Skip to content

Multi-Region Architecture

1. Decision

Choose Multi-Region Architecture (Active-Passive or Active-Active across geographical cloud regions) for global disaster recovery, compliance with sovereign data residency laws, and sub-50ms latency for globally distributed end users.

2. 30-Second Answer

A Multi-Region Architecture deploys application compute and database infrastructure across physically separated cloud geographic regions (e.g., AWS us-east-1, eu-west-1, ap-southeast-1). Active-Passive Multi-Region maintains a primary active region serving 100% of traffic while asynchronously replicating database state to a passive standby region for disaster recovery (DR). It is simple and avoids cross-region write conflicts, but incurs failover recovery time (RTO/RPO). Active-Active Multi-Region serves live read/write traffic simultaneously in all regions. It delivers ultra-low regional user latency (\(<30\text{ms}\)) and continuous availability, but forces complex distributed consensus, asynchronous cross-region replication lag, and conflict resolution (Last-Writer-Wins, CRDTs).

3. Decision Matrix

Option Best For Avoid When Latency Scalability Consistency Operational Complexity
Active-Passive Multi-Region Disaster Recovery (DR), business continuity compliance, global applications accepting 50-100ms cross-region RTT Zero-RPO/RTO requirements, ultra-low latency requirements for global users far from primary region High for users far from Primary region (100-200ms RTT); Low within Primary region High (Compute scales per region; DB writes bounded by Primary region) Strong (Primary region); Asynchronous lag to Passive standby Medium
Active-Active Multi-Region Global latency-critical web apps, zero-downtime regional failover, high availability, regional data residency Strict single-node financial ACID transactions, tight engineering budgets Ultra-Low (<30ms regional user RTT) Extremely High (Global scale across multiple independent cloud datacenters) Eventual Consistency; Requires conflict resolution (CRDT / LWW) Extremely High (Cross-region replication, split-brain mitigation, global DNS)

4. Selection Rules

Requirement Best Choice Rationale / Example
Disaster Recovery (DR) Compliance Active-Passive Primary region handles traffic; standby region receives async DB backups/WAL replication
Global Ultra-Low Latency (<30ms RTT) Active-Active Route users to nearest geographic region via Anycast / Route 53 Geolocation DNS
Strict Data Sovereignty (GDPR / India Data Law) Multi-Region Local Pinning User data stored strictly within sovereign regional database boundaries
Zero Downtime During Total Region Outage Active-Active If us-east-1 suffers a total outage, global DNS shifts traffic seamlessly to us-west-2
Conflict-Free Distributed Replicated Data Active-Active (CRDTs / CockroachDB) Use Conflict-Free Replicated Data Types or Spanner/CockroachDB Paxos consensus

5. Tradeoffs

  1. Global Low Latency vs Distributed Cross-Region Write Consistency
  2. Gain (Active-Active): Sub-30ms user response time worldwide.
  3. Lose (Active-Active): Writing to local region creates asynchronous replication lag window to other regions.
  4. Disaster Recovery Security vs Operational Cost & Complexity
  5. Gain (Multi-Region): System survives complete datacenter destruction or cloud vendor regional outage.
  6. Lose (Multi-Region): 2x to 3x infrastructure billing costs; complex multi-region CI/CD deployment pipelines.
  7. Regional Autonomy vs Cross-Region Data Conflicts
  8. Gain: Regional outages are isolated; region can operate autonomously if cross-region link breaks.
  9. Lose: Dual-writing same entity in 2 regions simultaneously causes conflicting write states.
  10. Anycast / Geo-DNS Routing Speed vs DNS Propagation Lag
  11. Gain: Route 53 Latency-Based routing directs users to optimal regional health endpoints.
  12. Lose: DNS TTL caching delays failover traffic shifts when a region fails suddenly.
  13. Asynchronous Cross-Region WAL Replication vs Synchronous Latency Penalty
  14. Gain (Async): Local writes commit instantly without waiting for cross-ocean network RTT.
  15. Lose (Async): Regional failure results in data loss of un-replicated write bytes (RPO > 0).

