From Daily Batch to Real Time: An NRT Generative Recommender with Chronon

Feature platforms are increasingly asked to serve not just features, but model outputs — embeddings that stay fresh as user behavior happens, not as of last night's batch run. This talk introduces two new capabilities we built into Chronon, Airbnb's open-source feature platform, to close that gap: Push Mode and NRT Model Transform.

We cover Push Mode, which triggers a downstream pipeline the instant a feature updates instead of waiting to be polled, and NRT Model Transform, which runs model inference directly inside Chronon's streaming pipeline, so a model's output becomes just another feature. We'll walk through how the two compose with an existing offline backfill path to keep a single feature continuously fresh with no changes to serving, and the design decisions behind making both capabilities general enough for any model, not just one use case.

Attendees will leave with a reusable pattern for adding near-real-time model serving to their own Chronon deployment. We'll ground it in a production case at Airbnb — a generative recommender for search ranking — where the switch from daily batch to near-real-time cut staleness from roughly two days to under a minute. Both capabilities are already in our public repo.