How Uber Scales Its Real-Time Ride‑Sharing Platform: Architecture Secrets

Statistics

Uber's geographic index aims for millions of writes per second and several times that rate for reads.

The dispatch system runs on thousands of nodes.

Platform

Node.js

Python

Java

Go

Native iOS and Android apps

Microservices

Re‑sharding system

Postgres

MySQL

Riak

Twitter Twemproxy

Google S2 geometry library

Ringpop – consistent‑hash ring

TChannel – RPC multiplexing and framing

Thrift

Overview

Uber connects passengers and drivers as a real‑time market.

Challenge: build a dynamic supply‑demand system that works instantly for both sides.

The dispatch system matches riders and drivers via mobile devices.

New Year’s Eve is Uber’s busiest day.

Rapid technological progress has turned once‑fictional tools like GPS into everyday utilities.

Architecture Overview

Drivers and riders using native mobile apps drive the system.

The backend processes information between mobile devices.

Clients connect to the dispatch system to match supply and demand.

The dispatch system is written almost entirely in Node.js.

* Uber originally considered moving to io.js before the projects merged.

* JavaScript can be used for interesting distributed‑system work.

* The enthusiasm of developers enables rapid task completion.

Old Dispatch System

Limitations of the old system began to restrict growth.

Most of the system needed rewriting.

It was designed for single‑passenger rides, assuming one passenger per car and only mobile users, which hindered expansion to UberPool, food, and package delivery.

City‑level sharding worked initially but became hard to manage as more cities joined.

Failures in one component could cascade to others.

New Dispatch System

To solve city sharding and support new product types, separate supply and demand services were created.

Supply service tracks quantity and state of all supplies (vehicles, seats, wheelchair access, etc.).

Demand service tracks all ride requests and their requirements.

DISCO service performs scheduling optimization, matching supply and demand, predicting future availability, and using geographic indexes for both supply and demand.

Scheduling Flow

Vehicles send location updates every few seconds to the "geo‑by‑supply" index.

DISCO queries this index to find nearby candidate vehicles.

Candidates are sent to the routing/ETA service, which computes road‑distance based estimates.

ETA results are returned to the supply service and then to drivers.

Special handling is required for airport queues and multi‑passenger scenarios.

Geographic Index

Designed for high scalability: millions of writes per second, with read throughput several times higher.

Uses Google’s S2 library to partition the earth into hierarchical cells with unique IDs.

Both supply and demand entities are indexed by these cell IDs, enabling fast proximity queries.

Scaling is achieved by adding more nodes and replicas for write and read load.

Routing Goals

Minimize extra driving, reduce rider wait time, and lower overall ETA.

Prefer drivers already carrying passengers over idle drivers far away.

Predictive models improve decisions for shared rides, package delivery, and food transport.

Scaling the Dispatch Service

Built with Node.js; requires stateful services, so traditional stateless scaling does not apply.

Node’s single‑process model is extended across multiple CPUs and machines using Ringpop, a gossip‑based consistent‑hash ring.

Ringpop provides AP semantics (availability over strict consistency) and integrates with Uber’s RPC layer, TChannel.

TChannel, inspired by Twitter’s Finagle, offers high‑performance, multiplexed RPC, outperforming HTTP by ~20×.

Persistent connections are used for gossip and data forwarding.

Dispatch Availability and Fault Tolerance

All operations are designed to be retryable and idempotent.

Services are partitioned into small, isolated components to limit failure impact.

Backup data centers and driver‑phone‑based state snapshots enable rapid failover.

Cross‑region replication and request shadowing mitigate latency spikes.

Translation by Feng Yahua; Proofreading by Wendy. Original article: How Uber Scales Their Real‑Time Market Platform