CS 329S: Machine Learning Systems Design - Stanford University

Delivers a comprehensive, sequenced curriculum covering the full lifecycle of machine learning systems from problem formulation through production deployment. The course uses a modular architecture organizing content into discrete units (data, modeling, evaluation, deployment, monitoring) with progressive complexity, enabling learners to build mental models of end-to-end ML system design rather than isolated techniques. Content is structured as interactive web pages with embedded code examples, case studies, and design patterns that scaffold understanding from foundational concepts to production-grade architectural decisions.

Teaches ML systems design through detailed analysis of real production systems and design decisions, using case studies that illustrate how companies solved specific architectural challenges. The curriculum embeds concrete examples (e.g., recommendation systems, fraud detection, autonomous vehicles) that demonstrate trade-offs between accuracy, latency, cost, and maintainability in actual deployed systems. This pattern-based learning approach helps practitioners recognize similar design challenges in their own work and understand the reasoning behind architectural choices rather than memorizing isolated techniques.

Teaches the design and implementation of data pipelines for ML systems, covering data collection, cleaning, validation, feature engineering, and data quality assurance. The curriculum explains how to structure data workflows to ensure reproducibility, handle data drift, manage data versioning, and maintain data quality at scale. This includes patterns for detecting and addressing data quality issues before they degrade model performance, and architectural approaches for integrating data pipelines with model training and serving systems.

Teaches how to evaluate ML models in production contexts, going beyond accuracy metrics to consider latency, throughput, cost, fairness, and business impact. The curriculum covers offline evaluation strategies, online evaluation (A/B testing, canary deployments), and how to choose appropriate metrics based on the business problem and user experience requirements. It explains the trade-offs between model complexity and inference cost, and how to structure evaluation pipelines that catch performance regressions before models are deployed to production.

Teaches the architectural patterns and design decisions for deploying ML models to production, covering batch serving, real-time serving, edge deployment, and model versioning. The curriculum explains how to structure serving systems for low latency, high throughput, and reliability, including patterns for A/B testing, canary deployments, and model rollback. It covers the trade-offs between different serving architectures (e.g., embedded models vs. microservices, synchronous vs. asynchronous serving) and how to integrate model serving with broader application architecture.

Teaches how to monitor ML systems in production, covering model performance monitoring, data drift detection, feature monitoring, and system health metrics. The curriculum explains how to structure monitoring to catch model degradation, data quality issues, and infrastructure problems before they impact users, and how to set up alerting and incident response for ML systems. It covers the unique challenges of monitoring ML systems compared to traditional software systems, including the difficulty of detecting model performance issues without ground truth labels.

Teaches how to optimize the cost and resource efficiency of ML systems across the full lifecycle, from data collection through serving. The curriculum covers trade-offs between model accuracy and inference cost, strategies for reducing computational requirements (model compression, quantization, distillation), and how to structure systems for cost-effective operation at scale. It explains how to measure and optimize the cost of data pipelines, model training, and serving infrastructure, and how to make architectural decisions that balance accuracy, latency, and cost.

Teaches how to identify, measure, and mitigate bias and fairness issues in ML systems, covering sources of bias (data bias, algorithmic bias, feedback loops), fairness metrics and definitions, and mitigation strategies. The curriculum explains how fairness concerns integrate into the full ML system lifecycle, from data collection through monitoring, and how to make trade-offs between fairness and other objectives (accuracy, cost, latency). It covers the business and ethical implications of biased ML systems and how to structure governance and decision-making around fairness.

Teaches a systematic framework for making architectural decisions in ML systems by analyzing trade-offs between competing objectives (accuracy, latency, cost, fairness, maintainability). The curriculum provides decision frameworks and heuristics for choosing between different architectural approaches based on system requirements and constraints, and explains how to structure decision-making processes that involve multiple stakeholders (engineers, product managers, business leaders). It covers how to evaluate architectural alternatives and make evidence-based decisions rather than defaulting to common patterns.