andygrove opened a new issue, #3778:
URL: https://github.com/apache/datafusion-comet/issues/3778

   ### What is the problem the feature request solves?
   
   Following on from the refactors to move shuffle code into a separate crate 
in https://github.com/apache/datafusion-comet/pull/3749 and 
https://github.com/apache/datafusion-comet/pull/3772, and the new standalone 
shuffle benchmark binary in 
https://github.com/apache/datafusion-comet/pull/3752, I propose that we start 
evaluating some different shuffle approaches to better understand the trade 
offs in terms of throughput and memory usage of different approaches.
   
   This is a very open-ended epic. There are some ideas that I am planning on 
experimenting with, such as moving away from the current interleave batches 
approach, which keeps batches buffered for a long time, and moving to a more 
immediate scatter approach that removes the need to buffer the input batches 
for such a long time. I expect other contributors may also have ideas that they 
would likt to explore.
   
   There is a already a trait for the shuffle partition-and-write operation, so 
it should be trivial to support different implementations behind a config for 
testing and evaluation, and allow the standalone benchmark to compare 
performance characteristics across implementations.
   
   
   
   
   ### Describe the potential solution
   
   _No response_
   
   ### Additional context
   
   _No response_


-- 
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