We wanted to move an expensive GraphQL store write off the JS thread. A background worklet looked like the answer. But the cost to serialize the data into the worklet made it a no-go. Here is the experiment.
I built one React Native markets app with six data layers — Relay, TanStack Query, RTK Query, Zustand, Jotai, and hand-rolled vanilla — and measured the JS-thread CPU each one burns on a low-end Samsung. The full-fledged server-state libraries buy you real features, but on every cache write (including every live-price tick) you pay for them. Here's the bill.
I benchmarked Animated (native driver), Reanimated 4, and react-native-ease across four animation types on Android — measuring UI + JS frame drops, per-thread CPU, and memory. The winner depends on whether a gesture is involved — plus what Worklets Bundle Mode does to Reanimated's memory.
A bare RN app vs a blank Expo app — the fixed cost of adopting Expo is ~36 ms and ~13 MB RAM. Then what each extra Expo module adds, and why the Hermes profiler over-attributes 100–240 ms to @expo.
Press → paint measured across four libraries on a trivial screen and a heavy 24-row list. JS call stacks reveal each library's architecture — and why 'first frame' means very different things when the destination actually has work to do.
Why Expo Router costs 3× cold start and RAM, what React Navigation adds on top of rn-navigation, the Hermes hot-function breakdown for each library, and a controlled experiment isolating Reanimated.
Four navigation libraries, one identical app, benchmarked on Android with Perfetto Systrace and the Hermes CPU profiler. Cold start, RAM, press-to-paint — and what's actually behind the numbers.
Verifying whether Expo Router's file-based routing actually solves the eager-aggregator problem — require.context getters, getComponent thunks, and a probe proving route discovery loads zero screens.
We turned on inlineRequires on a large production app and the startup profile barely moved. The reason: eager aggregators — modules that require dozens of others at init time, which the transform structurally cannot defer.
