Design the location backbone of a ride-hailing app. Every active driver's phone posts a GPS fix on a 30-second timer — 360,000 drivers at the Friday-night surge — and every open rider app polls for the set of drivers near its map viewport.
Like presence data, a driver position is ephemeral: it expires in thirty seconds and the next ping replaces it, so the loss window is generous and nobody needs a quorum round. Unlike presence data, a city's dispatch runs on it — a lost copy shouldn't blind downtown mid-surge, and the fleet is big enough that "an availability zone failed" is a Tuesday, not a hypothetical. The contract splits the difference, and reading exactly where it's generous and where it isn't is half the problem.
The other half is that both streams are heavy at once. Twelve thousand writes per second reach the store no matter what you build in front of it. Twenty-five thousand reads per second sound like cache food — but only the hot downtown cells are shared between riders; the suburban long tail is effectively private, and the cacheable fraction stops at 40%. Your store tier carries a write floor and a miss stream simultaneously, and the interesting question is what a cache in front of it is actually worth.