zhengchenyu commented on PR #1660:
URL: 
https://github.com/apache/incubator-uniffle/pull/1660#issuecomment-2072211403

   > Thanks for your work. @zhengchenyu After reviewing the design doc and 
concrete code briefly, I have some question about this feature.
   > 
   > ### Spark client
   > 1. What's the diference of RecordBlob and RecordBuffer? The combine and 
sort difference is not reflected on the names
   > 2. Why not making the `RMWriteBufferManager` extend the 
`WriteBufferManager`,  `RMWriterBufferManager` can almost replace the 
WriteBufferManager, we could control whether sort by config.
   > 3. Is the local sort in block level enoguh? Can we make the block bigger 
if spill to file?
   > 
   > ### Shuffle-Server
   > 1. If the merge failed, the reducer should failover to original mechanism
   > 2. The merge process will introduce too much random read. If this happens 
on the HDD, the whole process is terrible. So the key point is to make the 
block bigger (Sort merge them before flushing to disk to create a bigger 
block?) But this looks will break the original blockId mechanism
   
   Spark client
   (1) RecordBlob is used for combine and RecordBuffer is used for sort
   (2) RMWriteBufferManager should be extended from WriteBufferManager. I just 
don't want to affect other code. If the proposal is accepted, I will do it 
later.
   (3) If we wanna bigger block, best way is increase the block size.
   
   Shuffle-Server
   (1) Yes. We need failover to original rss, especially if we can't load class 
of key, value, comparator.
   (2) I firstly test in HDD, it is terrible. Then I test in SSD cluster, then 
better. We can try to merge the block, but need lots of modification. I just 
wanted to do it quickly, so I reused the original mechanism 


-- 
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.

To unsubscribe, e-mail: [email protected]

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