Azure ML Deployment Methods

This repo demonstrates different quick ML deployment methods on Azure.

Anyone needing to deploy using Azure can pick the most suitable deployment method and use the sample code here to by swapping in their model. Each notebook takes the model and deploys it to an endpoint hosted in Azure cloud for real-time inference. We show a range of lightweight methods using various Azure cloud model hosting and deployment tools:

1. Azure Managed Endpoint (with Azure Machine Learning)
2. Azure Function App
3. Azure Container Instance
4. Azure Web App
5. Azure Kubernetes Endpoint

A sample model ready for deployment is provided in the models folder. This model simulates a model in a production setting coming from:

• A continuous training stage in a CI/CD pipeline: see notebook for sample training code.
• A registered model in a ML tracking registry such as MLFLow or Azure ML Workspace: see notebook for sample download code.

Get started

• Create environment: python -m venv venv && source venv/Scripts/activate
• Install requirements: pip install -r requirements.txt
• Setup Azure CLI: sudo apt-get install --only-upgrade azure-cli; az login

Model tracking

Models can be tracked using Azure Machine Learning studio (TODO insert link) and can be registered or downloaded as appropriate at any time. Alternatively, you can use an external MLFlow (TODO insert link) model tracking workspace. However, at time of writing there is no support for deploying models from an external MLFlow to Azure. Note that, at time of writing, you can use mlflow.deployments via the azureml-mlflow plugin to wrap MLFlow syntax around deploying an Azure Machine Learning workspace-tracked model to Azure Managed Endpoints, which is the same method as (1). See this article for latest MLFlow support.

Model deployment

1. Deploy to Azure Managed Endpoints

See notebook 01 - Deploy Azure Managed Endpoint

This registers a ML model in the Azure Machine Learning workspace (TODO add link) and deploys it to Azure Managed Endpoints using the Azure Machine Learning Python SDKTODO: add link. Note that Azure Machine Learning is fairly expensive. Notes:

• This uses the azure-ai-ml Python SDK v2 for interacting with Azure Machine Learning. This is the most up-to-date method and deprecates using azureml-core.
• Direct deployment from Azure Machine Learning to Azure Container Instances (ACI) is deprecated, see this blog for info.

2. Deploy to Azure Function App - serverless approach

See notebook 02 - Deploy Azure Function App

This containerises a ML model and deploys to an Azure Function App, which is a cheaper and lightweight serverless way of exposing a model to an inference endpoint.

3. Deploy to Azure Container Instances with custom container (BYOC)

See notebook 03 - Deploy Azure Container Instance

This gives the most flexibility without creating a dedicated compute instance. The model, inference script and API endpoint are all bundled into a Docker image, so that scoring, environment, load balancing can all be configured in the image. The image is pushed to Azure Container Registry (ACR) and deployed with Azure Container Instances (ACI).

4. Deploy to Azure App Service

tutorial

5. Deploy to Azure Kubernetes Service

Azure Kubernetes endpoint with AKS (sample), more inspiration

Development

• Render docs: cp README.md index.qmd && quarto render