Design a Nearby Friends Service

The prompt arrives sounding familiar: "When a user opens the app, show which of their friends are nearby, and notify them when a friend shows up. Everyone is moving all the time." Within seconds most candidates have drawn a quadtree. They have recognised the shape of the question — points on a map, a radius, a search — and reached for the index that shape implies.

That recognition is the trap. In a Yelp-style nearby search the data sits still and the reads dominate, so you build a geospatial index once and query it. Here every user is simultaneously a moving write and a reader, the data is worthless within a minute of being written, and the search set is not everyone near me — it is my friends, a list of about four hundred identifiers that the social graph already maintains. The friend list is the index. A candidate who notices that in the first ten minutes designs a different and much better system than one who does not.

The Trap Built Into a Familiar Shape

Two failures show up in this question again and again:

• The Index Reflex: builds a geospatial index over all ten million active sharers and runs a radius query on every app open, then intersects the result with the friend list. It can be made to notify on arrival — query the index at write time and intersect the candidates with the friend list — but the costing is brutal: thousands of candidate rows scanned per update against the roughly thirty-eight keyed cache reads the friend list needs, and every one of those scans touches the rows of strangers who were never party to the question. Signal: pattern-matched the shape instead of reading the problem.
• The Realtime Maximalist: a durable log, a stream processor, exactly-once delivery and sub-second updates — engineering for permanence and precision on a payload with a thirty-second shelf life, for a feature whose human-perceivable freshness need is about five seconds. Signal: cannot state what guarantee the product actually requires.

The passing performance computes the fan-out arithmetic early, lets that number choose push over search, and then spends the rest of the hour on the things that genuinely bite: notifications that flap when a friend cycles along the boundary, stadiums that empty into a single map cell, an off switch that has to erase and not merely stop, and ten million phones reconnecting at once.