Yelp

Location-Based Reviews

Design a local business directory where people find places near them, read and write reviews, and see a rating they can trust.

Overview

A local review service answers one question most of the time: what is near me that matches what I want? The user sends a location, a search term such as 'coffee' and some filters, and expects a ranked list of businesses in a fraction of a second. Everything else (business pages, reviews, photos) is ordinary read-heavy web serving.

What makes it different from plain text search is the location. A normal database index sorts along one dimension, but "near" needs two (latitude and longitude). The core of the design is a geospatial index: a way to turn a 2D location into keys that a normal index can sort, so a proximity query reads a small part of the data instead of all of it.

The second theme is trust. Reads outnumber writes by about a thousand to one, so the average rating is computed once when a review is written, not on every read, and reviews pass through filtering before they count.

Questions to ask first

How many businesses and searches? — 10 million businesses; 200 million searches a day
What does a search combine? — A location and radius, optional text and category, filters such as price and open now
How fresh must results be? — A new business or edited hours may take a minute to show in search; a new review shows on its page within seconds
Where do users search? — Worldwide, mostly in cities, so business density varies a lot
Are reviews moderated? — Yes: automated filtering decides which reviews count toward the rating

What you will be able to do

Explain how geohash, quadtree, S2 and H3 turn 2D locations into indexable keys, and pick one.
Walk a proximity query from cells to candidates to exact distance to ranking, including the cell-boundary trap.
Keep an aggregate rating correct without recomputing it on every read.
Size the system and say why the search index, not the database, is the part to scale.

Key concepts

Geohash
Quadtree
S2 and H3 cells
Neighbor cells
Haversine distance
Bounding box
Search index
Pre-computed ratings
Smoothed average
Review filtering

Functional requirements

Search businesses near a location, with text, category, price and open-now filters
Show a business page: details, hours, photos and reviews
Write one review per business, with 1 to 5 stars and text
Show an aggregate rating and review count on every result
Upload photos for a business

Non-functional requirements

Search responds in under 200 ms at p99 (a target we chose)
Business pages respond in under 100 ms at p99 from cache (a target we chose)
A new review appears on the business page within seconds; search ranking may lag by a minute
Read paths stay up if the review pipeline is down (degraded freshness, not an outage)
Ratings reflect only reviews that pass filtering

Interview questions

How do you find businesses within 1 km of a user?
Why not just index latitude and longitude?
Geohash, quadtree, S2 or H3?
How do you keep the average rating correct and fast?
How does search stay fresh when businesses change?
What happens in a very dense area like downtown Manhattan?
How would you scale this to many regions?