Question

You operate a Vertex AI online endpoint serving a recommendation model for an e-commerce site. Product owners report that during peak evening traffic, some users experience slow recommendations, but the model's accuracy metrics look normal. You need to detect and be alerted when serving performance degrades so you can act before customers are affected. Which monitoring approach best addresses this operational concern?

Accepted Answer

Configure Cloud Monitoring on the endpoint's prediction latency and error-rate metrics, set an SLO on p99 latency, and create an alerting policy that fires when the threshold is breached.

Answer

Enable Vertex AI Model Monitoring for feature skew and drift, and alert when input distributions shift beyond a threshold.

Answer

Schedule a nightly batch prediction job and compare its output distribution against the previous night's results.

Answer

Enable request-response logging to BigQuery and manually query the logs each morning to review response times.