Writing Everywhere
When users span continents, a single write leader makes far away writes slow. Multi region designs spread leadership and replicas to balance write latency, read latency, and survival of a whole region failing.
Common Survival Goals
- Zone survival: tolerate one datacenter loss; replicas can sit in one region for low latency.
- Region survival: tolerate an entire region loss; the Raft majority must span at least three regions, so each commit pays a cross region round trip.
You cannot get both region survival and single region commit latency at once, because surviving a region loss requires a majority outside any one region.
Placement Strategies
- Geo partitioned tables keep each row leader near its users for local writes.
- Global tables replicate everywhere with fast reads but pay write latency, suited to rarely changed reference data.
- Regional by row lets each row choose its home region.
The right mix depends on whether your workload is write local, read global, or must survive a regional outage.
Key idea
Multi region writes force a tradeoff: surviving a whole region loss needs a majority across three regions and cross region commit latency, while local low latency writes give up region survival.