Consume messages from Kafka in batches of 20.
Instead of one-by-one. It's probably more efficient, and seems to partially mitigates T2034.