...

In the rare scenario where user thread sends a message when the metadata is not available, but never calls producer.send() after metadata is available, the message will never be put into the per-partition queue with the changes above. In order to take care of this scenario, this KIP should additional modify the producer's IO thread such that, if no new ProduceRequest can be generated and there exists non-empty per-topic queue whose topic's partition information is available, IO thread should do the step 3) above to move these messages into the per-partition queue.

Theoretical Performance Analysis

Here we analyze how the changes proposed in this KIP would change the performance of producer:

1) Worst case time of producer.send() suppose the producer queue is not full.

Before this KIP, the producer.send() may block up to max.block.ms for metadata to become available. This KIP improves the latency in this scenario by putting the message in the per-topic queue without blocking producer.send().

This KIP may also increase the time of producer.send() in the following scenario. The first time producer.send() is called after metadata becomes available, according to step 3) in the proposed changes above, producer.send() will compress messages and put these messages into the batch for the corresponding partitions. If there is a lot of messages in the per-topic queues, this can take some time. But amount of time to compress messages should be considerably less than the time needed to wait for metadata (and compress messages) because user threads are better utilized by not waiting for metadata.

2) Throughput

This change does not affect the rate at which IO thread can send messages to producer. Thus the difference in throughput should mainly be caused by how we do computation intensive tasks (i.e. serialization and compression) and how we wait for metadata.

If producer.send() is called infrequently, throughput should not be an issue. If producer.send() is called frequently, all computation intensive task, i.e. compression and serialization, will still be done by the user threads, which is the same as the existing producer implementation. Thus the change proposed in the KIP sohuld not reduce throughput terms of computation.

On the other hand, since user thread will not longer block on metadata, user thread can be better utilized to do other useful work. So the overall throughput should be better with the change proposed in this KIP.

3) Latency of message from the time producer.send() is called to the time the message is acknowledged

If metadata is available, the latency should be the same.

If metadata is not available, messages can not be sent. latency should be dominated by the time to wait for metadata and will still be the same. Once the metadata becomes available, user thread will be busy compressing these messages and putting them into the per-partition queue. Since the throughput should be better (see the analysis above), the latency in this case should also be slightly better.

Compatibility, Deprecation, and Migration Plan

...