MLOps and Production Machine Learning

About this course

How machine-learning systems are run for real. Learners build a reproducible training pipeline with experiment tracking and a model registry, add data contracts and validation gates, deploy models for batch and online inference, automate testing and continuous training, monitor for drift and decay, respond to model incidents, and choose platforms and GPU capacity on cost, reliability and team fit.

For: ml practitioner, software engineer, data professional.

Learning outcomes

On successful completion you will be able to:

1. 1Explain the ML lifecycle in production and the technical debt specific to ML systems, including data dependencies, feedback loops and train-serve skew.
2. 2Build reproducible training pipelines with experiment tracking, data and model versioning, validation gates and a model registry.
3. 3Deploy models for batch, online and streaming inference using appropriate serving patterns, containers and rollout strategies.
4. 4Design CI/CD and continuous-training workflows with automated tests and quality gates for code, data and models.
5. 5Monitor models in production for data drift, performance decay and operational health, and design the response.
6. 6Evaluate platform and tooling choices (feature stores, orchestrators, serving stacks, managed or self-hosted, GPU capacity) on cost, reliability and team fit.
7. 7Take ownership of a production model's operation, including incident response, rollback and the documentation auditors and regulators expect.

Modules and lessons

1. Week 1ML systems in production⌄

   • The ML system and its hidden technical debt
   • Experiment tracking, versioning and the model registrylab
2. Week 2Data and features for production⌄

   • Feature stores and point-in-time correctnesscase study
   • Data validation, contracts and orchestrationlab
3. Week 3Serving models⌄

   • Model serving patterns
   • Optimising inference cost and latencylab
4. Week 4Automation and quality gates⌄

   • CI/CD and continuous training
   • Testing ML systemslab
5. Week 5Monitoring and operations⌄

   • Monitoring models and detecting drift
   • Incident response, rollback and retraining
6. Week 6Platforms and economics⌄

   • ML platform architecture and tooling choices
   • GPU capacity and the cost of ML

Assessment

• Quiz (15%)

  Six weekly quizzes drawn from the lesson quiz banks; best five count.

• Lab (35%)

  Four labs with automated checks - registry promotion, pipeline validation gates, a load-tested model service and a CI pipeline that blocks a regression.

• Project (35%)

  Production-grade ML service with a training pipeline, registry, deployment, monitoring dashboard, alerts and an operations runbook.

• Design review (15%)

  Platform decision record and a postmortem for a simulated model incident, defended live to two mentors.

Pass mark 60%. Grades: Distinction, Merit, Pass.