Interview question

Your backend consumes queue batches containing four or five messages, and three or four backend service instances are running. How would you support parallel processing while preserving sequential processing where required?