Custom Software & Cloud Infrastructure
Cloud Architecture Selection: Comparing AWS, Google Cloud, and Azure for Multi-Region Deployments
Compare AWS, GCP and Azure across global data services, networking, identity and failover design, with a decision workflow for multi-region topologies.
Multi-region is where cloud comparisons stop being about feature checklists. Every major provider offers compute, object storage, managed databases and a global load balancer. What separates them is how they handle the three genuinely hard problems: data replication topology, cross-region networking cost and latency, and identity plus failover semantics during a partial regional failure.
The business problem is usually stated as "we need high availability." That phrasing hides the decision that matters: are you building active-passive with a recovery objective measured in minutes, or active-active with globally distributed writes? The first is an infrastructure exercise. The second is a data-modelling exercise that changes application code, because conflict resolution and read-your-writes guarantees become your problem.
This analysis gives CTOs and platform leads a selection framework: workload classification, provider-specific strengths in global data services, networking and egress economics, the failover design that actually gets tested, and a 90-day plan to reach a verified regional recovery capability. The takeaway: choose the provider whose global data primitive matches your consistency requirement, and let compute and networking follow.
Classify the Workload First
Provider choice is downstream of these four determinations.
- Consistency requirement. Does the application require strong consistency across regions (financial ledgers, inventory) or is bounded staleness acceptable (catalogues, content, analytics)?
- Write distribution. Single-region writes with global reads is straightforward. Multi-region writes require either a globally consistent database or explicit conflict handling.
- Recovery objectives. Define RTO and RPO numerically. "Near zero" is not a specification; 15 minutes RTO and 1 minute RPO is.
- Data residency. Regulatory constraints on where records may live often eliminate architectures before performance does.
Only after these are fixed does provider comparison become meaningful.
Provider Strengths in Global Data
- AWS offers the broadest set of primitives: Aurora Global Database for low-lag cross-region replication with fast promotion, DynamoDB Global Tables for multi-region multi-writer with last-writer-wins semantics, and S3 cross-region replication. The tradeoff is assembly — you compose and operate the topology, and IAM plus VPC design complexity is real.
- Google Cloud leads when strong global consistency is a hard requirement. Spanner provides externally consistent multi-region writes with synchronous replication, removing conflict resolution from application code at higher baseline cost. Its global anycast load balancing and single-VPC-across-regions networking model reduce topology overhead significantly.
- Azure is strongest in enterprise identity integration and configurable consistency. Cosmos DB exposes five consistency levels — from strong to eventual — letting teams tune per-workload, with multi-region writes available. Entra ID integration and hybrid connectivity make it the pragmatic default where Microsoft estate and compliance frameworks already dominate.
Networking and Cost Realities
Cross-region traffic is where multi-region budgets fail. Three rules apply universally:
- Keep chatty traffic intra-region. Every synchronous cross-region call adds tens of milliseconds; a request that crosses regions three times is architecturally broken.
- Model egress before committing. Inter-region replication and internet egress are metered; replicating high-write datasets to three regions can cost more than the compute serving them.
- Terminate at the edge, compute near data. Use anycast or global load balancing for TLS termination and static delivery, then route dynamic requests to the region holding the authoritative data.
Failover Design That Gets Tested
Untested failover is documentation, not capability.
- Externalise all state. Compute must be stateless and reproducible from infrastructure-as-code in any target region.
- Set health checks on business transactions, not TCP liveness. A database read failure must fail the check.
- Keep DNS TTLs low, or prefer anycast so failover does not depend on resolver behaviour.
- Pre-provision minimum capacity in the secondary region; cold regions fail scale-out precisely when demand spikes.
- Run scheduled regional evacuation drills — quarterly, in production, with measured RTO and RPO.
