Page History

...

For (B) we need another timeout to expire the batches in accumulator in addition to batching purpose. That means we probably need the following configuration.

      54. (B) batch.timeout.ms - In some cases, the leader broker information of a batch could be unavailable after the batch is ready to be sent (i.e. batch is full or reached linger.ms). This configuration define the maximum time the producer will wait for the necessary broker information before it fail the batches.

For (A) it makes sense to have a dedicated timeout for send() method to provide blocking bound. So it gives another configuration.

      65. (A) max.send.block.ms - This timeout defines how long the producer.send() method can block (we will enforce this for both metadata and blocking on buffer full in a separate ticket).

...

• (A/B) metadata.fetch.timeout.ms - reuse metadata timeout as batch.timeout.ms because it is essentially metadata not available.
  
• (C) replication.timeout.ms
  
• (C) network.request.timeout.ms

Proposal 2:

• (A) max.send.block.ms
• (B) batch.timeout.ms
• (C) replication.timeout.ms
• (C) network.request.timeout.ms

Proposal 1 has four configurations but does got guarantee of blocking time for send() when blocking on buffer full is turned on. So user needs to catch the exception in outside send and retry if buffer is full.

...