[GitHub] [kafka] ableegoldman commented on pull request #8892: KAFKA-10068: verify assignment performance with large cluster

[GitBox <gi...@apache.org>]

ableegoldman commented on pull request #8892:
URL: https://github.com/apache/kafka/pull/8892#issuecomment-691358663


   Alright this should finally be ready for a final pass and merge @vvcephei  @cadonna -- sorry for leaving this hanging for so long.
   
   After running a few different tests, it seems like the HATA assignor may actually be faster than the old StickyTaskAssignor in most cases. The HATA seems to scale slightly worse with increasing number of clients, but significantly better with partition count and num standbys
   
   The `testLargeNumConsumers` with 1,000 clients and 1,000 partitions (and 1 standby) took the HATA 20s for the full test, but only ~1-2s for the STA and FPTA. 
   
   The `testManyStandbys` with 100 clients,  1000 partitions, and 50 standbys took the HATA roughly 10s, and the STA/FPTA just under a full minute.
   
   The `testLargePartitionCount` with 1 client, 6,000 partitions, and 1 standby took the HATA under 1s. The STA and FPTA both ran out of time, surprisingly on the first assignment alone (taking just over 1.5 minutes)
   
   Decided to reduce the number of partitions in the `testLargePartitionCount` test to 3,000 rather than increasing the timeout for all tests, as it's already pretty high. Maybe we can drop the STA sooner or later and then tighten it up


----------------------------------------------------------------
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.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org