Disaster Recovery in the Cloud
Implementing backup and restore, pilot light, warm standby and active-active on AWS, Azure and Google Cloud with the managed services that support each.
See Disaster Recovery for RPO/RTO concepts and strategy definitions.
Service support per strategy
| Strategy | AWS | Azure | Google Cloud |
|---|---|---|---|
| Backup and restore | AWS Backup (cross-region, cross-account, vault lock), S3 replication | Azure Backup (immutable vaults), geo-redundant storage | Backup and DR Service, dual/multi-region Cloud Storage |
| Pilot light | Aurora Global Database / RDS cross-region replicas; infra via IaC | Azure SQL geo-replication / failover groups; Site Recovery | Cloud SQL cross-region replicas |
| Warm standby | Scaled-down ASGs/EKS in second region; Route 53 failover | Paired region deployment; Front Door failover | Regional MIGs/GKE; global LB failover |
| Active-active | DynamoDB global tables, Aurora DSQL, Route 53 latency routing | Cosmos DB multi-region writes, Front Door | Spanner multi-region, global LB |
| VM replication | AWS Elastic Disaster Recovery | Azure Site Recovery | (partner tools / Backup and DR) |
Example: warm standby on AWS
flowchart LR R53[Route 53 failover routing<br/>health checks] --> P[Primary region<br/>full capacity] R53 -. on failure .-> S[Standby region<br/>min capacity, auto scales up] P --> AP[(Aurora primary)] AP -- Global Database --> AS[(Aurora secondary)] S --> AS P --> S3P[(S3)] -- CRR --> S3S[(S3 replica)]
Failover runbook: confirm regional impact → promote Aurora secondary (managed failover or switchover) → scale standby compute → shift DNS → verify → communicate. Practise it.
Pitfalls
- Dependencies left in one region (identity provider, secrets, container registry, CI/CD, DNS management).
- Quotas in the DR region too low to scale up.
- Untested runbooks; data replication lag not monitored.
- Global services treated as immune (control-plane dependencies during regional events).
Interview questions and answers
How do you verify your DR region can actually take the load?
Regular failover exercises with production-like traffic, quota checks, pre-provisioned capacity reservations for critical tiers, and measuring achieved RTO/RPO each time.
Is multi-AZ enough?
For many workloads, yes: AZs are physically separate data centers. Multi-region is needed for regional disasters, regulatory requirements, or very high availability targets, at significant cost.