In production at Linkedin, we use a batch size of 200, which has worked
If the producer shuts down, it flushes out the messages in the internal
queue. If it crashes, it does not get a chance to do that and there is no
good alternative to this.