Can you sanity check this with the end-to-end latency test that ships with Kafka in the tools package?
https://apache.googlesource.com/kafka/+/1769642bb779921267bd57d3d338591dbdf33842/core/src/main/scala/kafka/tools/TestEndToEndLatency.scala On Saturday, June 25, 2016, Kafka <kafka...@126.com> wrote: > Hi all， > my kafka cluster is composed of three brokers with each have 8core > cpu and 8g memory and 1g network card. > with java async client，I sent 1000000 messages with size of 1024 > bytes per message ,the send gap between each sending is 20us，the consumer’s > config is like this，fetch.min.bytes is set to 1, fetch.wait.max.ms is set > to 100. > to avoid the inconformity bewteen two machines，I start producer > and consumer at the same machine，the machine’s configurations have enough > resources to satisfy these two clients. > > I start consumer before producer on each test，with the sending > timestamp in each message，when consumer receive the message，then I can got > the consumer delay through the substraction between current timesstamp and > sending timestamp. > when I set acks to 0，replica to 2，then the average producer delay > is 2.98ms， the average consumer delay is 52.23ms. > when I set acks to 1，replica to 2，then the average producer delay > is 3.9ms，the average consumer delay is 44.88ms. > when I set acks to -1, replica to 2, then the average producer > delay is 1782ms, the average consumer delay is 1786ms. > > I have two doubts，the first is why my consumer's delay with acks > settled to 0 is logger than the consumer delay witch acks settled to 1. > the second is why the delay of producer and consumer is so big when I set > acks to -1，I think this delay is can not be accepted. > and I found this delay is amplified with sending more messages. > > any feedback is appreciated. > thanks > > > >
