Cloud Foundations That Survive the Next Three Years
Architecture choices that keep platforms flexible as product scope, compliance, and traffic evolve.
Written byMuhammad Abdullah Chaudhary
Cloud sprawl usually starts with good intentions: one quick environment for a demo, a second for a client, a hand-configured server nobody dares touch. Three years later, no one can rebuild the platform from scratch.
The fix is a foundation that is opinionated where it matters and flexible where it doesn't. These are the choices we make on client platforms, and why.
The reference shape
Everything that defines the platform lives in a repository the client owns. Everything that changes production goes through the pipeline.
Prefer boring infrastructure
| Choose | Over | Because |
|---|---|---|
| Managed database | Self-run database on a VM | Backups, patching and failover come included |
| Terraform or similar | Clicking in a console | The environment can be rebuilt and reviewed |
| Containers on Kubernetes or serverless | Hand-configured servers | Portable between clouds and on-premise |
| One CI/CD system | Several scripts | One place to audit releases |
| Clear network boundaries | Flat networks | Limits the damage of any one mistake |
Novel infrastructure is fine when it solves a problem you actually have. Most product companies don't have that problem yet.
Boring infrastructure, defined as code, is what you can still rebuild in three years.
Own the definition, not just the servers
On our ERP migration for a German enterprise, the client's systems ran on vendor servers with no definition the client owned. Moving them meant first writing down what was running, much of which had never been documented. The real deliverable was not the move. It was that the environment became a definition (Terraform, Docker, Kubernetes) in the client's own repository.
Separate environments cleanly
Promotion from development to production should be documented, automated and reversible. For audited clients we add an approval gate where the developer who wrote a change cannot approve its release.
Let cost follow usage
For a group of companies in Japan, every website had its own hosting, so cost grew with the number of sites rather than traffic. Moving them onto one shared CMS on AWS serverless made cost track actual use.
Plan observability on day one
Logs, metrics and traces are not extras. They are how teams sleep after launch. The minimum we ship:
- Structured logs with a request ID across services.
- Four basic metrics per service: traffic, errors, latency, saturation.
- Alerts with an owner. An alert routed to nobody is noise.
- A restore test. A backup you have never restored is a hope, not a backup.
A three-year checklist
- Could you rebuild production from the repository alone?
- Can you deploy without the person who set it up?
- Do you know what each environment costs per month?
- Has anyone restored a backup in the last quarter?
- Could you move to another provider without a rewrite?
If any answer is "no", that is where the foundation needs work.
Sources and further reading
- The Twelve-Factor App, 12factor.net
- AWS Well-Architected Framework, AWS
- Azure Well-Architected Framework, Microsoft Learn
- Cloud security guidance, UK National Cyber Security Centre
- Monitoring distributed systems, Google SRE book
Related case study
Taking Back the Keys: ERP Migration from Vendor Servers to Client-Owned Kubernetes
