RN startup — phases, what cuts each, how to measure

react-native · folio

In one line: A cold start runs native init → JS runtime → bundle load → module factories (the top-level code of every imported module) → first render/mount = TTID → real content = TTFD/TTI. Every optimisation is “do less before the first frame”: precompiled bytecode, lazy modules, a thin root, deferred SDKs. Measure it on a release build on a low-end Android.

Download PDF Print view LaTeX source

RN startup — phases, what cuts each, how to measure — figure 1

How it works

  • Cold = new process (everything above); warm = process alive, activity recreated; hot = brought to front. Optimise and report cold.
  • Hermes compiles JS to bytecode at build time (release builds only); bytecode is loaded on demand, no parse. Hermes V1 (opt-in 0.82, default 0.84): Expensify measured Android low-end -3.2% bundle load / -7.6% TTI, iOS -9% / -2.5%.
  • Module factories: Metro wraps each module in a function; import runs it at first require. A root that imports every screen runs them all at start. Inline requires move each require to its first use, so unused paths never execute. RN CLI inlines requires by default; Expo: enable inlineRequires in metro.config.js.
  • RAM bundles (indexed/file bundles, parse on demand) — legacy, not supported with Hermes; Hermes’ on-demand bytecode replaces them.
  • TurboModules initialise on first access (legacy modules all at start); a module touched in the root still costs at startup.
  • Splash: iOS launch storyboard; Android 12+ always draws the system splash (installSplashScreen(), setKeepOnScreenCondition); Expo: SplashScreen.preventAutoHideAsync() → hideAsync() when the first screen has laid out.

Example — lazy paths + marks

// index.js - first line executed
performance.mark('js-start');
// screens off the first path load on navigation
const Settings = lazy(() => import('./Settings'));
let Charts = null;                 // lazy require
const openCharts = () => {
  Charts ??= require('./Charts').default; nav('charts'); };
function Home() {
  const feed = useFeed();          // cache first, then network
  useEffect(() => { if (feed.ready)
    performance.measure('ttfd', 'js-start'); }, [feed.ready]);
}  // PerformanceObserver ships 'measure' entries to analytics

Where the time really goes

  • Native: analytics/crash/ads SDKs in onCreate / didFinishLaunching, synchronous storage reads, big DI graphs — defer until after first frame or init on a background thread.
  • JS: moment/lodash-size deps, i18n loading every locale, a store rehydrating megabytes before render, polyfills, JSON config parsing.
  • First render: a navigator that builds every tab, providers that compute on mount, images decoded at full size.

Bundle analysis

  • npx react-native-bundle-visualizer@latest (--expo for Expo) — source-map-explorer over Metro’s output, treemap per package.
  • Expo Atlas: EXPO_ATLAS=true npx expo export, then npx expo-atlas .expo/atlas.jsonl.
  • source-map-explorer on a Hermes export needs --no-bytecode.

Traps

  • Timing a debug build: bundle streamed from Metro, no bytecode, dev checks — startup numbers mean nothing.
  • Inline requires change evaluation order: modules with side effects (polyfills, global setup) break — blockList / nonInlinedRequires.
  • A long splash makes TTID look great; users feel TTFD. Report both.
  • “Use RAM bundles” on a Hermes app — not supported, and not needed.
  • A lazy screen with module side effects (subscribing, globals) defeats laziness.

Remember

Less native init · bytecode · lazy modules · thin root · cached first paint — and measure TTID and TTFD on a release build.

Report a problem ·Discuss this folio