Case Study: Amazon Ads achieves ultralow-latency machine learning with Amazon Web Services

Delivering Ultralow Latency Machine Learning for Amazon Ads

Amazon Ads needed a way to deliver machine learning at massive scale while keeping latency ultralow. Its challenge was twofold: understanding stable product similarity with deep learning embeddings and capturing fast-changing shopping trends in near real time. To support this, Amazon Ads used Amazon Web Services, including Amazon ElastiCache, Amazon Kinesis, and Amazon Simple Queue Service (Amazon SQS).

Amazon Web Services helped Amazon Ads build a scalable hybrid architecture that caches popular product embeddings locally and less popular ones remotely, reducing network cost and supporting hundreds of millions of requests per second. AWS also enabled near-real-time trend processing with Kinesis and prioritized feature publishing with SQS, while Amazon Ads’ bot-traffic handling reduced noise in traffic analysis. The result was a more efficient, cost-effective system that delivers relevant recommendations at huge scale with ultralow latency.

Amazon Ads

Shenghua Bao

Senior anager of Applied Science