Airbnb Reduces Recommendation Data Latency from 2 Days to Under 1 Minute by Adding Push Mode and NRT Features to Chronon

Airbnb added two features, Push Mode and Near-real-time (NRT) Model Transform, to its open-source feature platform Chronon, significantly improving the real-time nature of search results. The existing system processed guest activities via a nightly batch job using a JourneyFormer-based sequence encoder, resulting in up to 2 days of data staleness due to scheduling and processing delays.

Key Technical Implementations

The newly introduced features were designed to overcome the limitations of batch processing.

• Push Mode: Emits lightweight notifications when streaming jobs commit new feature values. This allows downstream consumers to react immediately on an event basis without polling, ensuring that a guest's latest search or listing view instantly triggers embedding generation.
• NRT Model Transform: Executes inference for already deployed models (Transformer sequence encoder) within the streaming pipeline. This integrates the encoder into the real-time pipeline without duplicating training logic and records the results as features ready for immediate serving.

Performance and Impact

The integrated pipeline merges guest activity events with existing long-term/short-term sequence data, then sends signals via Push Mode to generate fresh embeddings through NRT Model Transform. This process yielded the following measurable improvements:

• Freshness: Data staleness was reduced from approximately 2 days to under 1 minute, with actual updates reflected within 10–30 seconds after activity.
• Quality: Offline evaluation showed a +1.67% improvement in NDCG compared to the daily-batch baseline. Online A/B tests showed an approximately 0.33% increase in uncancelled bookings.

Implications

This update demonstrates that the 'event reaction → state merge → immediate inference → serving' pattern can be generalized to other guest-facing models that rely on freshness. Push Mode and NRT Model Transform are general platform features published in the Chronon open-source repository, not exclusive to Airbnb, and are reusable for other near-real-time use cases.