Multi-Raft at Global Scale

The Challenge

Global consensus is hard. Traditional Raft requires a single leader, creating cross-region write latency proportional to RTT.

Multi-Raft with Leader Leases

We partitioned the keyspace into regional shards, each running an independent Raft group with:

• Leader leases to avoid stale reads
• Cross-region replication only for disaster recovery
• Automatic failover in under 3 seconds

Results

Combining Multi-Raft with leader leases, we achieved write latencies within 5ms of local-region RTT, with failover completing in under 3 seconds — even across a global footprint.

The key insight: don't fight the speed of light; route around it.