Evolving Feed Ranking at Whatnot

Ford Bohrmann, David Kawashima, Miguel Fernandez-Montes Cuberta30 Apr 2025

Problem

Whatnot, a livestream shopping marketplace, ranked shows in each buyer's For You Feed with daily batch predictions. Every day it scored watch and purchase likelihood for all buyer-seller pairs, which is n×m predictions. As users and sellers grew, compute and storage exploded, most scores went unused, and offline filtering of inactive users became a pile of hacks. The scores were also stale and blind to what a show sells right now. The feed carries the business: 50% of GMV comes from the For You Feed alone.

Example from the article: a buyer may want a show selling Air Jordans in their size now, but care much less once the seller switches to Crocs five minutes later. A score computed yesterday cannot see that.

Solution in brief

Move the ranking model from batch to online inference. At request time the feed service fetches the live shows, pulls features from a feature store, and scores only those buyer-show pairs. Features known only at request time travel with the request and are logged for training.

Questions this case answers
  • How to rank a fast-changing set of candidate items for each user when precomputing scores for all user-supplier pairs grows quadratically and ignores request-time signals
  • How to move a ranking model from scheduled batch scoring to online inference without training-serving skew
  • How to rank a feed of live shows when daily batch predictions for every buyer-seller pair no longer scale?
  • How to use request-time features, such as the category a seller just set for a show, in a ranking model?

Tools and methods

  • Online inference endpoint — scores only the live shows for one buyer at request time
  • Feature store (Chalk) — builds the training set and serves features at inference
  • Pointwise likelihood model — predicted buyer-show watch or purchase likelihood is the ranking score
  • Negative sampling — random (easy) plus shown-but-ignored (hard) negatives
  • Elasticsearch — retrieves the currently playing shows
Continue reading — it's free

Google or a one-time email code. No password.