{"slug": "airbnb-cuts-recommendation-staleness-from-2-days-to-seconds-with-chronon", "title": "Airbnb Cuts Recommendation Staleness from 2 Days to Seconds with Chronon", "summary": "Airbnb extended its open source Chronon feature platform with Push Mode and Near Real Time (NRT) Model Transform, cutting sequence-recommender embedding staleness from nearly two days to under a minute, with typical updates arriving in 10 to 30 seconds. Airbnb reported a 1.67% offline NDCG improvement and roughly a 0.33% increase in uncancelled bookings in online A/B testing. The work builds on Airbnb's JourneyFormer Transformer sequence encoder, which previously ran as a daily batch job; the new streaming path runs the same deployed model inside the pipeline and writes embeddings to the existing low-latency store, leaving the online ranking path largely unchanged. Airbnb has contributed both capabilities to the public Chronon repository.", "body_md": "Airbnb has extended its open source Chronon feature platform with [Push Mode and Near Real Time (NRT) Model Transform](https://medium.com/airbnb-engineering/the-guest-journey-updated-in-real-time-extending-airbnbs-sequence-recommender-with-chronon-8f1582578553) to update sequence based recommendation embeddings within seconds of guest activity. The change reduced embedding staleness from nearly two days to less than a minute, with typical updates arriving within 10 to 30 seconds. Airbnb also reported a 1.67% improvement in offline NDCG and an approximately 0.33% increase in uncancelled bookings in online A/B testing.\n\nThe work builds on Airbnb’s JourneyFormer based sequence recommender, which uses a Transformer encoder to represent a guest’s booking, review and browsing history. Previously, the encoder ran as a daily batch job, generating embeddings that were stored in a low latency store. During a search, the ranking system retrieved the latest embedding and combined it with the live query, keeping the Transformer outside the latency sensitive serving path. However, activity occurring after the daily batch could remain absent from the embedding until the following run, with processing delays bringing effective staleness close to two days.\n\n*Airbnb’s initial architecture showing the daily batch sequence encoder, embedding store and online ranking path. (Source : [Airbnb Blog Post](https://medium.com/airbnb-engineering/the-guest-journey-updated-in-real-time-extending-airbnbs-sequence-recommender-with-chronon-8f1582578553))*\n\nAirbnb addressed the freshness problem by adding two capabilities to [Chronon](https://chronon.ai/). Push Mode causes a streaming job to publish a notification when a feature value is committed, allowing downstream processing to react to updated guest activity rather than polling or waiting for a scheduled job. NRT Model Transform allows an already deployed model to run inside that streaming pipeline. Airbnb uses the same Transformer sequence encoder for the near real time path, avoiding a separate inference implementation for the streaming workflow.\n\nThe resulting pipeline captures events such as searches and listing views, combines them with the guest’s existing short and long-term sequence data, and triggers model inference when the updated feature state is available. The generated embedding is then written to the same low-latency store used by the batch system. Consequently, the online ranking path remains largely unchanged while the representation it retrieves becomes more current.\n\n[Pengyu Hou](https://www.linkedin.com/in/pengyuhou/), Senior Software Engineer, Infrastructure at Airbnb, described the platform change as addressing the gap between event-driven data and model inference:\n\nNRT Model Transform lets Chronon call an already deployed model ... directly from within the streaming flow.\n\nThe approach follows a broader feature serving pattern in which predictions can be updated when underlying features change rather than computed on every user request. Feast, another open-source feature platform, documents cached prediction architectures where inference is triggered by feature updates, trading additional writes and pipeline complexity for lower serving latency. Chronon instead integrates streaming computation and model transformation into its feature platform and emphasizes consistency between offline and online computation.\n\n*Airbnb’s near real-time Chronon architecture* *(Source: [Airbnb Blog Post](https://medium.com/airbnb-engineering/the-guest-journey-updated-in-real-time-extending-airbnbs-sequence-recommender-with-chronon-8f1582578553))*\n\nAirbnb has contributed Push Mode and NRT Model Transform to the public Chronon [repository](https://github.com/airbnb/chronon/). Chronon supports both batch and streaming computation and provides online feature serving designed to maintain consistency between training and production inference.", "url": "https://wpnews.pro/news/airbnb-cuts-recommendation-staleness-from-2-days-to-seconds-with-chronon", "canonical_source": "https://www.infoq.com/news/2026/10/airbnb-chronon-recommendation/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=global", "published_at": "2026-10-28 13:53:00+00:00", "updated_at": "2026-10-28 14:12:35.368138+00:00", "lang": "en", "topics": ["machine-learning", "ai-infrastructure", "mlops", "large-language-models", "ai-research"], "entities": ["Airbnb", "Chronon", "JourneyFormer", "Pengyu Hou", "Feast"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/airbnb-cuts-recommendation-staleness-from-2-days-to-seconds-with-chronon", "markdown": "https://wpnews.pro/news/airbnb-cuts-recommendation-staleness-from-2-days-to-seconds-with-chronon.md", "text": "https://wpnews.pro/news/airbnb-cuts-recommendation-staleness-from-2-days-to-seconds-with-chronon.txt", "jsonld": "https://wpnews.pro/news/airbnb-cuts-recommendation-staleness-from-2-days-to-seconds-with-chronon.jsonld"}}