Case Study: Yelp achieves real-time store visit predictions with Ververica's Apache Flink

How Yelp uses Flink for predicting store visits in real time

Yelp wanted to send timely, relevant mobile notifications that help users discover nearby local businesses, but needed to process tens of thousands of location pings per second in real time, handle out-of-order and late mobile data, and avoid duplicate notifications. Using Ververica’s Apache Flink stream processing platform, Yelp built a system for predicting store visits as users move near businesses.

Ververica’s Flink solution clusters location pings into visits, uses stateful processing and event-time handling to filter late data, and delegates machine learning inference to an external Python service via Async I/O. The result was a 10x increase in visit recall, the ability to scale predictions to thousands per second, a reduction from hundreds of Python instances to 14 Flink instances, and a 5x improvement in cost efficiency.

Yelp

Luca Giovagnoli