- Learn
- Machine Learning
- MLOps and Production ML
ML-M15
MLOps and Production ML
Taking models from a notebook to a product: framing the problem, platform architecture, experiment tracking, feature stores, pipelines, serving, monitoring, compression and distributed training.
A model that scores well in a notebook is a small part of a working ML product. Around it sit data pipelines, feature definitions that must match between training and serving, scheduled retraining, a way to serve predictions within a latency budget, and monitoring that notices when the world changes under the model.
You will start before any model exists, by deciding whether ML is needed and which metric it should move. Then you will tour the parts of an ML platform, from experiment tracking and feature stores to training pipelines and the main serving patterns. The module ends with what keeps models healthy and affordable in production: drift monitoring and retraining, compression to make models smaller and faster, and how GPUs and distributed training make large models trainable at all.
Lessons
- ML-15.01 Framing ML problems for products Before building a model, confirm ML is needed: could a rule or heuristic do the job? Intermediate 12 min
- ML-15.02 ML system architecture overview A production ML system has data pipelines that produce features, a feature store, training pipelines, a model registry, serving infrastructure and monitoring, with separate offline and online paths that must stay consistent. Intermediate 14 min
- ML-15.03 Experiment tracking and reproducibility Experiment tracking records the code version, data version, hyperparameters, metrics and artifacts of every training run, so results can be compared and reproduced. Beginner 10 min
- ML-15.04 Feature stores and training-serving skew A feature store computes features once and serves them both to training (historical values) and to online inference (latest values), which prevents training-serving skew, where the model sees features computed differently in production. Advanced 12 min
- ML-15.05 Training pipelines and orchestration Training pipelines turn ad hoc notebooks into repeatable workflows: extract data, validate it, build features, train, evaluate and register the model if it passes quality gates. Intermediate 10 min
- ML-15.06 Model serving patterns Batch serving precomputes predictions on a schedule and stores them for lookup, which is cheap and simple. Intermediate 14 min
- ML-15.07 Monitoring, drift and retraining Models degrade as the world changes. Intermediate 14 min
- ML-15.08 Model compression Large models are expensive to serve. Advanced 12 min
- ML-15.09 GPUs and distributed training GPUs run thousands of operations in parallel, which suits the matrix math of deep learning. Advanced 16 min