Databricks ML Professional Practice Question 17: AI Gateway Inference Tables
Databricks Certified Machine Learning Professional · Test 2 · question 7 of 10
A custom model is served on a Databricks endpoint. For later drift analysis the team needs every request payload and the corresponding response landed in a Unity Catalog Delta table. Client applications must not change, and the team does not want to maintain logging code of their own.
What should they do?
Show answer and explanation
Answer:
AI Gateway-enabled inference tables are the current payload-logging feature for Model Serving. You turn them on in the endpoint's AI Gateway configuration, naming a catalog and schema, and the platform writes request and response payloads to a Delta table in Unity Catalog. It is configuration rather than code, it is invisible to callers, and because the output is a governed Delta table it feeds a data profile for drift monitoring directly. Databricks recommends it over the legacy feature because it covers custom model, foundation model and agent endpoints alike.
Option A names the superseded feature and inverts the recommendation. The legacy inference tables documentation is retired, and Databricks directs new endpoints to the AI Gateway version.
Option B is the maintenance burden the requirements exclude. Logging inside predict() couples the model artifact to a table schema, adds write latency to every scoring call, and has to be reimplemented in every model the team deploys.
Option C looks in the wrong place. Audit logs record that an endpoint was invoked and by whom, not the request and response bodies, so the payloads drift analysis needs are simply not there.
Reference: AI Gateway-enabled inference tables