Skip to content
datarekha

Course · 4 chapters · 35 lessons

MLOps

The lifecycle, Docker, CI/CD, MLflow, FastAPI/BentoML serving, monitoring, and the incident-response playbook the textbook skips.

The MLOps journey 0 / 35 completed
  1. Chapter 01

    Lifecycle

    5 lessons
  2. 01 The real ML lifecycle Train once, ship forever — that's the textbook story. The real lifecycle is a loop, and every stage has its own way of biting you. Beginner5 min
  3. 02 Establishing baselines Before you train a model, train the dumbest possible model. It's a floor, a leakage detector, and the thing that most often saves you from shipping… Beginner5 min
  4. 03 Data-centric AI The fastest way to improve a stuck model is almost never another hyperparameter sweep. It's fixing the 50 mislabeled rows that are dragging the sli… Intermediate6 min
  5. 04 Data contracts & quality Most ML failures are silent data failures. A data contract is an enforced agreement on a dataset's schema and meaning, so bad data fails the pipeli… Intermediate6 min
  6. 05 Labeling and production feedback loops How to turn messy human judgments and delayed production outcomes into training data you can trust. Intermediate13 min
  7. Chapter 02

    Tooling

    9 lessons
  8. 06 MLflow experiment tracking Stop guessing which notebook produced the model in prod. MLflow is the duct tape between 'I trained a model' and 'I can reproduce, compare, and rol… Intermediate7 min
  9. 07 Data & model versioning Version datasets and model artifacts so an old ML run can be rebuilt, inspected, and defended months later. Why a git commit is not enough, how con… Beginner6 min
  10. 08 Lineage and provenance Trace one prediction back through the exact data, code, environment, configuration, and deployment that produced it. Intermediate12 min
  11. 09 Model registry & promotion The release record for model versions: lineage links, approval gates, aliases, rollback, and the boundary between approving a model and deploying it. Intermediate6 min
  12. 10 Reproducible environments and builds Pin the dependency graph, the native runtime, and the build inputs so your model does not depend on one lucky laptop. Beginner12 min
  13. 11 Docker for ML Works-on-my-laptop is not a deployment strategy. Docker is how you ship a Python ML service that runs the same in CI, staging, and production — and… Intermediate7 min
  14. 12 CI/CD with GitHub Actions One repo, two pipelines. Code CI runs on every PR. Model CD runs on a schedule, retrains, validates against baselines, and blocks the deploy if it… Intermediate6 min
  15. 13 Testing ML & the ML Test Score Unit tests on code are not enough. Learn how to test ML data, model behavior, infrastructure, and monitoring, then use Google's ML Test Score to ag… Intermediate7 min
  16. 14 Pipeline orchestration How an orchestrator turns fragile ML scripts into a scheduled, retryable, observable pipeline, and how to choose between Airflow, Dagster, Prefect,… Intermediate7 min
  17. Chapter 03

    Serving & Monitoring

    12 lessons
  18. 15 Serving with FastAPI A REST API for your model in 60 lines — with the four things juniors miss the first time: Pydantic validation, lifespan startup, batch endpoints, a… Intermediate8 min
  19. 16 BentoML & Ray Serve FastAPI is a great starting point. Past a certain scale you want automatic batching, multi-model hosting, versioned model storage, and autoscaling… Advanced7 min
  20. 17 Batch vs real-time inference A practical guide to batch, real-time, async, streaming, and hybrid inference: choose by freshness, response latency, throughput, burstiness, relia… Intermediate6 min
  21. 18 Deployment Strategies A new model can pass every offline test and still break in production. Deployment strategies control the blast radius when it does — canary, blue-g… Intermediate9 min
  22. 19 A/B testing & experimentation Deploying a model gets it live; A/B testing proves it's actually better. Hypothesis testing, primary vs guardrail metrics, sample size, and varianc… Intermediate7 min
  23. 20 Feature stores (Feast, Tecton) The online/offline skew problem in one sentence, then a runnable Feast-shaped feature pipeline. Plus the honest answer to 'do we need a feature sto… Advanced10 min
  24. 21 Drift & monitoring Your model isn't broken. Your data changed. Here's the difference between data drift and concept drift, how to detect each, and when 'retrain' is a… Advanced7 min
  25. 22 Training-Serving Skew Why a model that scores beautifully offline dies the day it goes live. The two mechanisms — online-offline skew and point-in-time leakage — why you… Advanced9 min
  26. 23 Retraining & continual learning How to choose retraining triggers, build a fresh training window, compare a challenger with the production champion, and automate updates without s… Intermediate7 min
  27. 24 Incident response A model is misbehaving in production. Customers are noticing. Here's the playbook — rollback first, debug later — and the postmortem template that… Advanced6 min
  28. 25 LLMOps — operating LLMs MLOps assumed you trained the model. With LLMs you usually didn't — so the artifact you version is the prompt, evals replace the accuracy number, a… Advanced11 min
  29. 26 ML observability and SLOs A practical guide to four-layer ML telemetry, delayed labels, SLOs, error budgets, and alerts that wake people for customer impact. Intermediate12 min
  30. Chapter 04

