Hi all, A correction to my numbers: my benchmark script restarted the brokers in a way that left one broker leading all 3,000 partitions. With balanced leaders (1,000 per broker) the gain is smaller:
PLAINTEXT packets/s 22,274 -> 8,207 (-63%), broker CPU and latency the same as on trunk SSL packets/s 37,631 -> 8,992 (-76%), broker CPU and latency the same as on trunk I also changed the PR: a send that is written with one call (a ByteBufferSend without SSL) is not changed anymore, because there the toggle only cost CPU. Details are in the PR: https://github.com/apache/kafka/pull/24170 Best regards, Andrei Iatsuk
