Databricks App Scores Transactions in Milliseconds Using Model Serving and Lakebase

A sample Databricks App (FastAPI + React) demonstrates scoring credit card transactions for fraud in real-time. It utilizes Model Serving with route optimization for low-latency inference and Lakebase Postgres for online feature and profile lookups.

Customer profile data, such as daily limits or international transaction toggles, can be updated in Lakebase and become immediately visible to the next transaction. This eliminates the need for redeployment or cache invalidation for profile changes.

Route optimization shortens the network path between the application and the inference container in Databricks Model Serving. This leads to faster, more direct communication, unlocking higher queries per second and more stable, lower latencies for interactive use cases.

Lakebase, a managed Postgres solution, now features autoscaling to handle demand fluctuations. This prevents the database from becoming a bottleneck and ensures stable performance as application load changes.

The application employs connection pooling and OAuth token rotation patterns for interacting with Lakebase. This maintains stable latency under load by keeping database connections open and utilizing secure, short-lived tokens for authentication.

The sample app achieved end-to-end response times of 27 ms at p50 and 37 ms at p95 across 5,000 requests, with median Lakebase feature lookups of 8.9 ms. The full application code is available on GitHub for users to deploy and test in their own Databricks workspaces.