Inside Uber's Real‑Time Dispatch: How the Company Scales Its Marketplace

Statistics

Uber's geospatial index aims for one million writes per second, with reads much faster.

The dispatch system runs on thousands of nodes.

Platform

Node.js, Python, Java, Go, iOS and Android native apps.

Microservices, Redis, Postgres, MySQL, Riak, Twemproxy, Google S2 library, Ringpop, TChannel, Thrift.

Overview

Uber connects passengers and drivers; the core challenge is real‑time matching of dynamic supply and demand.

The dispatch system is a real‑time market platform that uses mobile phones for communication.

New Year’s Eve is the busiest period of the year.

Architecture Overview

Clients run native apps; the backend serves mobile traffic over the public internet.

Dispatch (DISCO) coordinates drivers and riders, built mainly in Node.js.

Scaling relies on a consistent‑hash ring and a gossip protocol (Ringpop).

Geospatial Index

Designed for one million writes per second from driver updates every four seconds; reads must be significantly faster.

Uses Google S2 library to partition the earth into level‑12 cells, each identified by a 64‑bit ID.

Cell IDs serve as partition keys for location updates and queries.

Routing and Matching (DISCO)

DISCO matches supply and demand using geo‑supply and geo‑demand indexes.

Considers ETA, empty‑drive reduction, and overall wait time.

Supports future planning, route changes, and multi‑product requests such as rides, parcels, and food delivery.

Scalable Dispatch

Node.js processes are stateful; scaling is achieved with Ringpop’s gossip‑based consistent‑hash ring.

TChannel provides high‑performance RPC, about twenty times faster than HTTP, and supports Thrift.

Availability and Failure Handling

All operations are idempotent and retryable; services are partitioned into small, replaceable units.

Data‑center failover uses driver phones as the source of travel state, allowing seamless continuation after a switch.

Limitations

Node.js fan‑out can cause high latency under heavy load.

Cross‑server request cancellation in TChannel mitigates latency spikes.