Machine learning systems design, visualized
What happens around the model: the data it learns from, how it ships, and how it fails once real traffic changes under it.
Start with the overview and problem framing. The ML design study guides below are interview-style walkthroughs of complete systems.
Start with ML Systems: An Overview 17 topics
Start here
Beginner topics, in the order they build on each other.
- ML Systems: An OverviewBefore the model, a harder question: should this be machine learning at all? Step the five-part definition as a checklist, compare research against production on five axes, and watch one 3,000 ms request drag an average to 390 ms.Beginner18m
- ML System RequirementsConnect an ML metric to a business one with Netflix's take-rate, meet the four requirements every ML system owes its users, and follow the thirteen-step ad-serving workflow that ends by going back to step one.Beginner17m
- Framing ML Problems"Slow customer support is a problem, but it is not an ML problem." Turn a complaint into inputs, outputs and an objective function, then reframe an app recommender so adding a new app stops forcing a retrain.Beginner18m
Next
Intermediate topics.
- Data Engineering for MLWatch a row-major read touch every column to return four of them, price the same file as CSV and Parquet, and see stateful stream processing reuse 29 days of work a batch job throws away every night.Intermediate19m
- Sampling Training DataSelf-driving data is mostly from sunny Phoenix and the sunny Bay Area, and sentiment models learn from people willing to write reviews. Step the four nonprobability methods, then compute the odds that simple random sampling loses a 0.01% class entirely.Intermediate18m
- Labeling & Weak SupervisionAccurate phonetic transcription takes 400 times longer than the speech, so one hour is almost three months of work. Watch three annotators find three different answers, drag a labeling window that quietly undercounts clicks, and see eight hours of labeling functions match a year of hand labels.Intermediate19m
- Class ImbalanceTwo cancer-screening models, both with 90% accuracy — one catches 10 of 100 cancers, the other catches 90. Flip which class counts as positive and watch the same model's F1 move from 0.17 to 0.95.Intermediate19m
- Feature Engineering & Data LeakageAmazon had over two million brands in 2019, and three obvious encodings each fail one step later. Watch the hashing trick absorb an unbounded category space, then toggle scale-before-split to see a statistic carry the test set into training.Intermediate20m
- Model Development & Offline EvaluationAn F1 of 0.90 means nothing when the positive class is 90% of the data. Step the four phases, pick a baseline that gives a number meaning, calibrate a forecast, and watch two models swap places on overall accuracy while their per-slice scores never move.Intermediate21m
- Model Deployment & Prediction ServiceDeploying is easy if you ignore all the hard parts. Four myths against the Uber, Booking.com and Algorithmia numbers, batch against online prediction, a compression slider that reproduces the Roblox 7x result, and why the browser is an edge device.Intermediate22m
- The Human Side of MLThree properties separate an ML system from ordinary software, and each one creates a distinct UX problem. Pin Booking.com's filter suggestions with a consistency rule, watch the odds climb as you show more candidates, cap a latency tail with a backup model, and pick between two team structures.Intermediate22m
- Responsible AI in PracticeA grading model that matched human examiners at 60% agreement and failed anyway, and a heatmap of anonymized data that exposed military bases. Step the three Ofqual failures, flip a privacy default, and drag a compression slider until the headline metric and the long-tail metric come apart.Intermediate22m
Advanced
For when the rest feels solid.
- Data Distribution Shifts & MonitoringA model performs best the day it is trained. Separate software failures from ML failures, name the shift by which term of the joint distribution moved, break a degenerate feedback loop with a randomisation slider, and watch a detection window too short to see a weekly cycle.Advanced23m
- Continual LearningContinual learning is not per-sample updating, and it is not about how often you retrain. Slide a micro-batch size, toggle stateless retraining against the stateful training that cut Grubhub 45x, climb the four stages, and price data freshness with the book's own experiment.Advanced22m
- Testing in ProductionOffline evaluation cannot tell you a model works. Fan a request out to a shadow deployment, size an A/B test until the detectable lift shrinks, ramp a canary and abort it, redraw a team-draft interleaving list, and watch a bandit reach a verdict in 12,000 samples where A/B testing needed 630,000.Advanced22m
- MLOps InfrastructureMost companies are at "reasonable scale" — terabytes, not petabytes. Pick your scale band, climb the four layers, read compute by utilization instead of capability, drag a burstiness slider past the point where owning hardware wins, and separate cron from a scheduler from an orchestrator.Advanced24m
- ML Platform: Model & Feature StoresA DevOps engineer is paged and there are twenty data scientists. Rule out the four causes she cannot distinguish from the weights alone, click through the eight artifacts a model store should hold, separate the three unrelated problems called a feature store, and run the build-versus-buy staffing math.Advanced22m
Design walkthroughs
Interview-style study guides: requirements, estimates, API, data model and trade-offs. They open in the System Designer, which needs a desktop-width screen.