hiboyang edited a comment on pull request #31715:
URL: https://github.com/apache/spark/pull/31715#issuecomment-790013265


   > I do not think this flag (indicating whether to unregister all the shuffle 
blocks created by the failed executor) must be set at the application level 
(the way this PR does).
   > 
   > I think this must be a property kept for the replicated shuffle block so 
somewhere close to `MapStatus`.
   > 
   > Imagine a mixed solution where the disaggregated storage just a fallback 
and still locally stored blocks can be used.
   > Then it make sense to unregister the shuffle blocks for the failed 
executors and leave the disaggregated storage access for the blocks intact.
   
   Yeah, if user uses a mixed solution where the disaggregated storage just a 
fallback, it is better to have MapStatus to track availability of shuffle 
blocks. This PR does not block this scenario.
   
   There is non-mixed solution where people totally rely on remote shuffle 
service (like Facebook/Uber's one). In that case, people only need to set this 
flag in the application level. This PR will help this scenario.
   
   


----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
[email protected]



---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to