6. Interview Counter Questions

  1. What is RTO (Recovery Time Objective) and RPO (Recovery Point Objective) in Multi-Region Disaster Recovery?
  2. How does AWS Route 53 Latency-Based Routing and Geolocation Routing work?
  3. How do you resolve write conflicts in an Active-Active Multi-Region architecture (Last-Writer-Wins vs CRDTs)?
  4. Why is synchronous cross-region database replication (e.g. N. Virginia to Frankfurt) rarely used in OLTP?
  5. How does Google Spanner achieve global multi-region ACID transactions using TrueTime atomic clocks?
  6. What is Split-Brain syndrome in multi-region failover and how do you prevent it?
  7. How do DynamoDB Global Tables and Cassandra Multi-Region Clusters handle multi-region replication?
  8. How do you manage static assets and CDN (CloudFront/Cloudflare) in a Multi-Region setup?
  9. What is Anycast IP Routing and how does AWS Global Accelerator speed up multi-region traffic?
  10. How do you execute database schema migrations cleanly across a Multi-Region Active-Active deployment?

7. Gold Standard Answers

  1. RTO: Maximum acceptable downtime duration following an outage (e.g. recover within 15 mins). RPO: Maximum acceptable data loss volume measured in time (e.g. lose max 5 seconds of un-replicated data).
  2. Latency-Based Routing uses global AWS network latency probes to route users to the region with lowest RTT. Geolocation Routing maps user IP addresses to specific geographical regions for data sovereignty compliance.
  3. (1) Last-Writer-Wins (LWW): Overwrite based on wall-clock timestamp (risks clock drift data loss). (2) CRDTs (Conflict-Free Replicated Data Types): Mathematically merge concurrent updates. (3) User Pinning: Route specific user write traffic strictly to 1 home region.
  4. Cross-ocean network round-trip time (RTT) between US and Europe is ~70-100ms. Synchronous 2PC cross-region commits force every write to wait >100ms, severely degrading application throughput SLAs.
  5. Google Spanner uses TrueTime API backed by GPS receivers and atomic clocks in every datacenter. TrueTime provides a bounded clock uncertainty (\(\epsilon \approx 1-7\text{ms}\)), enabling global External Consistency (Serializable) without cross-region locks.
  6. Split-Brain occurs when cross-region network links break, causing both regions to assume the other died and promote local DBs to Primary. Prevent by using odd-numbered regional quorum voting (e.g. 3 regions) via Raft/Paxos consensus.
  7. DynamoDB Global Tables uses asynchronous multi-master replication across regions with Last-Writer-Wins conflict resolution. Cassandra uses masterless peer-to-peer ring replication with configurable LOCAL_QUORUM consistency.
  8. Static assets reside in S3 buckets in primary regions, fronted by CloudFront/Cloudflare CDN. Edge PoPs fetch and cache assets globally; if an origin region fails, CDN origin groups failover to S3 in the secondary region.
  9. AWS Global Accelerator assigns static Anycast IP addresses routed to the nearest AWS edge location over BGP. User traffic enters AWS's private global fiber network immediately, bypassing congested public internet hops.
  10. Use Multi-Phase Non-Breaking Schema Evolution: (1) Add new column as nullable in DB across all regions, (2) Deploy code reading old+new columns across all regions, (3) Deploy code writing to new column, (4) Drop old column after all regions synchronize.

8. Common Mistakes

  • Attempting to run synchronous database commits across global cloud regions, causing 100ms+ write latencies.
  • Deploying Active-Active Multi-Region without user region pinning or conflict resolution, corrupting user data via concurrent multi-region writes.
  • Relying purely on DNS failover without setting short DNS TTLs (e.g., 60s), leaving users stuck on a dead region endpoint.
  • Ignoring Cross-Region Data Egress costs, causing massive surprise AWS data transfer billing.
  • Testing Disaster Recovery failover plans only during actual production outages rather than conducting periodic GameDay chaos drills.

9. Keywords

Multi-Region, Active-Passive, Active-Active, RTO, RPO, Disaster Recovery (DR), Route 53, Geolocation Routing, Anycast, AWS Global Accelerator, CRDTs, Last-Writer-Wins, Split-Brain, TrueTime, Spanner, DynamoDB Global Tables, Non-Breaking Schema Evolution.

10. Cheat Sheet

  • Active-Passive: 1 Active region (100% traffic) + 1 Standby region (Async DB replication). Simple, no write conflicts, RPO/RTO > 0.
  • Active-Active: Multi-region live traffic. Ultra-low regional user latency (<30ms), high availability. Requires conflict resolution & user pinning.
  • RTO vs RPO: RTO = How long system is down; RPO = How much data loss occurred (measured in time).
  • User Pinning Rule: Route all requests for user_123 to their designated "Home Region" to eliminate cross-region write conflicts.
  • Network Speedup: Use AWS Global Accelerator (Anycast IP) to route user traffic onto AWS private fiber network at the nearest edge.
  • Schema Migration: Always use 4-phase non-breaking schema deployments across multi-region environments.

11. Related Decisions