Blue-Green Rollout of a ML Model¶

This tutorial shows how Iter8 can be used to implement a blue-green rollout of ML models hosted in a KServe modelmesh serving environment. In a blue-green rollout, a percentage of inference requests are directed to a candidate version of the model. The remaining requests go to the primary, or initial, version of the model. Iter8 enables a blue-green rollout by automatically configuring the network to distribute inference requests.

After a one time initialization step, the end user merely deploys candidate models, evaluates them, and either promotes or deletes them. Optionally, the end user can modify the percentage of inference requests being sent to the candidate model. Iter8 automatically handles all underlying network configuration.

In this tutorial, we use the Istio service mesh to distribute inference requests between different versions of a model.

Install the Iter8 controller¶

helm install --repo https://iter8-tools.github.io/hub iter8-traffic traffic

kubectl apply -k 'https://github.com/iter8-tools/hub.git/kustomize/traffic/clusterScoped?ref=traffic-templates-0.1.1'

kubectl apply -k 'https://github.com/iter8-tools/hub.git/kustomize/traffic/namespaceScoped?ref=traffic-templates-0.1.1'

Deploy a primary model¶

Deploy the primary version of a model using an InferenceService:

cat <<EOF | kubectl apply -f -
apiVersion: "serving.kserve.io/v1beta1"
kind: "InferenceService"
metadata:
  name: wisdom-0
  labels:
    app.kubernetes.io/name: wisdom
    app.kubernetes.io/version: v1
    iter8.tools/watch: "true"
  annotations:
    serving.kserve.io/deploymentMode: ModelMesh
    serving.kserve.io/secretKey: localMinIO
spec:
  predictor:
    model:
      modelFormat:
        name: sklearn
      storageUri: s3://modelmesh-example-models/sklearn/mnist-svm.joblib
EOF

Inspect the deployed InferenceService:

kubectl get inferenceservice wisdom-0

When the READY field becomes True, the model is fully deployed.

Initialize the Blue-Green routing policy¶

Initialize model rollout with a blue-green traffic pattern as follows:

cat <<EOF | helm template traffic --repo https://iter8-tools.github.io/hub traffic-templates -f - | kubectl apply -f -
templateName: initialize-rollout
targetEnv: kserve-modelmesh
trafficStrategy: blue-green
modelName: wisdom
EOF

The initialize-rollout template (with trafficStrategy: blue-green) configures the Istio service mesh to route all requests to the primary version of the model (wisdom-0). Further, it defines the routing policy that will be used by Iter8 when it observes changes in the models. By default, this routing policy splits inference requests 50-50 between the primary and candidate versions. For detailed configuration options, see the Helm chart.

Verify network configuration¶

To verify the network configuration, you can inspect the network configuration:

kubectl get virtualservice -o yaml wisdom

To send inference requests to the model:

1. In a separate terminal, port-forward the ingress gateway:

   kubectl -n istio-system port-forward svc/istio-ingressgateway 8080:80
   

2. Download the proto file and a sample input:

   curl -sO https://raw.githubusercontent.com/iter8-tools/docs/v0.13.18/samples/modelmesh-serving/kserve.proto
   curl -sO https://raw.githubusercontent.com/iter8-tools/docs/v0.13.18/samples/modelmesh-serving/grpc_input.json
   

3. Send inference requests:

   cat grpc_input.json | \
   grpcurl -plaintext -proto kserve.proto -d @ \
   -authority wisdom.modelmesh-serving \
   localhost:8080 inference.GRPCInferenceService.ModelInfer
   

Note that the model version responding to each inference request can be determined from the modelName field of the response.

Deploy a candidate model¶

Deploy a candidate model using a second InferenceService:

cat <<EOF | kubectl apply -f -
apiVersion: "serving.kserve.io/v1beta1"
kind: "InferenceService"
metadata:
  name: wisdom-1
  labels:
    app.kubernetes.io/name: wisdom
    app.kubernetes.io/version: v2
    iter8.tools/watch: "true"
  annotations:
    serving.kserve.io/deploymentMode: ModelMesh
    serving.kserve.io/secretKey: localMinIO
spec:
  predictor:
    model:
      modelFormat:
        name: sklearn
      storageUri: s3://modelmesh-example-models/sklearn/mnist-svm.joblib
EOF

Verify network configuration changes¶

The deployment of the candidate model triggers an automatic reconfiguration by Iter8. Inspect the VirtualService to see that inference requests are now distributed between the primary model and the secondary model:

kubectl get virtualservice wisdom -o yaml

Send additional inference requests as described above.

Modify weights (optional)¶

You can modify the weight distribution of inference requests using the Iter8 traffic-template chart:

cat <<EOF | helm template traffic --repo https://iter8-tools.github.io/hub traffic-templates -f - | kubectl apply -f -
templateName: modify-weights
targetEnv: kserve-modelmesh
trafficStrategy: blue-green
modelName: wisdom
modelVersions:
  - weight: 20
  - weight: 80
EOF

Note that using the modify-weights overrides the default traffic split for all future candidate deployments.

As above, you can verify the network configuration changes.

Promote the candidate model¶

Promoting the candidate involves redefining the primary InferenceService using the new model and deleting the candidate InferenceService.

Redefine the primary InferenceService¶

cat <<EOF | kubectl replace -f -
apiVersion: "serving.kserve.io/v1beta1"
kind: "InferenceService"
metadata:
  name: wisdom-0
  namespace: modelmesh-serving
  labels:
    app.kubernetes.io/name: wisdom
    app.kubernetes.io/version: v2
    iter8.tools/watch: "true"
  annotations:
    serving.kserve.io/deploymentMode: ModelMesh
    serving.kserve.io/secretKey: localMinIO
spec:
  predictor:
    model:
      modelFormat:
        name: sklearn
      storageUri: s3://modelmesh-example-models/sklearn/mnist-svm.joblib
EOF

Delete the candidate InferenceService¶

kubectl delete inferenceservice wisdom-1

Verify network configuration changes¶

Inspect the VirtualService to see that the it has been automatically reconfigured to send requests only to the primary model.

Clean up¶

Delete the candidate model:

kubectl delete --force isvc/wisdom-1

Delete routing artifacts:

cat <<EOF | helm template traffic --repo https://iter8-tools.github.io/hub traffic-templates -f - | kubectl delete --force -f -
templateName: initialize-rollout
targetEnv: kserve-modelmesh
trafficStrategy: blue-green
modelName: wisdom
EOF

Delete the primary model:

kubectl delete --force isvc/wisdom-0

Uninstall the Iter8 controller:

helm delete iter8-traffic

kubectl delete -k 'https://github.com/iter8-tools/hub.git/kustomize/traffic/clusterScoped?ref=traffic-templates-0.1.1'

kubectl delete -k 'https://github.com/iter8-tools/hub.git/kustomize/traffic/namespaceScoped?ref=traffic-templates-0.1.1'