MLOps: deploy Azure ML Models to Azure Kubernetes Service (AKS)

In this article I’ll walk through an end-to-end process of deploying a model from Azure Machine Learning (Azure ML) to Azure Kubernetes Service (AKS), using Docker and Azure Container Registry (ACR).

Prerequisites

An Azure account with access to Azure ML, ACR, and AKS.
Docker installed on your machine.
Azure CLI and kubectl installed on your machine.

Workflow Overview

Export the Model from Azure ML
Create a Docker Image with the Model
Push the Image to Azure Container Registry (ACR)
Deploy the Image to AKS

Step 1: Export the Model from Azure ML

Train and Register Your Model

Train your model using Azure ML workspaces. For example, use Azure ML SDK to train a simple scikit-learn model and register it in the Azure ML Model Registry.

from azureml.core import Workspace, Model
from sklearn.ensemble import RandomForestClassifier
import joblib

# Connect to Azure ML Workspace
ws = Workspace.from_config()

# Assume X_train, y_train are training data and labels
model = RandomForestClassifier().fit(X_train, y_train)

# Save the model locally
joblib.dump(model, 'model.pkl')

# Register the model in Azure ML
registered_model = Model.register(model_path='model.pkl',
                                  model_name='my_awesome_model',
                                  workspace=ws)

Download the Registered Model

Once the model is registered, you can download it from the Model Registry to your local environment.

from azureml.core.model import Model

# Download the model
model = Model(ws, 'my_awesome_model')
model.download(target_dir='.', exist_ok=True)

Step 2: Create a Docker Image with the Model

Create a Docker container that includes your model and the necessary runtime environment.

Write a Dockerfile

# Dockerfile
FROM python:3.8-slim
COPY . /app
WORKDIR /app
RUN pip install scikit-learn joblib flask
ENTRYPOINT ["python", "app.py"]

Build the Docker Image

docker build -t mymodel:latest .

Step 3: Push the Image to Azure Container Registry (ACR)

az acr login --name <RegistryName>
docker tag mymodel:latest <RegistryName>.azurecr.io/mymodel:latest
docker push <RegistryName>.azurecr.io/mymodel:latest

Step 4: Deploy the Image to AKS

Deploy your model as a service in AKS using the image stored in ACR.

Kubernetes Deployment YAML

Create a Kubernetes deployment file that specifies the use of your image from ACR.

# deployment.yaml
apiVersion: apps/v1
kind: Deployment
metadata:
  name: mymodel-deployment
spec:
  replicas: 2
  template:
    spec:
      containers:
      - name: mymodel
        image: <RegistryName>.azurecr.io/mymodel:latest
        ports:
        - containerPort: 80

Deploy to AKS

Apply the Kubernetes manifests to your AKS cluster.

kubectl apply -f deployment.yaml
kubectl apply -f service.yaml