Read Local Write Global
Replicating read-only state to every edge location while directing writes to a single authoritative region, which is the arrangement most edge data stores actually provide.
Edge compute is only fast if its data is nearby, and data replicated to two hundred locations cannot be strongly consistent without coordination that would cost more than the distance saved.
The workable arrangement separates the paths. Reads are served from a local replica at single-digit milliseconds. Writes are forwarded to one authoritative region and propagate outward, typically within a second but without a guarantee.
The consequence to design for is read-your-own-writes. A user who changes something and immediately reloads may read a replica that has not yet received the change, and will conclude the save failed. The mitigations are ordinary: pin that user's session to the write region briefly, read from the authoritative region after a write, or reflect the change from local client state rather than refetching.
The data that suits this shape is read-heavy, tolerant of a second's staleness and not subject to a global invariant — configuration, feature flags, routing rules, content, personalisation attributes. Anything requiring uniqueness, a balance, or an atomic decrement belongs in one place, and no amount of replication changes that.