Synchronous replication moves the cost to latency

Making a write wait until another region confirms it means no acknowledged write lives in only one place. The price is paid on every write, on quiet days as well as bad ones.

Distance sets that price. Light in fibre takes about 28 ms one way from New York to London before routers add their share, and no software recovers it: The speed of light is a design constraint. Work that makes many database calls pays it many times. After GitHub's 2018 failover moved its primaries to the other side of the US, applications on the East Coast could not cope with a cross-country round trip on most of their database calls.

The second cost is availability. If every copy must confirm, one unreachable copy makes writes fail, so across two regions an outage in either stops writes in both, unless something decides to stop waiting. That something is a failover decision in disguise.

Consensus protocols such as Raft commit once a majority has acknowledged, which is how Consensus gives you one leader at a time and Quorum reads and writes overlap keep writing through the loss of a minority.