Web Performance

Purpose

Make pages fast on the devices people actually use. Performance work without measurement on representative hardware is guesswork, and it is usually spent on the wrong thing.

When to Use

• Core Web Vitals failing in field data.
• A page that feels slow, or a bundle that has quietly grown.
• Before launching anything that matters commercially.
• Diagnosing jank, layout shift, or slow interactions.

Capabilities

• Core Web Vitals: LCP, INP, CLS — diagnosis and remediation.
• Bundle analysis and code splitting.
• Image, font, and asset optimization.
• Caching and CDN header configuration.
• Runtime profiling: long tasks, layout thrashing, memory growth.

Inputs

• Field data (Chrome UX Report, RUM), not just a lab score.
• The bundle, the network waterfall, and a CPU profile.
• The device and network profile of the actual audience.

Outputs

• A ranked list of fixes with expected impact.
• Before/after measurements on throttled hardware.
• Budgets in CI so the improvement holds.

Workflow

1. Measure the field, not the lab — Lighthouse on a developer laptop tells you nothing about a mid-range Android on 4G. Start with real-user data and reproduce it with 4x CPU throttling and a Slow 4G profile.
2. Find the LCP element — It is nearly always a hero image or a heading blocked by a font or a script. Preload it; do not lazy-load it.
3. Fix INP by breaking up long tasks — Interaction latency comes from the main thread being busy. Find tasks over 50ms in the profile and yield, defer, or move them off-thread.
4. Fix CLS by reserving space — Every image needs width and height or an aspect-ratio. Every ad slot, embed, and late-loading banner needs a reserved box.
5. Cut the bundle — Analyze it. One date library, one icon set, or one un-tree-shaken utility library is usually a third of the payload.
6. Set the cache headers — Immutable, fingerprinted assets get a one-year max-age. HTML gets no-cache and revalidates.
7. Budget it in CI — Otherwise it regresses within two sprints.

Best Practices

• Lazy-loading the LCP image makes LCP worse. loading="lazy" belongs below the fold, never above it.
• font-display: swap prevents invisible text but causes a shift; size-adjust with a matched fallback metric prevents both.
• Third-party scripts are the largest performance cost in most real applications, and the least examined. Audit them; load them with async, in a worker, or not at all.
• A 200 KB JavaScript bundle costs far more than a 200 KB image: the image decodes off-thread, the JavaScript must be parsed and executed on the main thread.
• Preconnect to critical third-party origins, but only two or three — each one costs a connection.
• Do not optimize what the profile does not show. The bottleneck is rarely where intuition says it is.

Examples

Preloading the LCP image and the font that blocks it:

<link rel="preload" as="image"
      href="/hero-960.avif"
      imagesrcset="/hero-640.avif 640w, /hero-960.avif 960w, /hero-1440.avif 1440w"
      imagesizes="(max-width: 768px) 100vw, 960px"
      fetchpriority="high" />

<link rel="preload" as="font" type="font/woff2"
      href="/fonts/inter-var.woff2" crossorigin />

<img src="/hero-960.avif" width="1440" height="810"
     alt="" fetchpriority="high" />   <!-- dimensions reserve space: no CLS -->

Breaking a long task to fix INP:

// Before: 400ms of blocking work on click. INP is terrible.
button.addEventListener("click", () => {
  const results = items.map(expensiveTransform);   // blocks the main thread
  render(results);
});

// After: yield to the browser between chunks so input stays responsive.
async function processInChunks(items, chunkSize = 50) {
  const out = [];
  for (let i = 0; i < items.length; i += chunkSize) {
    out.push(...items.slice(i, i + chunkSize).map(expensiveTransform));
    await scheduler.yield();   // hand control back; the browser can paint and respond
  }
  return out;
}

Notes

• INP replaced FID as a Core Web Vital in March 2024. It measures every interaction, not just the first, and it is a far harder bar — most sites that passed FID comfortably do not pass INP.
• Compression matters more than minification: Brotli at level 11 on a 300 KB bundle typically beats aggressive minification by a wider margin than the minifier does.
• content-visibility: auto skips rendering work for off-screen content and is close to free on long pages. It is one of the highest-value single-property changes available.