MLOps

See why models fail after they leave the notebook, learn the principles MLOps uses to prevent that, meet the parts of a production ML system, and the languages that glue it together.

1. Why Models Break in Production12 min
2. The Core MLOps Principles14 min
3. The Components of an MLOps System14 min
4. Python, Bash and Go in MLOps10 min

Pull training data reproducibly with SQL, keep code and configs in Git, version large data with DVC, and chain steps into a DVC pipeline that records lineage.

1. SQL for Reproducible Training Data14 min
2. Git and GitHub for ML Projects12 min
3. DVC: Versioning Data15 min
4. DVC Pipelines and Data Lineage15 min

Learn what continuous integration and delivery mean when the product is a model, build a GitHub Actions workflow that trains and tests on every pull request, and post model reports back to the pull request with CML.

1. What CI/CD Means for ML13 min
2. GitHub Actions for ML Workflows15 min
3. CML: Model Reports in Pull Requests13 min

Choose cloud compute and storage for training and serving, know what managed ML platforms do for you, package a model in a Docker image, and run training and serving on Kubernetes.

1. Cloud Compute and Storage for ML13 min
2. Cloud-native ML Services12 min
3. Docker Images for Models14 min
4. Kubernetes for Training and Serving14 min

Track every training run, promote the winner through a model registry, schedule the whole workflow as a pipeline, and serve the model safely in batch or online.

1. Experiment Tracking14 min
2. MLflow Model Registry14 min
3. Orchestrating Pipelines with Airflow15 min
4. Serving Models: Batch and Online15 min

Decide what to watch in a live ML system, detect when the world drifts away from the training data, and wire up Prometheus metrics and Grafana dashboards with alerts.

1. What to Monitor in an ML System13 min
2. Data Drift and Concept Drift15 min
3. Metrics with Prometheus and Grafana15 min