Lessons learned from performance monitoring in Chrome
summary
Covers concrete lessons: metric design favors user‑facing moments and simplicity, Chrome builds largest contentful paint from paint events and filters splash and background candidates, and LCP correlates ~0.83 with Speed Index on HTTP Archive. Lab practice uses WPR recordings not VMs, buys same‑lot hardware, monitors CPU temperature, freezes Math.random/Date, and adds diagnostic metrics. First meaningful paint shows ~20% outliers. Regression detection treats performance as distributions and uses Mann‑Whitney U. first meaningful paint is real but often inaccurate.
AI-written summary · about
chapters (27)
- 0:00Intro
- 1:28Agenda for our trip
- 7:20Properties of a Good Top Level Metric
- 12:11Key Moments in Page Load
- 12:48Goal: When is Main Content Loaded?
- 14:34Main Content Painted Metric Priorities
- 15:29Brainstorm Simple Metrics
- 15:46And Test Them for Accuracy
- 16:41What About Splash Screens?
- 17:27What About Background Images?
- 18:24What is a "Text Paint"?
- 18:47What About User Actions During Load?
- 19:12Validating using HttpArchive
- 21:24Monitoring stages
- 21:41Lab Testing Pros and cons
- 23:46Benchmark Realism: Real-world Stories
- 25:14Better Reproducibility: Test Environment
- 26:34Better Reproducibility: Diagnostic Metrics
- 28:03Better Reproducibility: Reference Runs
- 28:47Better Reproducibility: Change Detection
- 31:06A/B Testing Pros and cons
- 31:34Should be "Controlled Experimentation"
- 33:26Best practices for experiment groups
- 34:11Real User Monitoring
- 37:13Metric Breakdowns
- 37:41Make RUM more like A/B Testing
- 38:28Takeaways
related talks
44mAiming for the Stars
49mEffortless Performance Debugging
20mReal metrics for front-end performance
29mPerfecting Perf: Rapid, Rigorous and Reproducible Performance Experiments
48mFrom Here to There
54mThe Latest in Metrics & Measurement
49mDemystifying web performance tooling and metrics
19mPerceived Performance, And Why It Matters