Uber - System Design

First created 2021年01月18日21:50

A system with Write heavy & Read heavy, Transisent data

ttl can be short

 

Requirements:

Drivers: tell the service the real-time location and availability

Customers: request a ride

Drivers: get notified / accpet a nearby customer

            finish a ride and get available for the next ride

Customers: be charged fee/ give feedback and ratings

 

DAU:

drivers: 500K

customers: 1M

 

QuadTree: not suitable

since Uber drivers will update their real-time location very frequently (every 3 seconds) from one grid to another

 

Push Model

Broadcast the location of one driver to all its nearby customers 

on publisher/subscriber model

 

Servers Infrastructure Picture

 

customer -(make ride request)-> [Demand Server] -> [Location Server] (different cell IDs within a circle and a given radius (using Google S2 library))

cell ID is performed as the sharding key. So the Location Service querys a few replicas.

All the nearby drivers' info returned to the Demand Server and the Demand Server do a aggregations and sorting.

Then the Demand Server calls Notification Server to pushes message to all the selected drivers. Wait for one of them to choose acceping this ride.

 

drivers -(accept/reject ride request)-> [Demand Server]

If accept, then the Demand Server calls Notification Server to pushes message to other candidate drivers that the request has been taken.

 

Credits Server is used for charging fee after the customer finishes the riding.  

 

Notification Server

websocket

 

Notification Server

websocket

 

Schema

RideDB

requestId (primary key) | customerId | driverId | latitude | longitude | status | created_timestamp

DriverDB

driverId (primary key) | latitude | longitude | status | updated_timestamp

CustomerProfileDB

customerId (primary key) | profile(email, phone, ...) | paymentInfo | rideHistory [requestId1,2,3,...]

 

Replication

primary, secondary, ...

 

Reference

[1] https://www.educative.io/courses/grokking-the-system-design-interview/YQVkjp548NM