Eight Seconds Gone

An order service replicates asynchronously from its primary region to a standby, and monitoring shows replication lag hovering around 8 seconds under peak load. The primary region loses power and the team promotes the standby. What should they expect?

The promoted region refuses writes until the original primary comes back and streams the missing segment of its log

Roughly the last 8 seconds of committed orders are missing from the promoted region, an RPO the business has to reconcile

No data is lost, because the primary acknowledged those commits, and an acknowledgement implies the standby had already received them

No data is lost, but the promoted region serves stale reads for about 8 seconds while it applies the backlog it already holds