Case Study: Whatnot achieves real-time personalized recommendations with Chalk

Whatnot powers 99.9% real-time personalized feeds with Chalk

Whatnot, a live shopping marketplace, faced challenges with its batch-based machine learning system as it scaled. The system suffered from cold starts delaying personalization by up to 24 hours, dropping prediction coverage to ~90%, and omitting real-time user signals. To overcome these issues and power its real-time recommendations, Whatnot turned to the vendor Chalk and its unified feature platform.

Chalk provided a solution by serving as both a feature store and a real-time compute engine. This enabled Whatnot to compute over 300 million features per second with low latency, delivering fresher, personalized experiences. The results were significant: cold start delays were reduced from 24 hours to under one hour, prediction coverage increased to 99.9% of users, and model iteration time sped up from weeks to days.

Whatnot