Exactly-once delivery in distributed systems: myth or achievable?

the concept of exactly-once delivery in distributed systems feels like a holy grail. textbooks always say it's impossible, yet kafka advertises exactly-once semantics for stream processing. i'm trying to wrap my head around what kafka's guarantee *actually* means in practice. does it hold up when messages cross system boundaries, say from kafka to an http service, then to a database? my understanding is that once you leave the kafka ecosystem, you're back to at-least-once. practically, we implement idempotency at our service boundaries to handle retries and ensure no duplicate processing. is that the only real world solution for true exactly-once behavior end-to-end, or am i missing something about how kafka's guarantees extend beyond its internal world?