Do not buffer messages. This is the default and most resilient configuration.
Selecting no buffer means the output layer is directly coupled with the input layer. This is the safest and lowest latency option since acknowledgements from at-least-once protocols can be propagated all the way from the output protocol to the input protocol.
If the output layer is hit with back pressure it will propagate all the way to the input layer, and further up the data stream. If you need to relieve your pipeline of this back pressure consider using a more robust buffering solution such as Kafka before resorting to alternatives.