Core ideas: deploy, release and risk
A deployment strategy decides how a new version replaces the old one and how many users are exposed while you find out whether it works. Senior interviews go well beyond definitions: they expect you to explain the Kubernetes mechanics that make rolling updates safe, design blue-green and canary releases on EKS and AWS load balancers, automate canary analysis, handle database changes without downtime, and have a clear rollback plan. This topic covers all of it, using a Flux-based GitOps setup on EKS as the running example.
Deployment is not release
- Deployment — putting new code onto infrastructure.
- Release — exposing it to users.
Separating them is the heart of modern delivery. Blue-green deploys a new version without releasing it; canaries release to a few users first; feature flags release features independently of deployments; dark launches deploy code that nobody can see yet.
What you're trading off
| Dimension | Question |
|---|---|
| Blast radius | If the new version is bad, how many users are hurt before you notice? |
| Detection time | How quickly do metrics reveal the problem? |
| Rollback time (MTTR) | How fast can you get back to good? |
| Cost | Extra capacity needed during the release? |
| Complexity | Tooling, traffic control, metrics, team skill |
| Compatibility | Can old and new versions run side by side (APIs, schemas, sessions, message formats)? |
Every strategy is a different point on this map. Picking one is a risk decision, not a fashion choice.
Vocabulary you should use precisely
- Rollback — return to the previous known-good version. Roll forward — fix by deploying a new version (sometimes faster/safer when rollback is impossible, e.g. after an irreversible migration).
- Immutable deployment — replace instances/pods/images rather than modifying running ones. Rollback = redeploy the old artifact.
- Big bang — replace everything at once with no gradual exposure.
- Bake time — time spent watching metrics after a step before proceeding.
- Progressive delivery — gradual exposure combined with automated analysis and automated rollback.