jtuglu1 commented on PR #19025:
URL: https://github.com/apache/druid/pull/19025#issuecomment-3929707850

   > @jtuglu1 , to address the CI failures, would it make sense to not fully 
parallelize the allocation and instead use batches of size equal to the number 
of cores? Claude seems to think that this would reduce the contention in the 
JVM direct mem allocator.
   
   I think this can still run into issues since you're not guaranteeing and 
completion deadline on the buffer alloc tasks. This means you can still occupy 
the common thread pool threads which might cause other deadlock issues. To 
resolve this, I've created a temporary FJP to perform the allocs. Normally, 
doing this sort of thing would be prohibitive, however FJP threads are created 
lazily and there are only at most 2 production usages of this pool per node, so 
we're spinning up at most dedicate allocation pools per node on boot which I 
think is reasonable, LMK if you disagree.
   
   An alternative would be to make a static, shared FJP in the class.


-- 
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