咨询：如何在Azure部署含machine learning algorithms的自定义Python包为服务并供应用调用

Hey Amit, let’s walk through the most practical ways to deploy your custom machine learning Python package on Azure—perfect for both your batch jobs and website integrations. Since you already have an Azure license, we can skip the basics and jump straight to tailored strategies that fit your use cases:

1. Azure Functions (Lightweight, On-Demand/Event-Driven Calls) #

Ideal for your website’s real-time requests or small-scale batch tasks. Here’s how to set it up:

• Package your custom ML code as a Python wheel, or include the package directory directly in your Function project. For private packages, you can also host them in Azure Artifacts and reference them in requirements.txt.
• Create an HTTP-triggered Azure Function and select your target Python version (3.8+ recommended).
• Wrap your ML algorithm logic in the Function’s entry point—handle HTTP request inputs, call your package, and return formatted results.
• Test it out with tools like Postman or a simple browser call to the Function’s URL.
• For batch jobs: Use a time-triggered Function, or pair it with Azure Logic Apps to schedule and orchestrate larger batch runs.

2. Azure Container Apps (Scalable, Containerized Deployments) #

Great if your package has complex dependencies or needs to scale with traffic. Steps to deploy:

• Write a Dockerfile to containerize your package: Install dependencies, copy your custom package, and set a runtime command. Example snippet:

  FROM python:3.10-slim
  COPY ./your-custom-package /app/your-custom-package
  RUN pip install /app/your-custom-package
  CMD ["python", "/app/inference.py"]
  

• Push the built image to Azure Container Registry (ACR).
• Create an Azure Container Apps instance: Pull your image from ACR, configure HTTP ports, and set resource limits.
• For batch workloads: Use Container Apps’ Job mode to submit one-off or scheduled batch tasks, or integrate with Azure Batch for large-scale job scheduling.
• Perk: Auto-scales based on traffic, so it’s perfect for both high-concurrency website calls and heavy batch processing.

3. Azure Machine Learning (ML-Focused, End-to-End Workflows) #

If you want built-in model management, monitoring, and versioning, this is your go-to platform:

• Upload your custom package to your Azure ML Workspace: Use the CLI command az ml package create or drag-and-drop via the Azure ML Studio UI.
• Deploy as two types of endpoints:
  • Online Endpoints: For real-time website calls—define a deployment with your package, inference script, and CPU/GPU resources.
  • Batch Endpoints: Purpose-built for batch jobs. Point it to input data in Blob Storage/ADLS, schedule runs, and output results to your preferred storage.
• Bonus: Azure ML includes out-of-the-box monitoring for model performance, logs, and drift—super helpful for iterating on your ML algorithms later.

4. Azure App Service (Integrate with Existing Web Apps) #

If your website is already hosted on Azure App Service, you can directly embed your ML package:

• Add your custom package to the App Service’s Python environment: Reference it in requirements.txt (use a local path if uploading the package directly to the App Service file system, or point to Azure Artifacts for private packages).
• Configure the App Service’s startup command to ensure your web app loads the package correctly.
• For batch jobs: Use App Service WebJobs to run scheduled or on-demand batch tasks that call your ML package.

Quick Pro Tips for All Deployments #

• Dependency Consistency: Use requirements.txt or pyproject.toml to lock in all dependencies—avoids "it works on my machine" issues.
• Monitoring & Logs: Enable Azure Monitor to track service performance, errors, and resource usage—critical for troubleshooting.
• Security: Add API keys or Azure AD authentication to your endpoints to block unauthorized access.
• Cost Optimization: Match your pricing tier to your workload: Use Azure Functions’ Consumption Plan for low-traffic scenarios, and spot VMs for non-time-sensitive batch jobs.

内容的提问来源于stack exchange，提问作者Amit