data-fetching

Data fetching

Stack-agnostic principles, concrete examples in React 19 and React Native. The patterns port to Vue/Svelte and other query libraries; the judgments don't change.

Version targets: React 19, React Native 0.81 (current stable) / Expo SDK 54, TanStack Query v5. Snippets grounded against Context7 (/reactjs/react.dev, /tanstack/query, /facebook/react-native) at authoring time — see references/snippets.md.

The core question

Every data-fetching decision reduces to: is this server state or client state? Server state lives on a server, is fetched asynchronously, can be shared, and goes stale — it's a cache you borrow, not state you own. Treat it as a cache (dedup, revalidate, expire) and most loading/error/staleness bugs disappear. Genuine client state (form input, which tab is open) is the only thing you truly own with useState.

Principles

• Server state is a cache, not useState. The useEffect-fetch-into-useState pattern re-implements caching badly: no dedup, no revalidation, request races, stale closures, a refetch on every mount. Reach for a cache — TanStack Query / SWR, or React 19 use() + Suspense — that owns fetching, caching, and refetching.
• Every async read has four states. Render all of them. Loading, error, empty (resolved but zero rows), and success. A component that assumes data is present crashes on the first slow network or empty result. Make the states explicit — a discriminated result, or Suspense + an error boundary.
• Key data by its inputs; serve stale-while-revalidate. A query is identified by its key (['user', id]). Same key → shared, deduped, cached. Show cached data instantly and revalidate in the background; set staleTime so you don't refetch what's still fresh.
• Prefer declarative boundaries to per-component if (loading) ladders. React 19's use() reads a promise and suspends; a <Suspense> ancestor renders the fallback and an error boundary catches rejection. One place for loading, one for errors — not ternaries scattered through every component.
• Mutations: update optimistically, then reconcile. Show the result immediately (React 19 useOptimistic / a form action, or a cache write via onMutate), confirm against the server, and roll back on failure. Invalidate the affected queries so the cache re-syncs with the truth.
• Co-locate fetching with the component that needs it — but keep it out of presentational components. The data owner fetches; the presentational shell takes data as props (the component-design "separate layout from data" rule). Don't prop-drill server state through five layers — read it from the cache where it's used.
• Mobile parity. The same cache-not-state rules hold in React Native; the differences are platform wiring — refetch on app foreground (AppState) and on reconnect (NetInfo), there's no SSR, and every request needs a timeout/abort because mobile networks drop silently. See references/pitfalls.md.

How to use this skill

1. Run references/checklist.md against the read or mutation you're writing or reviewing.
2. Reach for a pattern in references/snippets.md (React 19 use() + Suspense, TanStack Query read, optimistic mutation, React Native wiring).
3. Check references/pitfalls.md before you ship — the useEffect fetch and the missing error state catch almost everyone.

Related

• Where data lives once fetched / sharing it across components → state-management
• The component that renders it → component-design
• A form that submits it → forms-and-validation
• Testing fetch/mutation behavior → frontend-testing
• Waterfalls, over-fetching, slow lists → frontend-performance