Lakebase and Serving Endpoints Implementation in Real-Time Pricing

Klaudia Ratkowska17 Jun 2026

Problem

ERGO Hestia, a Polish insurer, runs more than 100 pricing models on more than 1,000 variables. The old stack loaded Databricks tables into an external Azure PostgreSQL database, and the pricing engine read them through an adapter application that built SQL and cached results. Large data volumes could be loaded only at night, because the database was overloaded during the day even with the adapter. New models also went live only at night, and there was no real-time model monitoring.

Example from the article: a data scientist who saw a model degrade had to wait for the night and for a software engineer to deploy a new version.

Solution in brief

Serving moves into Databricks. Final tables sync into Lakebase, and the pricing engine reads them through REST data endpoints. Models are registered in Unity Catalog and served from model endpoints, with CI/CD in Azure DevOps and inference tables for monitoring. Data endpoints respond in about 20 ms, and the team rolled it out in small steps.

Questions this case answers
  • How to serve real-time features and model predictions to a decision engine with low latency from the lakehouse instead of an external database and adapter layer
  • How to deploy and monitor a new predictive model in production during the day without an engineering release
  • How to serve real-time pricing model features and predictions with low latency from the lakehouse
  • How to replace an external Postgres database and adapter layer with a Databricks-native serving stack

Tools and methods

  • Lakebase — Databricks database that stores synchronized tables, with provisioned or autoscaling compute.
  • Feature spec and data serving endpoints — define source table, lookup keys and returned features.
  • Lakeflow Spark Declarative Pipelines — sync tables to Lakebase.
  • MLflow, Unity Catalog, model serving endpoints — train, register, version and serve models.
  • Azure DevOps and inference tables — automated deployment and request monitoring.
Continue reading — it's free

Google or a one-time email code. No password.