Uber

Search Context

Uber’s search infrastructure primarily serves Uber Eats — a food, grocery, and retail delivery marketplace operating globally. The search problem is inherently geo-aware: results must be filtered and ranked by delivery radius, ETD (Estimated Time of Delivery), and store availability in real time.

The same infrastructure powers multiple surfaces: Home Feed, Search, Suggestions, Ads.

Search Challenges

• Geo-awareness at scale: sharding a global index by location without hotspots
• Multi-category catalog: restaurants (hundreds of items) vs. grocery (100k SKUs/store) require different index layouts
• Latency under expansion: growing delivery radius from 10–15 min to 50+ miles multiplied candidate set size
• Real-time constraints: ETD changes constantly; indexing must reflect current delivery windows

Tech Stack

• Custom search platform built on Apache Lucene with Lambda architecture (Spark batch + Kafka streaming)
• H3 hexagonal geosharding for geographic partitioning
• Four-layer architecture: Base Search → NER Query Builder → Catalog API → Search API + Algorithm Gateway

People

• Janani Narayanan — ML Engineer
• Karthik Ramasamy — Senior Staff SWE

Use Cases

• Uber Eats - Scaling Search for Food Delivery

Published Articles

• Optimizing Search at Uber Eats
• Food Discovery with Uber Eats - Building a Query Understanding Engine