What Happens in the Milliseconds After You Tap Pay
Databricks, Thursday, July 16th, 2026
A technical walkthrough of a Databricks fraud detection app scoring card transactions in under 50 milliseconds.
This engineering post walks through a real-time fraud detection application built on Databricks that scores credit card transactions in under 50 milliseconds.
The system pairs Model Serving with route optimization for faster inference, Lakebase managed Postgres for feature lookups, and connection pooling with OAuth token rotation to sustain performance at scale.
Benchmarks across 5,000 requests recorded a median round-trip latency of 27ms at p50 and 37ms at p95, with feature lookups staying under 14ms even at the 95th percentile. The authors ship a deployable sample application showing how these components combine to meet checkout-grade latency requirements while preserving governance.