MLOps — from research to production

Machine Learning Operations (MLOps) is about shipping and maintaining ML models in production. Here is how to build a solid MLOps pipeline.

The MLOps challenge

Common issues

• Models that work in dev but fail in prod
• Data drift
• Lack of reproducibility
• Manual, risky deployments

Solution: automation

MLOps applies DevOps ideas to ML so the full lifecycle can be automated.

End-to-end MLOps pipeline

1. Development

# Version code
git init
git add .
git commit -m "Initial ML model"

# Experiment with MLflow
import mlflow
mlflow.start_run()
mlflow.log_param("learning_rate", 0.01)
mlflow.log_metric("accuracy", 0.95)
mlflow.log_model(model, "model")

2. Automated tests

# Unit tests
def test_model_prediction():
    model = load_model()
    prediction = model.predict([[1, 2, 3]])
    assert prediction is not None

# Performance tests
def test_model_performance():
    model = load_model()
    accuracy = evaluate(model, test_data)
    assert accuracy > 0.9

3. CI/CD pipeline

# .github/workflows/ml-pipeline.yml
name: ML Pipeline
on: [push]
jobs:
  train:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v2
      - name: Train model
        run: python train.py
      - name: Test model
        run: pytest tests/
      - name: Deploy if tests pass
        run: ./deploy.sh

4. Deployment

Options:

• Batch: Cloud Functions, Cloud Run
• Real-time: Vertex AI Endpoints, SageMaker
• Edge: TensorFlow Lite, ONNX Runtime

5. Monitoring

Watch:

• Performance: latency, throughput
• Quality: accuracy, F1
• Data drift: feature distributions
• Model drift: prediction quality over time

# Example with Evidently
from evidently import dashboard
from evidently.dashboard import Dashboard

dashboard = Dashboard(tabs=[DataDriftTab()])
dashboard.calculate(reference_data, current_data)
dashboard.save("report.html")

MLOps tooling

Orchestration

• Kubeflow: Kubernetes-native
• MLflow: tracking and registry
• Weights & Biases: experimentation

Deployment

• Vertex AI: Google Cloud
• SageMaker: AWS
• Azure ML: Microsoft

Monitoring

• Evidently AI: data drift
• Prometheus: metrics
• Grafana: dashboards

Best practices

1. Version code, data, and models
2. Test unit, integration, and performance
3. Automate with CI/CD
4. Monitor metrics continuously
5. Document models and processes

Conclusion

MLOps is key to reliable ML in production. Automate your pipeline to reduce risk and improve quality.