I am using Kafka 0.7.2
I have a consumer that tries to fetch messages from several partitions on
the same broker (and topic).
The consumers issues a fetch request per partition, if there are messages
to read, it fetches the first one and goes to the next partition to fetch
another message and so on (in a loop).
When there are not messages in a partition, the consumer will issue another
fetch request on the next cycle etc.
I noticed that starting with the 2nd fetch request, response time becomes
unacceptable, i.e. reading back the response from the broker takes ~0.04
second (while it took ~0.0006 sec on the first time).
This behavior is consistent over different runs.