Idempotency in Distributed Systems: Essential or Overkill?

hey everyone, i've been thinking a lot about idempotency lately, especially when designing apis for distributed systems. we all know it's crucial for things like message queues and retries to prevent duplicate operations. but how far do we really need to go, are there scenarios where implementing full idempotency across all services becomes more of a burden than a benefit, what are your experiences with making systems idempotent, and when do you decide it's 'good enough'