Vinod Kumar Vavilapalli commented on YARN-3021:

bq. RM can simply inspect the incoming renewer specified in the token and skip 
renewing those tokens if the renewer doesn't match it's own address. This way, 
we don't need an explicit API in the submission context.
bq. Seems regardless of this jira, we could have do the above change, right? 
any catch?
I think this will work, and is a preferable solution to me. What do others 
think? /cc [~qwertymaniac]/[~jianhe]

bq. In our simple tests the app did run through successfully with such an 
approach, but there was multiple factors we did not test for (app recovery, 
task failures, etc. which could be impacted). Would it be better if we added in 
a morphed DelegationTokenRenewer (which does NOP as part of actual renewal 
logic), instead of skipping adding in the renewer completely?
Never mind. I looked at the patch again, it would work, barring my other 
non-code comments.

> YARN's delegation-token handling disallows certain trust setups to operate 
> properly over DistCp
> -----------------------------------------------------------------------------------------------
>                 Key: YARN-3021
>                 URL: https://issues.apache.org/jira/browse/YARN-3021
>             Project: Hadoop YARN
>          Issue Type: Bug
>          Components: security
>    Affects Versions: 2.3.0
>            Reporter: Harsh J
>         Attachments: YARN-3021.001.patch, YARN-3021.002.patch, 
> YARN-3021.003.patch, YARN-3021.patch
> Consider this scenario of 3 realms: A, B and COMMON, where A trusts COMMON, 
> and B trusts COMMON (one way trusts both), and both A and B run HDFS + YARN 
> clusters.
> Now if one logs in with a COMMON credential, and runs a job on A's YARN that 
> needs to access B's HDFS (such as a DistCp), the operation fails in the RM, 
> as it attempts a renewDelegationToken(…) synchronously during application 
> submission (to validate the managed token before it adds it to a scheduler 
> for automatic renewal). The call obviously fails cause B realm will not trust 
> A's credentials (here, the RM's principal is the renewer).
> In the 1.x JobTracker the same call is present, but it is done asynchronously 
> and once the renewal attempt failed we simply ceased to schedule any further 
> attempts of renewals, rather than fail the job immediately.
> We should change the logic such that we attempt the renewal but go easy on 
> the failure and skip the scheduling alone, rather than bubble back an error 
> to the client, failing the app submission. This way the old behaviour is 
> retained.

This message was sent by Atlassian JIRA

Reply via email to