Best Practices

We end the bonus clips with a checklist. It brings together the best practices of this course, in one place. Use it when you review your own clusters, or when you prepare for production. It is long, but each line comes from a lesson you have already seen. Pause on each screen, and tick the boxes.

• Reliability
• Security
• Operations

Reliability checklist

First, reliability. At least two replicas for everything important, spread across nodes, and zones. Resource requests, and limits, on every container. Readiness, and liveness probes, with sensible timing. Rolling updates, with max unavailable zero, for services. Pod disruption budgets, for safe maintenance. Graceful shutdown, handling the term signal. And autoscaling, based on real load. These seven items remove most outages, before they happen.

Security checklist

Second, security. Run containers as non root, with read only file systems, and no extra capabilities. Use role based access control, with least privilege, and a separate service account per application. Keep secrets out of Git, and encrypt them at rest. Apply network policies, with a default deny. Enforce pod security standards, and admission policies. Pull images from trusted registries, scan them, and never use the latest tag. And keep Kubernetes, and its nodes, upgraded.

Operations checklist

Third, operations. Everything as code, in Git, with changes reviewed, and delivered by a Git ops tool. Monitoring of metrics, logs, and events, with alerts on user visible symptoms, and a runbook for each. Regular backups of et see dee, and of volumes, with a rehearsed restore. Namespaces per team, with quotas, and labels for cost. Upgrades on a schedule, one minor version at a time. And a weekly health review, of dashboards, alerts, cost, and incidents. Boring operations are the goal.

Common anti-patterns

Finally, a few anti patterns to avoid. Running a single replica, of something important. Using the latest tag, so nobody knows what is running. Putting secrets in images, or in Git. Giving every pod cluster admin rights. Editing live objects by hand, without recording it. Skipping requests and limits, so one pod can starve the node. And running everything in the default namespace. Each of these has caused real outages, and real breaches. Now you know the alternative, for every one.

Summary