ML systems
Module 24 of the System Design Masterclass. Every lesson here is free to read. No account needed.
- 7 lessons
- 2 h 51 min of reading
- 163The pieces of an ML system: data, features, training and servingThe parts of a system built around a learned model: events, training data, a training job, a versioned model file, a serving service and monitoring, what flows between them, and which decisions the system designer owns.24 minDepth
- 164Training-serving skew and feature storesWhy a model can be good in tests and worse live: the same feature computed in two ways, the future leaking into training data, and how a feature store with an offline and an online copy keeps one definition.23 minDepth
- 165Is the model any good? Offline metrics, A/B tests and feedback loopsHow to judge a model: scores on past data (accuracy, precision, recall, threshold), why they do not promise a live gain, A/B tests with sample size and whole weeks, shadow mode, feedback loops that bend the next training data, and drift.24 minDepth
- 166Serving a model at scale: batch or real-time predictions, latency and costBatch against real-time predictions, a latency budget in milliseconds, model size and hardware, dynamic batching, cold start, canary and fallback, and cost per 1,000 predictions.24 minDepth
- 167Search ranking basics: retrieve many, then rank a fewHow search puts results in order: a cheap first stage finds a few hundred candidates, a model scores them with more features, labels come from clicks, and NDCG and long clicks tell you if the order got better.25 minDepth
- 168Recommendation systems basics: candidates, embeddings and a rankerHow a recommender answers what a user should see next with no query: candidate sources, embeddings and nearest-neighbour search, a ranker, a re-rank for variety, cold start, exploration, and batch against real-time.23 minDepth
- 169Case study: Feed ranking end to endOne full design from a post to a ranked feed: candidates from follows and recommendations, features, a light and a heavy ranker, rules, logging, daily training, experiments, a time budget, costs and fallbacks.28 minDepth · Case study
Want help while you learn?
Reading is free for everyone. A free account adds the AI tutor, quizzes, notes and your progress.