Multi-Cloud Architecture

Learning Objectives

• Understand the trade-offs of multi-cloud vs. single-cloud vs. hybrid-cloud strategies
• Design workload portability using cloud-agnostic abstractions
• Implement federated cluster management across multiple cloud providers
• Configure cross-cloud networking, DNS failover, and traffic routing
• Document architecture decisions with explicit trade-off analysis

Requirements

You are required to design and partially implement a multi-cloud architecture, making and documenting every architectural decision:

1. Architecture Decision and Documentation
   • Write an Architecture Decision Record (ADR) for the multi-cloud strategy:
     • Why multi-cloud? Evaluate: resilience, cost optimization, regulatory requirements, avoiding lock-in
     • When does multi-cloud make sense - and when it does NOT (operational complexity, cost of abstraction)
     • Chosen strategy: Active-Active, Active-Passive, or Workload Distribution
     • Accepted trade-offs and long-term operational cost
   • Create an architecture diagram showing the full multi-cloud topology
   • Define which workloads run where and why
2. Cluster Federation
   • Provision two Kubernetes clusters in different cloud providers (e.g., GKE + EKS, or EKS + AKS)
   • Use Cluster API or cloud-specific tools to manage both clusters from a single control plane
   • Configure Kubefed or use ArgoCD ApplicationSets to deploy to both clusters simultaneously
   • Verify the same application runs identically on both clusters
   • Document the federation model and its operational overhead
3. Cloud-Agnostic Abstractions
   • Use Crossplane or Terraform with provider-agnostic modules to provision:
     • Object storage (S3 on AWS, GCS on GCP, or Blob on Azure)
     • Managed Kubernetes clusters
     • DNS zones
   • The same Terraform/Crossplane code must work against both providers with minimal changes
   • Document which abstractions are genuinely portable and which required provider-specific code
4. Cross-Cloud Networking and Traffic Management
   • Configure a global load balancer or DNS-based failover (e.g., Route53 latency routing, Cloudflare Load Balancing)
   • Implement health checks that automatically redirect traffic to the healthy cluster
   • Simulate a full cluster outage in one cloud and verify traffic fails over automatically
   • Measure and document: failover detection time, DNS propagation time, user impact window
5. Observability Across Clouds
   • Configure a centralized observability stack that receives metrics and logs from both clusters
   • Use Thanos or Victoria Metrics for cross-cluster Prometheus federation
   • Create a unified Grafana dashboard showing both clusters side by side
   • Define alerts that trigger on the aggregate, not per-cluster

Stretch Goals

• Implement service mesh federation (Istio multi-cluster or Linkerd multi-cluster) for encrypted cross-cluster communication
• Configure Velero to back up one cluster and restore to the other
• Evaluate and document the total cost of the multi-cloud setup vs. a single-cloud equivalent

Deliverables

• ADR documenting the multi-cloud decision with explicit trade-offs
• Architecture diagram of the full topology
• Terraform/Crossplane code for cloud-agnostic resource provisioning
• Cluster federation configuration with deployment evidence on both clusters
• Failover test documentation: scenario, commands, timing, and results
• Unified observability dashboard screenshot

References

Books

• Cloud Strategy - Gregor Hohpe
• Kubernetes Patterns - Bilgin Ibryam & Roland Huß

Courses and Guides

• Google Cloud: Multi-Cloud Architecture Guide
• AWS: Multi-Region Application Architecture
• CNCF Multi-Cloud Working Group
• Crossplane Documentation
• Cluster API Documentation

Articles

• The Pragmatic Engineer: Multi-Cloud at Spotify
• Thanos: Highly Available Prometheus
• Cloudflare Load Balancing

Once you complete this task you will design infrastructure that survives a full cloud provider outage - and more importantly, you will know when multi-cloud is the right answer and when it is unnecessary complexity.