Hi, I have two questions when writing patch for KAFKA-2142 and needs some help:
1. In send(), we actually might call callback.onCompletion(), this might break the guarantee that callbacks are fired in sending order of a particular partition. 2. In RecordAccumulator, current logic is that if one batch to a broker is ready (full or reaches linger.ms), we drain all the batches to the same broker even if other batches are not full or reaching linger.ms yet. Guozhang mentioned that this is by design, but it seems different from the java doc and might have negative effect on batching. Was there any concern on this? Thanks. Jiangjie (Becket) Qin
