Definition
Geo-replication is the practice of replicating data across geographically separated data centers so that copies exist in multiple regions simultaneously.
How it works
Cross-region network latency is significant—US East to EU West is roughly 80–100ms round-trip. This makes synchronous replication impractical for most workloads, so geo-replication is typically asynchronous. A write commits locally first, then replicates to remote regions in the background.
Active-passive setups designate one region as primary for writes. All other regions receive replicated data and serve reads only. AWS RDS cross-region read replicas work this way. Failover promotes a replica to primary if the primary region goes down, but there is a window of data loss equal to the replication lag (typically seconds).
Active-active setups allow writes in any region. CockroachDB and Azure Cosmos DB support this model. The hard problem is conflict resolution—if two regions update the same row simultaneously, the system needs a deterministic rule. Cosmos DB uses Last Writer Wins (LWW) with timestamps. CockroachDB uses serializable transactions with global consensus, paying a latency cost for cross-region coordination. Cassandra uses vector clocks or LWW at the column level.
Some systems offer region-pinning: data is assigned to a home region for low-latency writes, with read replicas elsewhere. Google Spanner lets you configure placement policies that keep user data in specific geographies, satisfying GDPR requirements while still providing global reads.
When to use it
- Serving a global user base from a single US region adds 200ms+ latency for Asian users. Geo-replicate to
ap-southeast-1to cut read latency to under 20ms. - Meeting regulatory requirements: EU user data must stay in EU data centers, but US services need read access for analytics.
- Disaster recovery: if
us-east-1goes down (it has, multiple times), a replica inus-west-2can take over within minutes. - S3 Cross-Region Replication for serving static assets from the nearest bucket to the user.
Trade-offs
Gains: Lower read latency for global users, high availability across region-level failures, and compliance with data residency regulations.
Costs: Asynchronous replication means stale reads—a user who writes in US East and immediately reads from EU West may not see their own write. Cross-region bandwidth costs add up: replicating 1 TB/day across AWS regions costs roughly $20/day in data transfer fees. Active-active conflict resolution adds application complexity. And more replicas mean more storage cost—3 regions tripling your storage bill.
Example
Configuring CockroachDB zone constraints for geo-pinned data:
ALTER TABLE users CONFIGURE ZONE USING
num_replicas = 5,
constraints = '{"+region=us-east-1": 2, "+region=eu-west-1": 2, "+region=ap-southeast-1": 1}',
lease_preferences = '[[+region=us-east-1]]';
Related terms
Geo-replication often triggers rebalancing when regions are added or removed. In event streaming architectures, exactly-once semantics become harder to guarantee across geo-replicated Kafka clusters.