Engineering Online Experimentation and ML Evaluations

Description

Online experimentation is now essential for modern software and machine learning teams. This book provides an engineer-first, end-to-end guide to building and operating production-ready experimentation platforms.

The book begins with Part I establishing the core foundations of credible experimentation, including hypothesis testing, power analysis, sample sizing, metric design, and common pitfalls such as peeking, multiple testing, and novelty or learning effects. Part II focuses on platform engineering—traffic and identity management, mutual exclusion, event and logging design, ETL/ELT pipelines, building a stats engine with SciPy and statsmodels, SRM detection, integrating deployments with feature flags and canaries, and setting up guardrail and health monitoring. Part III presents advanced designs that improve speed and sensitivity: sequential testing with alpha spending, bootstrap intervals for ratios and quantiles, A/B/n testing with ANOVA, interleaving for ranking systems, switchback and geo experiments, and multi-armed bandits. Part IV connects experimentation to ML workflows, covering offline, shadow, canary, and A/B evaluation pipelines; Bayesian optimization for adaptive experimentation; counterfactual and IPS methods for learning from logs; and safe retraining supported by strong governance.

What you will learn:

• Design trustworthy experiments with proper metrics, guardrails, α/power/MDE settings, and safeguards against peeking and multiple-testing errors
• Build a production-ready experimentation stack with assignment, identity/diversion, logging, ETL/ELT, a stats engine, and SRM checks
• Run advanced designs at scale, including sequential tests, bootstrap CIs, interleaving, switchback/geo experiments, and multi-armed bandits
• Evaluate ML systems from offline to online, leverage experiment logs for learning, and enable safe retraining with governance