Data viz

Data Visualization

Library policy

• One charting library per repo. If the repo already has one installed, use it — do not add a second. Otherwise the default is Recharts for standard charts (bar/line/area/pie/scatter); escalate to visx only when a design needs lower-level control Recharts can't express, and to D3's math modules (scale/shape, React owning the DOM) only for genuinely bespoke visualizations. For plots that stay dense after aggregation, use uPlot (canvas). Each escalation is a deliberate, ask-first dependency choice, not a default.
• Charting libraries are heavy: the AGENTS.md dependency ask-first rule applies, and the chart bundle should be code-split with the dashboard route, not in the shared bundle.

Rendering & SSR

• Charts measure the DOM → render client-side. Next.js: next/dynamic with ssr: false, called from inside a Client Component wrapper — it errors in Server Components (see nextjs for pushing the boundary down). Astro: client:only island — the sanctioned exception to the astro skill's last-resort rule, since a DOM-measuring chart cannot SSR; the fixed aspect-ratio container below neutralizes its layout-shift risk. SPA: React.lazy.
• Reserve space while loading: fixed aspect-ratio container + skeleton — a chart popping in is CLS.
• Responsive sizing via the library's responsive container or one ResizeObserver — never window resize listeners.

Accessibility — a chart is information, not decoration

• Every chart has a text alternative: a one-sentence summary (what the data shows, not "a bar chart") plus access to the underlying data — a visually-hidden or toggleable <table>. The table covers the data as presented (the decimated/aggregated series, or link the raw dataset as a download) — never one row per raw point. Screen-reader users get the data, not a canvas mystery.
• Color is never the only encoding: pair palette with direct labels, patterns, or shape (see a11y). Use a colorblind-safe categorical palette defined as tokens, not library defaults.
• Hover-only tooltips hide data from keyboard/touch users — ensure the values are reachable elsewhere (labels, the table, or focusable points).

Performance

• ~1,000 points: aggregate or decimate server-side first — nobody reads 50k DOM nodes, and SVG dies long before that. Canvas rendering for genuinely dense plots.

• Live data: update via the library's data prop with stable references; don't re-mount the chart per tick.
• Large tables are virtualized (e.g. TanStack Virtual) past a few hundred rows; sorting/filtering of big datasets happens server-side.

Formatting & states

• All numbers/dates through Intl.NumberFormat / Intl.DateTimeFormat (locale-aware, handles compact notation) — never hand-built strings (see i18n for locale/currency rules).
• Timezones are explicit: store/transport UTC, render in a deliberately chosen zone, label it when ambiguous.
• Every chart designs its empty, loading, and error states — and each is a Storybook story (see storybook); "blank white rectangle" is not an empty state.