Skip to main content
Machine Learning & AI

MLOps Tutorial with Examples.

MLOps tutorial for production ML engineering: experiment tracking, model registry, CI/CD for ML, monitoring and drift detection.

3 Beginner 5 Intermediate +4 From other topics

Architecture at a glance

Click any box to open the lesson that teaches it

MLOps architecture at a glance

data → train & register → serve ⟲ monitor, wrapped in CI/CD

MLOps architecture at a glance Data is versioned with DVC, then a training run is tracked and registered in MLflow behind a movable alias. Model serving loads that alias for a FastAPI endpoint and batch scoring; the same model can also deploy through PyTorch's ONNX export or TensorFlow's TF Serving. Monitoring watches live inputs for drift and feeds a retrain decision back into the training box. CI/CD wraps the pipeline: validate data, gate on a metric, then deploy. ML Lifecycle Data DVC Train & Register MLflow runsregistry alias Serving FastAPI API Monitoring drift: KS & PSI retrain on drift Pipeline & registry DVC pipeline prepare → train Registry aliases champion ↔ challenger Interview: reproducibility, rollback Also deploys with PyTorch → ONNX Runtime TensorFlow → TF Serving CI/CD: train → test → deploy CI/CD pipeline Validate data Gate on metrics Deploy to prod
How it fits together (6 steps)
  1. The lifecycle starts in Introduction to MLOps, then data is versioned with DVC — a small pointer file lives in Git, the data itself in a cache and a remote, restored with dvc pull and dvc checkout.
  2. A training run is tracked in MLflow 3: parameters and metrics are logged per run, search_runs compares them, and the winner is registered under a movable alias such as champion.
  3. Model serving loads that aliased model once at startup, answers predictions through a validated FastAPI endpoint, and batch-scores a whole Parquet file the same way.
  4. The same trained model also deploys a framework's own way: PyTorch exported to ONNX Runtime, or a Keras SavedModel served by TF Serving — neither path changes how it was trained or registered.
  5. Monitoring watches live inputs for drift with a KS test and PSI, catching what accuracy alone can miss — the same reproducibility and rollback discipline is what makes a drift alert, not a calendar, the trigger to retrain.
  6. Every stage ships through CI/CD: GitHub Actions validates incoming data, gates promotion on a metric threshold, and only then deploys.

Beginner

Start here — no prior experience needed

3 lessons
mlops beginner
5 min read

MLOps Learning Roadmap

A structured path from experiment tracking basics to fully automated ML platforms — covering CI/CD, monitoring, data versioning, and serving.

Read Tutorial
mlops beginner
5 min read

Introduction to MLOps

Understand the ML lifecycle, set up experiment tracking with MLflow, and build your first reproducible ML pipeline.

Read Tutorial
mlops beginner
5 min read

MLOps Projects

Projects that build production ML infrastructure — from experiment tracking to automated retraining pipelines and monitoring systems.

Read Tutorial

Intermediate

For developers with core concepts down

5 lessons
mlops intermediate
5 min read

MLOps Experiment Tracking

Track ML experiments with MLflow 3 locally: log params and metrics, compare runs with search_runs, per-epoch metrics, and a registry with aliases.

Read Tutorial
mlops intermediate
5 min read

Model Serving and Deployment

Serve a model with FastAPI and pydantic, load it once from an MLflow alias, batch-score Parquet in chunks, and gate promotion and rollback on a holdout.

Read Tutorial
mlops intermediate
5 min read

MLOps Model Monitoring

Detect data drift with KS tests and PSI, avoid the large-sample p-value trap, see concept drift that inputs can't show, and turn checks into alerts.

Read Tutorial
mlops intermediate
5 min read

MLOps CI/CD for Machine Learning

Automate ML workflows with GitHub Actions — test data quality, validate models, and deploy only when performance thresholds are met.

Read Tutorial
mlops intermediate
5 min read

MLOps Data Versioning with DVC

Version datasets and models next to Git with DVC: a local remote, dvc push and pull, a cached pipeline that skips unchanged stages, and params diffs.

Read Tutorial

Also part of MLOps

Taught in other topics. Each link opens the lesson in its home topic.

4 lessons
llm advanced
5 min read

LLM Evaluation

Why overlap metrics mislead, assertion-based graders, judge calibration with kappa, retrieval recall and MRR, and a regression gate, all run in Python.

Read Tutorial
pytorch advanced
5 min read

PyTorch Deployment with ONNX

Export PyTorch models to ONNX, verify them against PyTorch, serve with ONNX Runtime, measure latency, and replace TorchScript with torch.export.

Read Tutorial
tensorflow advanced
5 min read

TensorFlow Model Deployment

Deploy a Keras model for real: export a SavedModel, serve it with TF Serving in Docker, wrap it in FastAPI, and convert it to int8 TFLite run with LiteRT.

Read Tutorial
ai-interviews advanced
5 min read

MLOps and Production Questions

Drift detected with PSI, reproducibility that actually holds, rollback that works, and the retraining trigger that is not a calendar.

Read Tutorial