← Back to Index

1. Introduction

What is ML System Design?

ML System Design is the practice of architecting production-grade machine learning systems that reliably deliver predictions at scale. It encompasses data pipelines, training infrastructure, model serving, monitoring, and the operational processes that keep an ML system healthy over time.

Unlike building a model in a notebook — where the goal is to maximize accuracy on a static dataset — ML system design focuses on the end-to-end lifecycle: how does data flow from source to feature to prediction to feedback?

Why ML System Design Matters

ML Systems vs Traditional Software Systems

AspectTraditional SoftwareML Systems
BehaviorDeterministic, rule-basedProbabilistic, data-driven
TestingAssert expected outputEvaluate metrics (AUC, precision) on holdout data
FailuresExceptions, crashes, bugsSilent quality degradation, drift, stale features
MonitoringUptime, latency, errors+ prediction distribution, feature drift, model accuracy
DependenciesLibraries, databases, services+ Training data, features, model artifacts, ground truth labels
RollbackDeploy previous versionModel rollback + feature rollback + possibly data rollback

2026 State of ML Systems

Exercise: Think of an ML-powered feature you use (e.g., YouTube recommendations, spam filter, facial recognition). List three things that could go wrong in production that are not related to model accuracy. For each, suggest a way to detect it.