Design an experiment to measure latency impact

You work on YouTube. Engineering proposes a change that reduces video start latency by ~100 ms for some users, but it might increase error rate and change buffering behavior. Design an online A/B experiment to estimate the causal impact of latency on downstream business metrics. In your answer, cover: 1) **Hypothesis and metrics**: pick a primary metric and 2–3 guardrails. Explain tradeoffs (e.g., sensitivity vs. business relevance). 2) **Experimental unit & randomization**: user-level vs. device-level vs. request-level; how to avoid interference and contamination. 3) **Power / MDE**: how you would estimate sample size and duration (what inputs you need; how to handle heavy-tailed watch-time). 4) **Variance reduction**: at least two methods (e.g., CUPED, stratification) and when they are appropriate. 5) **Analysis plan**: how you would compute the treatment effect, handle multiple metrics, and interpret conflicting movements (e.g., watch time up but errors up). Assume: users are global, traffic varies by time-of-day/day-of-week, and latency changes may have heterogeneous effects by network type (WiFi vs. cellular).