In general, I like this idea, because before we would potentially end up with a 
log of additional RocksDB threads that might not be needed - they are all 
probably blocking on the same disk anyway!

Did you measure the performance changes in some scenarios (1 RocksDB instances 
vs. 10 or even more) to verify that we are actually not making things worse 
with this change?

Previously, with (for example) `FLASH_SSD_OPTIMIZED`, every RocksDB instance 
was using 4 threads and now all together will use 4?
-> Should we tweak the default settings for the number of threads, also in the 
other profiles?

[ Full content available at: https://github.com/apache/flink/pull/6603 ]
This message was relayed via gitbox.apache.org for [email protected]

Reply via email to