Problem
Whatnot is a livestream shopping marketplace. Its home feed recommends livestreams, which are ephemeral: yesterday's shows are useless, and a show is worth promoting only while something exciting is happening in it. The feed therefore needs signals that are seconds old. The team set three speed targets: signals ready for retrieval under 3 seconds after the event, under 50 ms to retrieve personalized candidates, and fast iteration on features and models.
Example from the article: livestreams have "lulls" between "waves", and the feed should promote a stream while its most exciting moments are happening.
Solution in brief
Whatnot built the stack on an event bus and two ways to prepare real-time features. In the pull model, recent events sit in an indexed store and are joined at query time. In the push model, a stream job aggregates events into denormalized entities that serve point lookups. A unified retrieval interface lets retrieval switch between backends.
Questions this case answers
- How to serve fresh, fast-decaying signals to a candidate retrieval stage within a tight per-request latency budget
- How to decide which real-time features to compute at query time and which to precompute in a stream so that both iteration speed and read latency stay acceptable
- How to recommend ephemeral content such as livestreams using signals that are only seconds old?
- How to retrieve personalized candidates for a home feed in under 50 ms while joining several fast-changing datasets?
Tools and methods
- Pull model — join real-time and batch datasets at feed-generation time in Rockset.
- Push model — aggregate events into rolling windows in a stream, then serve point lookups.
- Unified retrieval interface — one parameter set implemented by Elasticsearch DSL and Rockset QueryLambda, with fallback.
- Feature flags and dynamic config — change parameters and run parallel configurations for A/B tests.
- Elasticsearch — a sink for denormalized data.
Google or a one-time email code. No password.