An acknowledged write is a promise to the user

When the service replies that an order went through, the user has no reason to send it again, and may act on it. What they rely on is the acknowledgement, so the durability question for a regional outage is exact: at the moment of replying, where did the write exist?

What an acknowledgement means is a database setting, and defaults differ. DynamoDB replies once the partition leader has made the change durable, then copies it to the other replicas. MongoDB has waited for a majority of the replica set by default since version 5.0; earlier versions waited only for the primary. Redis replicates asynchronously unless the client asks it to wait, and even then the copies may exist only in replica memory.

Synchronous replication makes the reply wait for another copy. Asynchronous replication replies first and copies afterwards, so a write acknowledged just before a region fails can exist in that region alone.

The promise has a second edge. A client whose request timed out cannot tell whether it was applied, and its retry must not become a second order. Making that safe is the job of Idempotency keys make retries safe.