Cloud landing zone design on AWS, Azure and Google Cloud
The first deliverable is a landing zone, not a server. That means the account or subscription hierarchy, with production isolated from everything else, a central identity provider issuing short-lived credentials instead of long-lived keys, a hub-and-spoke network in which databases never get a public address, and organisation-level guardrails, such as AWS service control policies or Azure Policy, that make the insecure option impossible rather than discouraged. Every piece is a Terraform module reviewed through a pull request, with remote state locked and versioned, and the same modules build development, staging and production, so the evidence an auditor asks for is the Git history. Decisions above that layer follow the workload: Kubernetes for long-running services a team will operate daily, serverless for spiky or event-driven jobs, managed databases almost always, and data residency fixed per region where regulation requires it.
How a phased cloud migration avoids downtime and risk
A migration plan begins with a dependency map, because the outages people remember were caused by the system nobody listed: the cron job on an old server, the hard-coded IP, the licence tied to a MAC address. Each workload then gets a route. Rehost when the application is fine and only the hardware is old, replatform when a self-managed database can become a managed one, refactor only when the code genuinely blocks the move. Stateless services go first, data goes last. Databases replicate continuously to the new platform with change data capture, so cutover is a DNS switch on a short TTL after a freeze measured in minutes, and rollback is repointing the record while replication still runs. Parallel running does cost two bills for a few weeks; that is the price of comparing outputs before anyone commits.
FinOps and observability that keep the platform affordable and reliable
Cost control follows a fixed order, because doing it backwards locks waste in. First, tags are enforced at deploy time so every resource has an owner, an environment and a cost centre, and untagged resources fail the pipeline. Second, utilisation data from the first weeks of production sets rightsizing and autoscaling policies. Only then are committed-use discounts bought, sized to the remaining baseline, since a reservation on an oversized instance is a cheaper mistake. Budget and anomaly alerts go to the engineers who own the tag, not to finance a month later. Reliability is handled the same way: uptime and latency targets per workload, alerts on symptoms users would notice rather than every CPU spike, and each alert linked to a runbook. You keep the Terraform, the dashboards, the runbooks and the cost model, whether or not we stay on.