Jason Lowe commented on YARN-3103:

Sorry for the confusion, I'll try to go through it step-by-step.  It's similar 
to the problem we saw in MAPREDUCE-6230 which is the same kind of AMRM token 
updating code, just in the MR-AM code rather than the AMRMClientImpl code.

# The RM creates an AMRM token as part of setting up the AM's launch context.  
That token is created with an empty service name, see 
# The RM then pokes this token into the launch context credentials, using the 
token's service name as the key.  See AMLauncher.setupTokens and 
Credentials.addToken.  We now have an AMRM token in the credentials mapped 
under an empty service name as the key.
# When the AMRMClient receives an updated AMRM token it updates the service 
name of the token _before_ it pokes it into the credentials.  See 
AMRMClientImpl.updateAMRMToken.  Therefore we are adding the token using a 
non-empty service name as the key in the credentials token map.
# After updating the credentials we now have _two_ AMRM tokens in the 
credentials map, one with an empty service name key (but a non-empty service 
name on the token itself) and one where the service name key equals the token's 
service name.  In other words, we failed to replace the existing AMRM token in 
the credentials

Also for a secure setup, we're updating the wrong UGI.  IIUC the ClientRMProxy 
will use the UGI obtained from the current user when performing the connection 
to the RM, so updating the login user rather than the current user can update 
the wrong UGI if we're currently running in a doAs context where the UGI isn't 
the login UGI.

> AMRMClientImpl does not update AMRM token properly
> --------------------------------------------------
>                 Key: YARN-3103
>                 URL: https://issues.apache.org/jira/browse/YARN-3103
>             Project: Hadoop YARN
>          Issue Type: Bug
>          Components: client
>    Affects Versions: 2.6.0
>            Reporter: Jason Lowe
>            Assignee: Jason Lowe
>            Priority: Blocker
> AMRMClientImpl.updateAMRMToken updates the token service _before_ storing it 
> to the credentials, so the token is mapped using the newly updated service 
> rather than the empty service that was used when the RM created the original 
> AMRM token.  This leads to two AMRM tokens in the credentials and can still 
> fail if the AMRMTokenSelector picks the wrong one.
> In addition the AMRMClientImpl grabs the login user rather than the current 
> user when security is enabled, so it's likely the UGI being updated is not 
> the UGI that will be used when reconnecting to the RM.
> The end result is that AMs can fail with invalid token errors when trying to 
> reconnect to an RM after a new AMRM secret has been activated.

This message was sent by Atlassian JIRA

Reply via email to