Hybrid Cloud and Multi-Cloud
When hybrid and multi-cloud make sense, architectural patterns (portable platform, best-of-breed, DR across clouds), control planes like Anthos and Azure Arc, and the real costs.
Definitions
- Hybrid cloud: on-premises (or private cloud/edge) combined with public cloud, operating as one estate.
- Multi-cloud: using two or more public cloud providers.
Valid reasons
| Reason | Example |
|---|---|
| Data residency / latency / sovereignty | Factory systems on-prem, analytics in cloud; country with only one provider region |
| Mergers and acquisitions | Acquired company runs on another cloud |
| Best-of-breed services | Transactions on AWS, analytics on BigQuery, AI on Azure OpenAI |
| Regulatory concentration risk | Financial regulators requiring credible exit plans |
| Commercial leverage | Negotiation, but rarely worth the operational cost on its own |
Patterns
flowchart LR
subgraph Portable platform
K1[Kubernetes on AWS] --- GIT[GitOps + policy + observability<br/>single control plane]
K2[Kubernetes on Azure] --- GIT
K3[Kubernetes on-prem] --- GIT
end
subgraph Best of breed
APP[Core app on Cloud A] -- events --> DATA[Analytics on Cloud B]
end- Portable application platform: Kubernetes everywhere with a common control plane (GKE Enterprise/Anthos, Azure Arc, Red Hat OpenShift, Rancher) and GitOps.
- Best-of-breed per workload: each workload on one cloud, integrated via APIs/events; avoids lowest-common-denominator design.
- Cross-cloud DR: rarely justified given complexity; used by some regulated firms for critical services.
- Edge/hybrid extensions: AWS Outposts, Azure Local (formerly Azure Stack HCI), Google Distributed Cloud.
Costs and risks
- Duplicate skills, tooling, security controls and landing zones.
- Egress charges between clouds.
- Lowest-common-denominator architecture if portability is forced everywhere.
- Identity federation and consistent policy across providers.
[!CTO] Prefer "multi-cloud by design for specific reasons" over "multi-cloud by default". Portability has a price; pay it only where risk or value justifies it.
Interview questions and answers
How do you design for cloud exit without full multi-cloud?
Use open formats and standards (Kubernetes, PostgreSQL, Iceberg/Parquet, OpenTelemetry, Terraform), isolate provider-specific services behind interfaces where cheap, keep data exportable, document an exit plan with estimated effort and test parts of it.
What is Azure Arc used for?
Projecting servers, Kubernetes clusters and data services running outside Azure (on-prem or other clouds) into Azure Resource Manager so they can be governed with Azure Policy, monitored and managed centrally.