Design VM-based reliability (publish-consume vs publish with reply, persistent queues participating in transactions, redelivery on failure) and predict message durability/loss behavior on worker restart or crash in each deployment option.