World map

 

BM

World K7: Self-Driving

Progressive delivery

You will learn: how Argo Rollouts and Flagger send a new version to a few users first, and roll back by themselves.

Checked with Kubernetes 1.37 on 31 Oct 2026

The problem

The rolling update finished. Every pod is ready and green. Yet 8 in 100 orders now fail. What did the rollout check?

Try it

progressive delivery: Serve the new dish to a few tables first, check the numbers, then to everyone. Bad numbers roll it back. sends a few users first. Step the new version up. Watch the errors. Roll back before 150 users get hurt.

Send a few users to v2. Watch the graph. Roll back if errors jump.

Users now: 100% on v1 (old), 0% on v2 (new).

Error rate per minute
0%6%11%alarmminutes →

v1 (old) v2 (new)

Minute 0

Users who hit a v2 error: 0 (limit 150)

Numbers, minute by minute

No traffic yet.

Watch

Argo Rollouts writes those steps in YAML. A Rollout takes the place of a Deployment. Tap the lines that send more users to the new version.

Tap every line that sends more users.

0 of 2 found

Try it

Add a robot judge that reads the error graph. Put its loop in order.

First step at the top. Put these in order. Use the arrows, or drag.

  • Errors low: send 50%
  • Send 20% of users to v2
  • Errors high at any step: back to v1
  • Read the error rate
  • Still low: send everyone

Check

Two popular tools do this. Sort the facts.

Drag a card into a box. Or tap the card, then tap the box.

        Predict

        At 20%, the judge sees errors above the limit. What happens next?

        You did it!

        You can now ship to a few users first and let numbers decide.

        Go deeper (optional)

        Without a traffic router, a weight is only roughly a share of pods. With Gateway API, a service mesh or an ingress, the split is exact. Both tools also do blue/green. Analysis needs metrics, often from Prometheus. Real rep (optional): try the Argo Rollouts demo on a kind cluster.