kubernetes model serving rollback

If you want to roll back model serving in Kubernetes, the exact steps depend on how you deployed it. Common patterns are:

1) If you deployed with Helm

Check release history:

helm history <release-name> -n <namespace>

Roll back to the previous revision:

helm rollback <release-name> <revision> -n <namespace>

Example:

helm rollback model-serving 3 -n ml-prod

Verify:

kubectl get pods -n ml-prod
kubectl get svc -n ml-prod

---

2) If you deployed with plain Kubernetes manifests

Apply the previous known-good YAML:

kubectl apply -f previous-manifest.yaml

Or if using GitOps, revert the commit in Git and let Argo CD / Flux sync it.

---

3) If you use a Deployment for the model server

You can rollback to a previous ReplicaSet:

kubectl rollout history deployment/<deployment-name> -n <namespace>
kubectl rollout undo deployment/<deployment-name> -n <namespace>

Rollback to a specific revision:

kubectl rollout undo deployment/<deployment-name> --to-revision=<rev> -n <namespace>

Example:

kubectl rollout undo deployment/model-server --to-revision=2 -n ml-prod

Check status:

kubectl rollout status deployment/model-server -n ml-prod

---

4) If you use canary / blue-green deployment

Rollback usually means shifting traffic back to the stable version:

• restore old service selector
• revert ingress / virtual service weights
• promote stable revision again

Examples:

• Argo Rollouts:

  kubectl argo rollouts undo rollout/<name> -n <namespace>
  

• Istio VirtualService: set traffic weight back to 100% stable
• KServe: update the InferenceService to the previous predictor image or model URI

---

5) For model-specific rollback

If the deployment is fine but the model artifact/version is bad:

• point to the previous model version in storage
• update the serving config to older model URI
• redeploy or restart pods if they cache model files

Example:

kubectl set env deployment/model-server MODEL_VERSION=v1 -n ml-prod
kubectl rollout restart deployment/model-server -n ml-prod

---

Safe rollback checklist

Before rollback:

• identify the last known-good model version
• confirm whether rollback means code, image, or model artifact
• check if schema changes or feature flags make old version incompatible
• watch logs and metrics after rollback

Useful commands:

kubectl get pods -n <namespace>
kubectl describe pod <pod> -n <namespace>
kubectl logs -f deployment/<deployment-name> -n <namespace>

If you tell me what you use to serve the model — Helm, Deployment, KServe, Seldon, Argo Rollouts, or something else — I can give you the exact rollback commands.