Cheat sheet

From Notebook to Endpoint — Cheat Sheet

An MLOps end-to-end on Microsoft Azure. From a Kaggle scikit-learn model to a live FastAPI inference endpoint with CI/CD, scale-to-zero, and a budget alert.

← Read the full projectUpdated June 2026
1

The problem

A trained model in a Kaggle notebook is not a product. To use it from anywhere, it needs to be:

  • Packaged into a reproducible artefact (Docker image).
  • Served over HTTP with a stable contract (FastAPI).
  • Deployed on infrastructure that scales when called, sleeps when idle.
  • Re-deployable safely on every code change (CI/CD).
  • Monitored so you don't wake up to a €300 cloud bill.

This project ticks all five — on Microsoft Azure specifically, because the CV signal I needed was Azure / MLOps / cloud / CI-CD.

2

The model

  • Dataset: Forest Cover Type Prediction (Kaggle / UCI). 581k samples, 54 features. 7 forest cover classes.
  • Algorithm: Random Forest (scikit-learn). Tuned with RandomizedSearchCV.
  • Accuracy: ~87 % on held-out test set.
  • Serialisation: joblib.dump() → loaded by FastAPI on startup.

The model itself is intentionally simple — the project is about the infrastructure around it, not the algorithm.

3

The architecture

Code (GitHub) → GitHub Actions
   ├─ Train + test (pytest)
   ├─ Lint (ruff)
   ├─ Build Docker image
   ├─ Push to GHCR
   └─ Deploy to Azure Container Apps
                       ↓
       https://<endpoint>.azurecontainerapps.io
            /health · /predict
  • FastAPI — /health for liveness, /predict for inference.
  • Docker — python:3.11-slim, copy model + code, run uvicorn.
  • GitHub Actions — two workflows: CI on every push, deploy on merge to main.
  • OIDC federated credential — no long-lived secrets stored in GitHub.
  • Azure Container Apps — managed, scale-to-zero, autoscale on traffic.
4

The cost discipline

Cloud bills are the silent risk. Three guardrails:

  1. Scale-to-zero — --min-replicas 0. Idle cost: €0.
  2. Budget alert at €5/month. Email + Slack on spend.
  3. GHCR public package — Container Apps pulls without registry credentials, no extra ACR cost.

Real monthly cost: €0–2. Worst case (scale-to-zero fails): ~€30/month.

5

What I learned

  • The IE alumni Microsoft account is a mess — IE tenant, blocked trial. Used a personal Gmail-tied account instead.
  • OIDC federated credentials > stored secrets. Setting it up takes 20 min, saves you from a leaked key every quarter.
  • scale-to-zero is non-negotiable for portfolio projects. Without it, your "live endpoint" is a slow bleed.
  • The model serialisation is the boring part. The boring part is what breaks in production.
6

Tech stack

LayerTool
Modelscikit-learn (Random Forest)
ServingFastAPI + Uvicorn
ContainerDocker, python:3.11-slim
RegistryGitHub Container Registry (ghcr.io)
CI/CDGitHub Actions (CI + Deploy workflows)
AuthOIDC federated credential (Azure AD ↔ GitHub)
HostingAzure Container Apps
Cost mgmtAzure budget alerts
Testpytest
Lintruff