Performance

Core Web Vitals measure loading, responsiveness, and visual stability. Hit LCP ≤ 2.5s, INP ≤ 200ms, and CLS ≤ 0.1 at the 75th percentile of real users.

Keep main-thread tasks under the 50 ms long-task threshold: break up long work, yield with scheduler.yield(), offload heavy computation to a worker, and attribute what remains with Long Animation Frames in the field.

Serve images in modern formats (AVIF, WebP), at the right size for the viewport, with explicit dimensions. Images are the largest payload on most pages.

Native lazy loading defers off-screen images, iframes, and (recently) video until the user scrolls near them. Use loading="lazy" — but never on the LCP element.

Resource hints let you tell the browser what is coming. Preload the LCP image and critical fonts, preconnect to third-party origins, prefetch the next navigation.

An informational HTTP response that carries Link headers — preload, preconnect — to the browser before the final response is ready, putting server think-time to work.

Cache-Control tells browsers and CDNs how long to keep a response. Use immutable + max-age=31536000 for fingerprinted assets and short or no-cache for HTML.

Vary lists the request headers your server branched on when it chose this representation. Get it wrong and a cache hands one visitor's response to another; list too much and the cache stops working at all.

Send a validator — ETag or Last-Modified — on every cacheable response, and honour If-None-Match / If-Modified-Since so unchanged resources return an empty 304 instead of the full body.

The `No-Vary-Search` response header tells browsers and caches that some URL query parameters (tracking, UTM, sort order) do not change the response. The cached entry for the canonical URL is reused for variants — fewer fetches, better prefetch hits, less duplicate work.

Compress text responses with brotli where supported, gzip everywhere else. zstd is emerging. Don't compress already-compressed media.

Self-host WOFF2 fonts, subset them, set font-display: swap so text is readable while the font loads, and preload the critical face only when it styles above-the-fold content.

Inline the CSS needed for above-the-fold content and defer the rest. Render-blocking resources in <head> are the single biggest cause of slow first paint.

Choose the right script-loading attribute for every <script>: defer for app code, async for independent third-party, type=module for modern code. Bare <script> in <head> is always wrong.

Serve over HTTP/2 at minimum and HTTP/3 where you can. Multiplexing eliminates head-of-line blocking; QUIC removes TCP handshake delays.

Domain sharding and image sprites were workarounds for HTTP/1.1's connection limit; under HTTP/2 and HTTP/3 they hurt — drop them. Bundling is the nuanced one: stop concatenating to cut requests, start doing it to cut bytes.

Tell the browser which links to prefetch or prerender before the user clicks. Done well, navigations feel instant; done carelessly, you burn bandwidth on pages nobody visits.

Five resource hints — dns-prefetch, preconnect, preload, modulepreload, prefetch — cover every stage of the request lifecycle. Pick the right one for the job.

Animate between states (same-document) or between pages (cross-document) with a single CSS opt-in. Replaces ad-hoc SPA animation libraries with a platform primitive.

Keep pages BFCache-eligible so back/forward navigation restores them instantly from memory, with no reload, no hydration, and no repaint.

Use `content-visibility` with `contain-intrinsic-size` to skip layout and paint for off-screen content, and Intersection Observer to drive lazy behaviour, instead of scroll and resize listeners.

Use `contain: layout paint style` (or the `contain: content` shorthand) to tell the browser that an element's internals cannot affect the rest of the page, so reflow and repaint stay isolated to that subtree.

Drive CSS animations from scroll position or element visibility with `scroll-timeline` and `view-timeline`, replacing JS scroll-listener libraries with compositor-thread animation.

Use scrollbar-gutter: stable to reserve scrollbar space and stop horizontal layout shift between pages or states that overflow vs. don't.

On mobile, 100vh is taller than the screen because it ignores the browser's collapsing toolbar. Use dvh, svh, and lvh to size full-height elements to the viewport that is actually visible.

Use a previously served response, or a dedicated dictionary, as a Brotli/Zstandard dictionary so updated assets compress to a fraction of their size. Pure progressive enhancement over ordinary compression.

Server-Timing surfaces backend metrics — database time, cache hits, edge processing — in browser DevTools and to RUM via the PerformanceServerTiming API. Send it when you measure server-side latency, and keep the values free of sensitive infrastructure detail.