Multi-Region Latency Spike Diagnosis

Overview

Scope before you touch anything

Multi-region latency issues are hard to diagnose because they span routing, replication and infrastructure layers. Systematic scoping comes first: which region, which endpoints, which percentage of users.

> [!IMPORTANT] Compare the affected region against a healthy one at every layer. The delta tells you where to look.

The layers to check

• DNS and routing — Failing health checks, a routing policy sending users to the wrong region
• Edge — Users not hitting the nearest CloudFront edge, slow origin fetches
• Cross-region calls — The affected region calling another region synchronously
• Regional data tier — Read-replica lag, cache latency
• Compute — CPU throttling, failing instances in one AZ pushing traffic cross-AZ

Designing to avoid it

• Routing — Latency-based routing with health-check failover, or anycast via Global Accelerator
• Data — Multi-region replication with a clear consistency model
• State — No session affinity; sessions in a shared store any region can read
• Cost — Active-active costs 2-3x active-passive; justify it with near-zero RPO and seconds of RTO

Interview questions

• A multi-region deployment is causing latency spikes for users in one region but not others. How do you isolate and resolve it?
• How do you design a multi-region active-active system and what are the consistency tradeoffs?

More SRE topics

• Live Capacity Planning Scenario
• Load Shedding
• Load Testing
• Logging Best Practices
• Modern Observability Stack
• Multi-cloud Active-Active
• NALSD — Non-Abstract Large System Design
• On-call Structure
• On-call Sustainability
• Pager Discipline
• Performance Optimization
• RED, USE & Four Golden Signals