Skip to content
1 min read By Vamsi Karuturi · Senior Backend Engineer at Salesforce

Multi-Region Architecture

Real Incident: AWS US-EAST-1 Outage (December 2021)

A networking issue in AWS's primary region took down thousands of services for 7+ hours — including Ring doorbells, Disney+, Tinder, and parts of AWS's own console. Companies with multi-region failover (Netflix, Slack) continued operating seamlessly. Single-region = single point of failure. Multi-region is the only path to true high availability.

Why This Comes Up in Interviews

Any system design targeting 99.99%+ availability or serving global users needs multi-region discussion. Interviewers want to hear:

  • Active-active vs active-passive trade-offs
  • Data replication and consistency challenges across regions
  • How you handle failover (DNS, traffic manager, health checks)
  • Cost vs availability trade-off decisions

Single-Region vs Multi-Region

Aspect Single Region Multi-Region
Availability 99.9% (8.7h downtime/year) 99.99%+ (52 min/year)
Latency High for distant users Low for all users
Blast radius Region outage = total outage Region outage = graceful failover
Complexity Simple Significantly complex
Cost 1x 2-3x (duplicate infrastructure)
Data Single source of truth Replication + conflict resolution

Multi-Region Patterns

1. Active-Passive (Warm Standby)

%%{init: {'theme': 'base', 'themeVariables': {'fontSize': '13px', 'fontFamily': 'Inter, -apple-system, sans-serif'}, 'flowchart': {'nodeSpacing': 30, 'rankSpacing': 50, 'padding': 12, 'curve': 'basis'}, 'sequence': {'actorMargin': 60, 'messageMargin': 40}, 'class': {'padding': 12}}}%%
flowchart LR
    DNS[DNS / Route53] -->|All traffic| R1[Region 1: US-EAST<br/>ACTIVE]
    DNS -.->|Failover only| R2[Region 2: EU-WEST<br/>PASSIVE]
    R1 -->|Async replication| R2

    style R1 fill:#22c55e,color:#fff
    style R2 fill:#6b7280,color:#fff
Aspect Detail
How Primary region handles all traffic; standby receives replicated data
Failover DNS health check fails → route traffic to standby (30-120s)
Data loss Possible (async replication lag) — RPO = seconds to minutes
Cost ~1.5x (standby running but idle)
Best for Cost-sensitive, can tolerate minutes of downtime

2. Active-Active (Both Serving Traffic)

%%{init: {'theme': 'base', 'themeVariables': {'fontSize': '13px', 'fontFamily': 'Inter, -apple-system, sans-serif'}, 'flowchart': {'nodeSpacing': 30, 'rankSpacing': 50, 'padding': 12, 'curve': 'basis'}, 'sequence': {'actorMargin': 60, 'messageMargin': 40}, 'class': {'padding': 12}}}%%
flowchart TD
    DNS[GeoDNS] -->|US users| R1[Region 1: US-EAST]
    DNS -->|EU users| R2[Region 2: EU-WEST]
    R1 <-->|Bidirectional replication| R2

    style R1 fill:#22c55e,color:#fff
    style R2 fill:#3b82f6,color:#fff
Aspect Detail
How Both regions serve traffic simultaneously, data replicated bidirectionally
Failover Automatic — healthy region absorbs failed region's traffic
Data conflicts Possible — need conflict resolution (LWW, CRDTs, merge logic)
Cost ~2x+ (both regions fully provisioned)
Best for Global low-latency, 99.99% availability requirements

3. Follow-the-Sun (Regional Primary)

%%{init: {'theme': 'base', 'themeVariables': {'fontSize': '13px', 'fontFamily': 'Inter, -apple-system, sans-serif'}, 'flowchart': {'nodeSpacing': 30, 'rankSpacing': 50, 'padding': 12, 'curve': 'basis'}, 'sequence': {'actorMargin': 60, 'messageMargin': 40}, 'class': {'padding': 12}}}%%
flowchart LR
    subgraph "Each user's writes go to their home region"
        US[US Region<br/>Primary for US users]
        EU[EU Region<br/>Primary for EU users]
        AP[APAC Region<br/>Primary for APAC users]
    end
    US <-->|Async replication| EU
    EU <-->|Async replication| AP
    AP <-->|Async replication| US

    style US fill:#3b82f6,color:#fff
    style EU fill:#22c55e,color:#fff
    style AP fill:#f59e0b,color:#fff

Each user has a "home region" based on geography. Writes go to home region, reads can be served from any region (eventually consistent).

