Step 01 of 06
Learn the concept
A deployment is not the same as a release. You can put new code in the cluster and still expose it to only five percent of users. Progressive delivery buys time for metrics to disagree with optimism.
The ideas this is made of
Rolling update controls capacity, not correctness
A Kubernetes Deployment rolling update limits unavailable and extra Pods. It does not know whether the new version returns wrong answers. It will happily roll out a fast, broken service.
Blue/green makes rollback simple at infrastructure cost
Blue/green keeps two complete versions ready and moves traffic from one to the other. Rollback can be a traffic switch. The price is duplicated capacity and compatibility with shared dependencies.
A canary needs real metrics
Sending five percent of traffic to a new version helps only if something evaluates the result. Error rate, latency and saturation should decide whether the rollout continues. Otherwise it is just a slow rollout.
Feature flags decouple deploy from release
A flag lets code be deployed while behavior remains disabled for most users. That gives finer control. It also creates debt: old flags need owners and removal dates.
5% shoppers -> new lane
watch: queue time and failed payments
if failures > 2%: close lane
else: 25%, 50%, 100%The lane exists before every shopper uses it. Release means sending traffic there, and the metric decides whether that traffic grows.
Release strategies compared
| Strategy | Traffic shape | Rollback | Cost |
|---|---|---|---|
Rolling | Pod by Pod | ReplicaSet | Low |
Blue/green | All at switch | Switch back | High capacity |
Canary | Small to large | Abort early | Needs metrics |
Feature flag | Per cohort | Turn off | Code complexity |
What these are called on the job
Canary — Partial release to a small traffic share before wider rollout.
Blue/green — Two complete versions with traffic switched between them.
AnalysisTemplate — Argo Rollouts object that defines metric checks for a rollout.
Feature flag — Runtime switch that changes behavior without a new deploy.