You received a training-serving skew alert from a Vertex AI Model Monitoring job running in production. You retrained the model with more recent training data, and deployed it back to the Vertex AI…

Using a lower sampling rate would reduce the amount of data analyzed, potentially masking actual skew or making detection less reliable, rather than resolving the root cause of the alert.

The model monitoring job compares live serving data against a baseline derived from training data; if the model was retrained with more recent data, the monitoring job's baseline must also be updated to this new dataset to accurately detect skew and avoid false positives from comparing against an outdated reference.

Temporarily disabling the alert only hides the problem without addressing the underlying misconfiguration of the monitoring job's baseline or the potential for actual skew.

Disabling the alert is a temporary workaround that does not fix the monitoring configuration, which is comparing serving data against an incorrect or outdated training data baseline.