message queue idempotency: how do you handle retries?

Hey everyone, I'm trying to figure out how to handle message queue idempotency in a distributed setup. Basically, if a message consumer messes up halfway through, or if a message gets sent again, how do you stop things from happening twice? I've seen some ideas, like using unique message IDs and having the consumer check them, or using transactional outboxes. What methods have worked well for you and scale up nicely when you have retries?