Handling duplicate messages in a distributed system can be tricky.

Lately, I've been thinking about how to deal with potential duplicate messages when using message queues in a microservices architecture. We're using Kafka, and while it offers at-least-once delivery, we sometimes see duplicate events processed by consumers. What are your strategies for idempotency? We've considered unique message IDs and checking against a distributed cache or database, but are there other common patterns or pitfalls I should be aware of? Trying to avoid reprocessing critical operations.