Data Replication Strategies

Strategy Consistency Latency Data Loss Risk
Synchronous replication Strong High (cross-region RTT added to every write) None
Asynchronous replication Eventual None (fire and forget) Seconds of data on failover
Semi-synchronous Strong for 1 replica Moderate None if at least 1 replica ACKs

Conflict Resolution (Active-Active)

Strategy How Trade-off
Last Writer Wins (LWW) Highest timestamp wins Simple but can lose writes
CRDTs Mathematically mergeable data types Complex but lossless
Application-level merge Custom logic per entity Most flexible but most complex
Avoid conflicts Route user's writes to single region Simplest — no conflicts at all

Failover Mechanisms

%%{init: {'theme': 'base', 'themeVariables': {'fontSize': '13px', 'fontFamily': 'Inter, -apple-system, sans-serif'}, 'flowchart': {'nodeSpacing': 30, 'rankSpacing': 50, 'padding': 12, 'curve': 'basis'}, 'sequence': {'actorMargin': 60, 'messageMargin': 40}, 'class': {'padding': 12}}}%%
flowchart TD
    HC[Health Checker] -->|Every 10s| R1[Region 1]
    HC -->|Every 10s| R2[Region 2]
    HC -->|"R1 fails 3 checks"| DNS[Update DNS]
    DNS -->|"Route 100% to R2"| R2

    style HC fill:#f59e0b,color:#fff
Mechanism Failover Time How
DNS failover (Route53) 30-120s (DNS TTL) Health checks → remove unhealthy IPs
Global Load Balancer (Cloudflare, GCP GLB) 10-30s Anycast — instant rerouting
Traffic Manager (Azure) 30-60s DNS-based with fast health checks
Client-side failover <5s Client knows both endpoints, retries on failure

What to Replicate vs What to Rebuild

Replicate (stateful) Rebuild on failover (stateless)
Database (users, orders) Application servers
Object storage (S3 cross-region) Load balancers
Cache warming data Container orchestration
Message queue state DNS entries
Configuration/secrets CDN edge caches

Cost Analysis

Scenario: Web application serving 100K req/s

Setup Monthly Cost Availability Latency (global)
Single region $50K 99.9% (8.7h/year down) 200ms for distant users
Active-passive $75K (+50%) 99.95% (4.4h/year) 200ms (same, until failover)
Active-active (2 regions) $110K (+120%) 99.99% (52 min/year) <50ms for 80% of users
Active-active (3 regions) $160K (+220%) 99.999% (5 min/year) <50ms for 95% of users

Interview Cheat Sheet

Question Answer
"How to achieve 99.99% availability?" "Multi-region active-active. GeoDNS routes users to nearest region. If a region fails, traffic shifts to remaining regions in <30s. Data replicated asynchronously with conflict resolution."
"Active-passive vs active-active?" "Active-passive: simpler, cheaper, but 30-120s failover + possible data loss. Active-active: instant failover, better latency, but need conflict resolution for concurrent writes."
"How to handle data consistency?" "Route each user's writes to their home region (avoid conflicts). Cross-region reads are eventually consistent. For critical data (payments): synchronous replication to at least one other region."
"What about cost?" "Active-active is ~2x cost. Justified when downtime costs exceed infrastructure costs. For e-commerce doing $1M/hour, 1 hour downtime > yearly multi-region premium."
"Conflict resolution?" "Avoid conflicts by design: user writes go to home region. If unavoidable: LWW for simple data, CRDTs for counters/sets, application merge for complex business logic."