Design a Nearby Places Search

Places are indexed by location with geohashes, S2 cells or a quadtree; a search reads the cells covering its radius including the neighbouring cells across boundaries, and cell size adapts to density so city cells stay small and desert cells large.

Searches are served from an in-memory or read-optimised index sharded by area and replicated per region; candidates are filtered (category, open now, rating) and ranked by true distance only after the cell lookup has bounded them.

When too few results are found, the search widens through rings of cells or coarser cell levels, stopping at 50 km, without rescanning cells it has already read.

Edits are written to a primary store and streamed (change data capture or a queue) to the index builders, meeting the 10-minute freshness target, while search keeps serving from its own copy if editing is down.

15 billion searches a month is about 5,800 a second on average and about 17,000 at peak; 200 million places at around 1 KB is about 200 GB, while the location index alone is a few GB and fits in memory per shard.

Every functional requirement in the brief is visibly served by something on the board, and the non-functional targets are addressed rather than ignored.

Components are labelled, data flows are drawn as connections between them, and the direction of each flow is unambiguous.