    Platform & Infrastructure

    9 lessons
  31. 27 The cloud: AWS, Azure & GCP Three providers, three hundred service names, one mental model. What 'the cloud' actually is, the four primitives you'll touch, a Rosetta stone acr… Intermediate11 min
  32. 28 Kubeflow Pipelines KFP v2 turns a Python function into a containerised pipeline step with typed inputs, output artifacts, and metadata tracking. It's powerful — and i… Advanced10 min
  33. 29 Kubernetes for ML Pods, Deployments, Services, GPU scheduling, and the HPA — the four primitives that explain 90% of what 'we deploy on K8s' actually means. Plus an… Advanced9 min
  34. 30 Cost & FinOps for ML/GPUs GPU spend can become a board-level concern, and MLOps shares responsibility for making it visible and controllable. Learn how allocated capacity, u… Intermediate7 min
  35. 31 Responsible-AI ops Make fairness, human oversight, and AI governance part of the delivery system: per-group audits, living model cards, registry evidence, production… Intermediate7 min
  36. 32 ML security (MLSecOps) A practical guide to securing ML systems against poisoned data, model theft, adversarial inputs, malicious model files, and the supply-chain failur… Intermediate7 min
  37. 33 Infrastructure as code for ML Turn GPU pools, buckets, IAM, and endpoints from undocumented clicks into reviewable, repeatable infrastructure. Intermediate13 min
  38. 34 Model and artifact supply-chain security Keep hostile weights, packages, and build outputs from becoming code execution in your training and serving environments. Advanced12 min
  39. 35 Designing an internal ML platform A practical reference architecture for composing ML systems, choosing what to build, and keeping the platform from becoming a ticket queue. Advanced13 min
  40. End of section 0 / 35 complete

    Make it stick — pass every quiz.

    Each lesson has a short quiz at the bottom. Passing the quiz is what marks the lesson complete and counts toward your certificate.

    Section complete 35 / 35 lessons

    Nice work — you finished MLOps.

    Certificates are earned per learning path, not per section. Here's where this section takes you:

FAQCommon questions

MLOps — frequently asked questions

Straight answers to the questions people ask most about mlops.

What is MLOps, in one sentence?

MLOps is the practice of reliably deploying, monitoring, and maintaining machine-learning models in production — applying DevOps discipline (version control, CI/CD, automation, monitoring) to the messier reality of data and models that drift over time.

Why do models that work in a notebook fail in production?

Common causes are training/serving skew (production data or preprocessing differs from training), data drift over time, environment and dependency differences, and no monitoring to catch the decline. Reproducible pipelines and Docker containers close most of these gaps.

What is model drift and how do I detect it?

Drift is when the live data (data drift) or the input–target relationship (concept drift) changes from what the model trained on, quietly degrading accuracy. Detect it by monitoring input distributions and prediction quality over time, and retrain or alert when they shift beyond a threshold.

Why use Docker for machine learning?

Docker packages your code, libraries, and system dependencies into one image that runs identically on your laptop, in CI, and in production. That eliminates 'it works on my machine' failures, which are especially common in ML because of heavy, version-sensitive dependencies.

What does MLflow do?

MLflow tracks experiments — logging parameters, metrics, and artifacts so you can compare runs — and provides a model registry to version and stage models for deployment. It's the system of record that keeps ML work reproducible instead of scattered across notebooks.