ai-eng-demo-deployment-i-steps.md

Machine Learning - Model Deployment Demo - STEPS WE'LL FOLLOW

Intro to Deployment

• Environments (dev / staging / production)
• Recap: what's an API
• Architecture
  • Frontend vs. Backend vs. Database

Initial setup

Repo (main branch):

• https://github.com/ironhack-rmt-AI-student-materials/model-deployment-demo-i

Implement an API with FastAPI

• Intro
• Code → see branch "feat/model-deployment"
• Run api → see readme (on branch "feat/model-deployment")
• See API documentation http://localhost:8000/docs
• Test API locally (Postman)
• Deploy on Render.com
  • Sign-in
  • New "Web Service"
  • Select the repository (may need to grant permissions)
  • Environment: Python
  • Root Directory
    • leave blank
  • Build Command:
    • pip install -r requirements.txt
  • Start Command:
    • Example: uvicorn api.main:app --host 0.0.0.0 --port $PORT
    • Explanation: "uvicorn my_dir.my_subdir.main:app" tells Uvicorn where to find your FastAPI app by following the Python module path (my_dir → my_subdir → main.py) and running the app object defined there.
  • Set environment variables (if needed)
    • e.g. API keys, database URLs, secrets...
  • Select "Instance Type"
    • Limitations of the free tier:
      • Free instances spin down after periods of inactivity.
      • Ephemeral file system (SQLite will not work, since it stores the data in files; can use PostgreSQL)
      • https://render.com/docs/free#free-web-services

    - A Free web service spins down after 15 minutes of no inbound traffic (HTTP requests or WebSocket messages).
    - The next request automatically spins it back up, but startup typically takes about one minute, so the first request is noticeably slower.
    - Paid instance types do not spin down.
    

• Test API in production (Postman)
• Debugging & reading errors
• Discuss: how can we do if we have more than 1 model.

Frontend

• Implement the Frontend (e.g., React app)
• Deploy the Frontend (e.g., Vercel)