When is idempotency truly critical for message queue consumers?

We all know idempotency is a 'best practice' for consumers processing messages from a queue, especially in distributed systems where retries and at-least-once delivery are common. But let's be real, sometimes the effort to make an operation truly idempotent can be significant. So, in your experience, when do you absolutely draw the line and say 'this consumer MUST be idempotent, no exceptions'? Are there specific types of operations or data models where the cost of non-idempotency far outweighs the implementation effort? Or conversely, scenarios where you've consciously decided against full idempotency because the downstream impact of a duplicate was negligible and the complexity wasn't worth it?