. the problem with the
current HS is that it is MR specific, we should generalize it for diff AM types. 

thx

Alejandro
(phone typing)

On Jun 23, 2013, at 23:28, Devaraj k <devaraj.k@huawei.com> wrote:

> Even if we handle all the failure cases in AM for Job End Notification, we may miss cases
like abrupt kill of AM when it is in last retry. If we choose NM to give the notification,
again RM needs to identify which NM should give the end-notification as we don't have any
direct protocol between AM and NM.
>  
> I feel it would be better to move End-Notification responsibility to RM as Yarn Service
because it ensures 100% notification and also useful for other types of applications as well.
>  
>  
> Thanks
> Devaraj K
>  
> From: Ravi Prakash [mailto:ravihoo@ymail.com] 
> Sent: 23 June 2013 19:01
> To: user@hadoop.apache.org
> Subject: Re: Job end notification does not always work (Hadoop 2.x)
>  
> Hi Alejandro,
> 
> Thanks for your reply! I was thinking more along the lines Prashant suggested i.e. a
failure during init() should still trigger an attempt to notify (by the AM). But now that
you mention it, maybe we would be better of including this as a YARN feature after all (specially
with all the new AMs being written). We could let the NM of the AM handle the notification
burden, so that the RM doesn't get unduly taxed. Thoughts?
> 
> Thanks
> Ravi
>  
>  
> From: Alejandro Abdelnur <tucu@cloudera.com>
> To: "common-user@hadoop.apache.org" <user@hadoop.apache.org> 
> Sent: Saturday, June 22, 2013 7:37 PM
> Subject: Re: Job end notification does not always work (Hadoop 2.x)
>  
> If the AM fails before doing the job end notification, at any stage of the execution
for whatever reason, the job end notification will never be deliver. There is not way to fix
this unless the notification is done by a Yarn service. The 2 'candidate' services for doing
this would be the RM and the HS. The job notification URL is in the job conf. The RM never
sees the job conf, that rules out the RM out unless we add, at AM registration time the possibility
to specify a callback URL. The HS has access to the job conf, but the HS is currently a 'passive'
service.
> 
> thx
>  
> On Sat, Jun 22, 2013 at 3:48 PM, Arun C Murthy <acm@hortonworks.com> wrote:
> Prashanth, 
>  
>  Please file a jira.
>  
>  One thing to be aware of - AMs get restarted a certain number of times for fault-tolerance
- which means we can't just assume that failure of a single AM is equivalent to failure of
the job.
>  
>  Only the ResourceManager is in the appropriate position to judge failure of AM v/s failure-of-job.
>  
> hth,
> Arun
>  
> On Jun 22, 2013, at 2:44 PM, Prashant Kommireddi <prash1784@gmail.com> wrote:
> 
> 
> Thanks Ravi.
> 
> Well, in this case its a no-effort :) A failure of AM init should be considered as failure
of the job? I looked at the code and best-effort makes sense with respect to retry logic etc.
You make a good point that there would be no notification in case AM OOMs, but I do feel AM
init failure should send a notification by other means.
> 
>  
> 
> On Sat, Jun 22, 2013 at 2:38 PM, Ravi Prakash <ravihoo@ymail.com> wrote:
> Hi Prashant,
> 
> I would tend to agree with you. Although job-end notification is only a "best-effort"
mechanism (i.e. we cannot always guarantee notification for example when the AM OOMs), I agree
with you that we can do more. If you feel strongly about this, please create a JIRA and possibly
upload a patch.
> 
> Thanks
> Ravi
>  
>  
> From: Prashant Kommireddi <prash1784@gmail.com>
> To: "user@hadoop.apache.org" <user@hadoop.apache.org> 
> Sent: Thursday, June 20, 2013 9:45 PM
> Subject: Job end notification does not always work (Hadoop 2.x)
>  
> Hello,
> 
> I came across an issue that occurs with the job notification callbacks in MR2. It works
fine if the Application master has started, but does not send a callback if the initializing
of AM fails.
> 
> Here is the code from MRAppMaster.java
> 
> .....
> .......
>       // set job classloader if configured
>       MRApps.setJobClassLoader(conf);
>       initAndStartAppMaster(appMaster, conf, jobUserName);
>     } catch (Throwable t) {
>       LOG.fatal("Error starting MRAppMaster", t);
>       System.exit(1);
>     }
>   }
> 
> protected static void initAndStartAppMaster(final MRAppMaster appMaster,
>       final YarnConfiguration conf, String jobUserName) throws IOException,
>       InterruptedException {
>     UserGroupInformation.setConfiguration(conf);
>     UserGroupInformation appMasterUgi = UserGroupInformation
>         .createRemoteUser(jobUserName);
>     appMasterUgi.doAs(new PrivilegedExceptionAction<Object>() {
>       @Override
>       public Object run() throws Exception {
>         appMaster.init(conf);
>         appMaster.start();
>         if(appMaster.errorHappenedShutDown) {
>           throw new IOException("Was asked to shut down.");
>         }
>         return null;
>       }
>     });
>   }
> appMaster.init(conf) does not dispatch JobFinishEventHandler which is responsible for
sending a HTTP callback (via shutDownJob()). If there was an exception at this time, the process
would simply terminate (via System.exit(1) )
> 
> appMaster.start() however rightly uses the JobFinishEventHandler and things work fine.
> 
> Shouldn't a failure on init(..) also send a callback suggesting the job failed?
> 
> Thanks,
> Prashant
>  
>  
> 
>  
>  
> --
> Arun C. Murthy
> Hortonworks Inc.
> http://hortonworks.com/
> 
>  
> 
> 
>  
> -- 
> Alejandro
>  

Mime
View raw message