Skip to content
SP.
EN DE
Back to Work
Reference design

A platform that ships with confidence

Turn a fragile release process into a repeatable, observable path from commit to production.

  • Kubernetes
  • Terraform
  • Argo CD
  • GitHub Actions
  • Prometheus
Deployment target
< 10 min
Rollback target
< 5 min
Artifact per release
1

Design targets · not production measurements

The problem

A release is difficult to trust when its state lives across a CI job, a shell script, and someone’s terminal. This reference architecture explores a small, auditable delivery platform. The numbers above are design targets, not measured client outcomes.

The system

Build once, promote the same immutable artifact, and let the cluster reconcile its desired state from Git. Keep build credentials separate from runtime credentials.

A commit triggers checks and an image build. A reviewed configuration change is reconciled by Argo CD into Kubernetes, with Prometheus observing the cluster.

The pipeline publishes an image; a reviewed configuration change selects its digest. Argo CD applies that desired state, and Prometheus supplies the feedback needed to judge the release.

Decisions and trade-offs

  • Terraform owns the foundation. Network, cluster, and identity changes go through a separate plan-and-review workflow.
  • Git owns application state. Production changes remain reviewable and reversible. Reconciliation introduces a small deployment delay.
  • Health gates are explicit. Readiness checks protect traffic, but application-level error rates decide whether a rollout is successful.
  • One artifact moves between environments. Rebuilding for production would invalidate the evidence gathered in staging.
# Deployment excerpt: replace with your application's measured requirements.
strategy:
  type: RollingUpdate
  rollingUpdate:
    maxUnavailable: 0
    maxSurge: 1
minReadySeconds: 15
progressDeadlineSeconds: 600

Failure and recovery

Revert the image digest in Git and let the controller reconcile. Pause automated promotion if error-budget burn increases. Database migrations must remain backward-compatible throughout this window; reverting a container cannot undo an incompatible schema change.

What to measure

Record commit-to-healthy duration, failed-change percentage, and time to restore service. Validate the targets through staging failure drills before setting a production objective.

Back to Work