How Uber Scaled Its Real-Time Ride-Sharing Platform: Architecture & Challenges

Uber's Early Architecture

Initially Uber used a simple three-tier architecture: a mobile client, PHP business logic, and MySQL storage. Drivers posted GPS coordinates every 4 seconds, which were stored in MySQL and queried by PHP to match riders.

Scaling the Architecture

To handle rapid growth, PHP processes were made multi-threaded, and the system added a message queue, Python APIs for business logic, and MongoDB for GPS logs, reducing load on the dispatch service.

Single-Point Failures and Redundancy

Master‑Slave hot‑standby, region‑based sharding, and single‑threaded processing were introduced to avoid a single point of failure.

Message Handling and Hot Upgrade

A simple request manager listening on port 9000 can be upgraded without dropping in‑flight messages by closing the port, allowing existing requests to finish before restart.

Ringpop Storage and Routing

Uber adopted Ringpop, a Cassandra‑like distributed platform where each node stores location data for a geographic region, enabling fast lookup and routing.

Communication Layer (TChannel)

TChannel provides high‑performance, language‑agnostic message forwarding, load balancing, and encapsulation of protocols.

Design Principles

Services must be retryable.

Services should be killable for fault‑injection testing.

Services should be fine‑grained.

Load Balancing and Ringpop

Ringpop also acts as a load‑balancer, eliminating a single load‑balancer failure point.

Latency “Bucket” Problem

When a composite request contains many sub‑messages, the overall latency is dictated by the slowest sub‑message, leading to high failure rates.

Mitigating Bucket Latency

Duplicate processing across parallel services with cancellation reduces tail latency.

Data‑Center Failover

Drivers cache encrypted summaries of their state; a new data center can request the summary from the driver to rebuild state instantly.

Key Tools

Google S2 for geographic indexing.

Ringpop for distributed storage and load balancing.

TChannel for remote procedure calls.

Reference: “Scaling Uber’s Real‑time Market Platform”, Bittiger.