Problem
Airbnb ranks search results using a Transformer-encoded embedding of each guest's bookings, reviews and browsing. That embedding was refreshed by a nightly batch job. Activity from earlier the same day did not appear until the next run, and processing lag pushed effective staleness to nearly two days. A guest actively planning a trip therefore searched with an outdated picture of their intent.
Example from the article: a guest browses a dozen listings on a Tuesday afternoon and searches again that evening, expecting the ranking to reflect that browsing.
Solution in brief
Airbnb added two capabilities to its open source feature platform Chronon. Push Mode makes a streaming job announce each committed feature value, and near-real-time (NRT) Model Transform calls the deployed sequence encoder inside the same streaming flow. A guest event now triggers a fresh embedding in the store the ranker already reads, cutting typical staleness from about two days to under a minute.
Questions this case answers
- How to refresh a heavy sequence-model user representation within seconds of new activity when it is currently recomputed in a daily batch job
- How to run a model trained offline inside a streaming feature pipeline while keeping its outputs comparable to the batch version
- How to keep user embeddings used by a ranking model fresh within seconds instead of a daily batch refresh?
- How to run an existing Transformer sequence encoder inside a streaming pipeline without duplicating its offline logic?
Tools and methods
- Chronon — Airbnb's open source feature platform, extended with the two new capabilities
- Push Mode — event-driven notification emitted when a streaming job commits a new feature value
- NRT Model Transform — calls a deployed model from inside the streaming pipeline and writes the output back as a feature
- Transformer sequence encoder — turns a guest's long-term and short-term event sequence into an embedding for ranking
Google or a one-time email code. No password.