Opscraft
Deployment strategies · Lesson 1 of 10 · 3 min read

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.

Enlarge diagramDeployv2 is running in production, but no users seeit yetReleaseusers get v2: all at once, by percentage, orby feature flagwhen ready
Figure 1. Deploying puts new code on servers; releasing shows it to users. Keeping them separate is what makes safe delivery possible.
Swipe sideways to see all of it, or tap to open it full screen.

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.

Enlarge diagramslower rollback →biggerblastradiusRecreateRollingBlue-greenCanaryFeature flag
Figure 2. Strategies on one map: how much each risks if the release is bad, and how fast you can get back.
Swipe sideways to see all of it, or tap to open it full screen.

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.