Provider Comparison for Multi-Region
| Capability | AWS | Google Cloud | Azure |
|---|---|---|---|
| Strongly consistent global DB | Limited (Aurora Global read replicas) | Spanner (external consistency) | Cosmos DB strong within region set |
| Multi-region write model | DynamoDB Global Tables (LWW) | Spanner (synchronous) | Cosmos DB multi-write, tunable |
| Global load balancing | Route 53 + CloudFront + ALB | Global anycast LB | Front Door + Traffic Manager |
| Network topology | Per-region VPC plus peering | Global VPC | VNet plus peering / Virtual WAN |
| Identity model | IAM policies, fine-grained | IAM plus resource hierarchy | Entra ID, strongest enterprise SSO |
| Managed Kubernetes | EKS | GKE (most mature autopilot) | AKS |
| Region breadth | Largest | Strong | Strong, deep in regulated markets |
| Typical strength | Service breadth and maturity | Data consistency and networking | Enterprise identity and compliance |
Reference Blueprint
Active-passive with a promotable global database, expressed as portable infrastructure intent:
topology: active-passive
objectives:
rto_minutes: 15
rpo_seconds: 60
regions:
primary:
id: us-east
compute: { min_instances: 6, autoscale_max: 40 }
database: { role: writer, engine: managed-postgres, backup: pitr }
secondary:
id: eu-west
compute: { min_instances: 2, autoscale_max: 40 } # warm, never cold
database: { role: reader, replication: async, promotable: true }
traffic:
entry: global-anycast-load-balancer
health_check:
path: /api/public/health
asserts: [database_read, cache_write, queue_publish]
interval_seconds: 10
unhealthy_threshold: 3
failover: automatic_on_unhealthy_primary
data:
object_storage: { replication: cross-region, versioning: true }
secrets: { replicated: true, rotation_days: 90 }
observability:
slo: { availability: 99.95, latency_p95_ms: 400 }
drills: { cadence: quarterly, mode: production_evacuation }The health check contract is the important detail: asserting a database read, cache write and queue publish means a promoted region that cannot serve traffic is never declared healthy.
Strategic Action Plan: 90 Days
- Days 1–15 — Specify requirements numerically. Document consistency needs per dataset, write distribution, RTO/RPO targets and residency constraints. Model cross-region egress cost for the top three datasets.
- Days 16–35 — Prototype the data layer. Test the candidate global database primitive with production-shaped write volume, measuring replication lag, promotion time and read-your-writes behaviour.
- Days 36–55 — Codify infrastructure. Express both regions in infrastructure-as-code with identical modules, replicate secrets, and provision warm minimum capacity in the secondary.
- Days 56–75 — Implement transaction-level health checks and failover. Wire global load balancing, low TTLs or anycast, and automated promotion with guardrails against flapping.
- Days 76–90 — Drill and measure. Execute a production regional evacuation, record actual RTO and RPO against targets, fix the gaps, and schedule the drill quarterly.
Frequently Asked Questions
Is active-active always better than active-passive?
No. Active-active removes failover time but requires either a globally consistent database or application-level conflict resolution, and it doubles the surface for data-integrity defects. For most enterprise workloads, active-passive with a warm secondary, asynchronous replication and a tested 15-minute RTO delivers the required availability at materially lower complexity and cost.
Which provider is best for multi-region PostgreSQL-style workloads?
If you need strong global consistency with SQL semantics and want conflict handling out of application code, Google Cloud Spanner is the strongest primitive. If asynchronous replication with fast promotion is acceptable, AWS Aurora Global Database is mature and cost-effective. Azure suits organisations standardised on Entra ID that can model data in Cosmos DB with tunable consistency.
How do we control multi-region cloud costs?
Reduce cross-region data movement first: keep synchronous calls inside a region, replicate only datasets that failover genuinely needs, and compress or batch replication traffic. Then right-size warm standby capacity rather than running symmetric full-scale regions, apply committed-use or savings-plan discounts to steady-state compute, and attribute egress cost per service so ownership is clear.
What actually needs to be tested in a disaster recovery drill?
Promote the secondary database, shift production traffic, and complete real business transactions in the secondary region while measuring elapsed recovery time and data loss. Verify secrets availability, queue drain behaviour, scheduled jobs, third-party allowlists and observability continuity. A drill that only checks whether instances start does not validate recoverability.