We are observing very poor performance with a Java Kafka Producer 0.9 client when sending small messages. The messages are not being accumulated into a larger request batch and thus each small record is being sent separately.
What is wrong with our client configuration? Or is this some other issue?
Using Kafka Client 0.9.0.0. We did not see any related postings in the Kafka unreleased 9.0.1 or 9.1 fixed or unresolved lists, so we are focused on our client configuration and server instance.
We understand the linger.ms should cause the client to accumulate records into a batch.
We set linger.ms to 10 (and also tried 100 and 1000) but these did not result in the batch accumulating records. With a record size of about 100 bytes and a request buffer size of 16K, We would have expected about 160 messages to be sent in a single request.
The trace at the client seems to indicate that the partition may be full, despite having allocated a fresh Bluemix Messaging Hub (Kafka Server 0.9) service instance. The test client is sending multiple messages in a loop with no other I/O.
The log shows a repeating sequence with a suspect line: "Waking up the sender since topic mytopic partition 0 is either full or getting a new batch".
So the newly allocated partition should be essentially empty in our test case, thus why would the producer client be getting a new batch?
We provided the following properties file:
The Kafka Client shows the expanded/merged configuration list (and displaying the linger.ms setting):