Performance and reliability

Budgets

• Measure before optimising. Set budgets early (page weight, JavaScript size, response time, query count) and enforce them in CI.
• Core Web Vitals at the 75th percentile: LCP ≤ 2.5 s, INP ≤ 200 ms, CLS ≤ 0.1. Aim for LCP ≤ 2.0 s on content sites.
• Lighthouse mobile lab scores: performance 90+ for content sites (85+ for apps), accessibility 95+, and SEO 100 for public sites.
• Set p95 latency targets for API endpoints and track them.

Front end

• Ship JavaScript only for interactive parts; content reads without it. Split code by route and lazy-load what is not needed for first render.
• The LCP element is in the initial HTML. An LCP image is never lazy-loaded and normally uses fetchpriority="high"; preload a CSS background LCP image.
• Images use AVIF or WebP, srcset and sizes, and explicit width and height (or aspect-ratio).
• Self-host and subset fonts with font-display: swap. Use at most 3 font files and preload the main one.
• Lazy-load off-screen images and iframes with loading="lazy". Preconnect only to critical third-party origins.
• Keep pages eligible for the back/forward cache: no unload handlers and no Cache-Control: no-store on normal pages.
• Reserve space for embeds, banners and late content so nothing shifts.
• Keep the main thread free: break up long tasks, debounce input handlers, and animate only transform and opacity.
• Every third-party script needs a written reason. Load it after the content, and only after consent where required.

Back end

• Avoid N+1 queries. Select only needed columns, index what you filter and sort on, and check query plans of slow queries.
• Paginate every list with a maximum page size.
• Move slow work (email, uploads, webhooks, LLM calls) out of the request into background jobs.
• Cache with a defined key, lifetime and invalidation. Content-hashed assets get max-age=31536000, immutable; HTML revalidates. Serve text with Brotli or gzip over HTTP/2 or HTTP/3.

Reliability

• Every network call has a timeout. Retry only idempotent operations, with backoff, jitter and a limit.
• Background jobs are idempotent and have retries and a dead-letter path.
• Keep servers stateless. Rate-limit public and expensive endpoints, and cap request and upload sizes.
• When a dependency fails, degrade gracefully: show cached or partial data with a clear message.
• Track errors with release versions and set up uptime checks before launch. Alert on what users feel (error rate, latency), not on every log line.
• Deploy from CI on main with health checks. Rollback takes minutes. Migrations run as an explicit step.