Predictive preloads for ultimate web performance?
summary
A predictive preload system is A/B tested on hundreds of millions of users and combines client-side service workers that fetch an anonymous cached HTML in parallel with a personalized backend response, then merge them; simple triggers like hover preloads hit ~71% recall but only ~6% precision, while Markov-chain models reach ~16–33% recall depending on top-K; production A/B tests halve median time‑to‑first‑byte and substantially improve LCP across real ecommerce traffic.
AI-written summary · about
chapters (13)
- 0:00Predictive preloads intro
- 2:06Core web bottlenecks
- 4:11What is predictive preloading
- 6:41Preloading system building blocks
- 9:30HTML caching for preloading
- 10:13Speedkit with service workers
- 13:03Preloading success criteria
- 15:30Hover preloads baseline
- 18:18Markov chains for predictions
- 23:53Markov model limitations
- 26:41ML-enhanced models
- 29:34A/B test results impact
- 32:28Future directions & hiring
related talks
44mPredictive Prefetching
46mSpeculations about Web Performance
28mBetter than loading fast… is loading instantly
14mMaking the Web Instant
31mBe lazy, but with strategy
39mWhat’s new in performance?
27mpreconnect, prefetch, preload, pre what? An intro to Resource Hints
09mBuild a UI that Learns - Intelligent Prefetching