Case study

Design Yelp Nearby Places

Search and rank nearby businesses by location, category, and reviews for 200M monthly users.

geospatialsearchreviews

Requirements

Functional: search businesses near lat/lng; filter by category, price, rating; read/write reviews
Sort by distance, rating, or relevance
Non-functional: search p99 < 200ms; 200M MAU; 100M businesses indexed
Reviews: 5M new reviews/month; display aggregated rating

Back-of-envelope Estimation

Metric	Calculation	Result
Search QPS	200M × 10 searches/mo / (30×86400)	~800/sec avg; ~3K peak
Business index size	100M × 2 KB	~200 GB metadata
Reviews storage	5M/mo × 1 KB × 120 mo	~600 GB text
Geo index	100M points	Geohash + Elasticsearch geo_shape

API Design

Endpoint	Description
GET /v1/search?lat=&lng=&radius=5km&category=pizza&sort=rating	Paginated results
GET /v1/businesses/{id}	Details, hours, photos, rating aggregate
POST /v1/businesses/{id}/reviews	{ rating, text } — authenticated
GET /v1/businesses/{id}/reviews?cursor=	Paginated reviews

Data Model

Entity	Index
businesses	business_id, name, lat, lng, categories[], price, avg_rating — ES + SQL
reviews	review_id, business_id, user_id, rating, text, ts — sharded by business_id
rating_aggregate	business_id, sum, count — updated async from review stream

High-level Design

Search path

Client ──► Search API ──► Elasticsearch (geo_distance + filters)
                              │
                    Redis cache (lat/lng geohash-6 + category + sort key)
                              │
                    Review Service (aggregate rating from cache/SQL)

Deep Dive: Geo Search in Elasticsearch

Index businesses with geo_point field. Query: geo_distance around user location, filter terms on category, range on rating. Sort by _geo_distance or custom score blending distance + rating + review count.

Cache key

Round lat/lng to geohash-6 (~1.2 km) for cache key — nearby users share results.

Deep Dive: Review Aggregation

New review → Kafka → aggregator increments sum/count atomically in Redis
Periodic flush to SQL for durability
Display avg = sum/count; bayesian average for low-review businesses to reduce noise
Spam detection ML before publishing

Failure Modes & Monitoring

SLO	Target
Search p99	< 200ms
Rating staleness	< 5 min after new review
Index freshness	< 1 hour for new businesses

System Design