A replica is only as current as its lag

That a copy exists in another region is usually true. That it is a copy of now holds only to within the replication lag, and a failover plan inherits that lag as lost data.

Lag is not a setting. On a normal day it may be a few milliseconds, and AWS quotes typical cross-region lag under a second for Aurora global databases. Its floor is the round trip between replicas, and load, lost packets and slow replicas stretch it further, so an asynchronous system cannot promise an upper bound. A dashboard showing it low describes today, and incidents are not today.

It shows up in two places. During failover it is the data missing after promotion. In normal running it is what a user sees when a read lands on a replica straight after writing, which is how Read-your-writes breaks quietly. Some guarantees also end where the region does: Strong consistency can stop at the region boundary.

And lag only measures what replication carries. A perfectly current copy is perfectly current with a bad migration too: Replication copies your mistakes too.