Question

Your team deployed a fraud-detection model to a Vertex AI online endpoint serving thousands of predictions per minute. Six weeks after launch, the business reports that fraud recall has silently degraded. You need to establish a serving-side capability so that future degradation can be diagnosed by analyzing the actual prediction inputs and outputs the model received in production, without redeploying or modifying the model. Which approach should you implement?

Accepted Answer

Enable request-response logging on the endpoint to capture a sampled fraction of live prediction inputs and outputs into a BigQuery table for later analysis and drift detection.

Answer

Increase the endpoint's minimum replica count so more instances handle traffic, improving the fidelity of prediction results.

Answer

Switch the endpoint from online to batch prediction so all inputs and outputs are written to Cloud Storage automatically.

Answer

Add a client-side logging library in the calling application to store request payloads, since Vertex AI cannot log endpoint traffic.