avamingli commented on PR #1178:
URL: https://github.com/apache/cloudberry/pull/1178#issuecomment-2990138113

   > The phenomenon showed that the deduplication effect of CBDB’s **Streaming 
Partial HashAggregate** was significantly worse. As a result, the **Finalize 
HashAggregate** operator in CBDB processed significantly more data compared to 
GPDB under the same dataset
   
   Thank you for your insights!
   
   The Streaming HashAggregate does not spill; it sends out tuples once the 
hash table is full. Its performance is closely linked to the distinct property 
of the data from subnodes within the in-memory hash range. However, evaluating 
this distinct property can be challenging.
   
   While HashAggregate may require more space and disk I/O, it can deduplicate 
more tuples. Conversely, Streaming HashAggregate might incur higher network 
I/O, deduplicating fewer tuples but potentially speeding up slice execution 
upon the motion node. Ultimately, the effectiveness of either approach depends 
on factors like network, disk performance, and, crucially, the distinct 
property of subnode data, especially given the in-memory limit of the hash 
table.
   
   I encountered a similar observation in #1173, where we discussed adding a 
GUC to control Streaming HashAggregate. This allows users to tailor the 
approach to their specific environments, which could be beneficial, especially 
in ORCA.
   
   Anyway, your work looks great!


-- 
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