Databricks brings predictive analytics to SQL
Databricks announced four new analytical functions and an upgraded forecast in beta on October 14, 2026. The company maintains the models for analysts.
In this article
The announcement covers forecasting, unusual observations, contributions to metric changes, and predictions of numbers or categories. Analysts need not build their own models, Databricks says. In Genie, its everyday-language interface, only ai_top_drivers was available at announcement; the other four functions were coming later.
Choosing the forecast
The AWS forecast documentation requires Predictive AI Functions preview enrollment for version 2 and a Pro or Serverless SQL warehouse. Explicitly select version 2; the default remains version 1. Dated historical values produce future estimates with bounds. To use known future factors, such as planned prices, supply future rows containing those factors and leave the target values NULL, meaning missing.
Explaining a change
For ai_top_drivers on AWS, an administrator must enable the Predictive AI Functions preview. Requirements include Runtime 16.1 or newer and a Photon-enabled Pro/Serverless SQL warehouse or Photon-enabled Runtime cluster. Input supplies a numeric metric and a boolean label: true for test, false for control. The function ranks segments, or data groups, contributing to the metric change; rows missing the metric or label are dropped.
Finding unusual observations
The AWS anomaly function also needs administrator-enabled Predictive AI Functions preview access and Pro/Serverless SQL; Classic is excluded. Each time series, meaning observations over time, needs unique timestamps and at least one finite observed value. Preceding observations determine an expected value, prediction range and anomaly score. The first eight rows provide context and receive no score. During beta, each series allows at most 8,192 input rows; exceeding that limit fails the query.