iCreatAI News
Skip to content
Trending storyUpdating

Databricks releases real-time NBO recommender tutorial

1 reports1 sources1 hr agoUpdated

Understand this story

AI overview

Databricks has released an open-source Solution Accelerator that uses Feature Views to build a real-time next-best-offer recommendation system for retail banking, with the same feature definitions used for both training and online serving. The solution has two parts: a recommender based on batch features, and an extension that layers on streaming in-session intent; both are deployed end-to-end as runnable notebooks on Databricks. According to Databricks, under a minimal configuration, streaming events become visible to the model in about 0.1 seconds. Everything runs on serverless compute governed by Unity Catalog; the second part also requires a Unity Catalog Kafka connection and Amazon MSK connectivity to incorporate real-time session activity. Use requires a Databricks workspace, a configured Databricks CLI profile, and the accelerator's GitHub repository. Databricks says that before going to production, teams must still implement controls such as eligibility, consent, suitability, and suppression themselves.

AI overview based on source reports · Updated 39 min ago

Coverage timeline

Follow the coverage to explore different perspectives.

Nov 17
  1. Databricks Blog
    Build a real-time next-best-offer recommender using Databricks Feature Views

    Databricks describes an open-source Solution Accelerator that uses Feature Views to build a real-time next-best-offer recommender for retail banking, defining batch and streaming features once for both point-in-time-correct training and low-latency serving. The accelerator combines a LightGBM ranker trained via the Feature Engineering SDK with a route-optimized Model Serving endpoint that scores 40 offers per request, and in its smallest benchmark configuration streaming events became model-visible in about a tenth of a second with identical training and serving values. The ranking endpoint sustained roughly 56 requests per second and the Feature Serving endpoint roughly 390 RPS before p95 latency exceeded 50 ms, though the authors note the accelerator uses synthetic data and does not claim revenue uplift.

Activity for this story

Not enough continuous observations to show a trend yet.