
Model Monitoring in Gemini Enterprise Agent Platform (formerly Vertex AI) tracks deployed endpoints for statistical shifts in incoming prediction data and in how features drive model outputs. On the Professional Machine Learning Engineer exam, monitoring questions usually come down to four configuration decisions: choosing between drift and skew, distinguishing input features from feature attributions, balancing sampling rate against monitoring frequency, and knowing when a training baseline dataset is required.
Agent Platform Model Monitoring supports two monitoring targets, input features and feature attributions, and each can be checked for either drift or skew. Understanding what each target measures helps separate data changes from shifts in model behavior.
Feature drift occurs when the statistical distribution of input features changes over time while the underlying relationships learned by the model stay the same. Consider a fraud model where desktop traffic falls from 80% to 40% between 2012 and 2016 while mobile traffic rises from 20% to 60%. Even if the overall fraud rate remains 2.1% and average transaction values stay at $85 for desktop and $25 for mobile, the model was originally trained on 80% desktop data and is now scoring 60% mobile traffic. That distribution shift can still affect model performance and signal that retraining on more recent representative data is warranted.
Feature attribution drift tracks shifts in how much each feature contributes to individual predictions, such as through SHAP values. In a fraud detection model, login distance might start as the dominant factor with a SHAP value around 0.3 while transaction speed sits near 0.15. Later, transaction speed might climb to around 0.4 while login distance drops to around 0.25. That shift can appear when fraudsters adapt their tactics or when legitimate transactions take on new patterns, even if the raw input feature distributions look similar. Feature attribution drift measures how the model behaves in production, which sets it apart from structural feature importance (how the model was built) and from concept drift (changes in real-world conditions).
Across both targets, the difference between drift and skew comes down to what is being compared:
Drift is always a temporal comparison using only production data. Skew is always a comparison across environments, training versus serving, which is why it is also called training-serving skew.
An Agent Platform endpoint can receive hundreds, thousands, or millions of prediction requests per day, and logging every request is not always necessary. The sampling rate defines the proportion of incoming requests that are logged and analyzed for monitoring. For example, setting a random sample configuration to 0.33 captures roughly one third of incoming requests (about two out of every six requests) for the monitoring dataset, while the remaining requests are still served normally without being logged for monitoring.
Lowering the sampling rate reduces compute and storage costs because fewer requests are stored and processed for analysis. On high-volume endpoints serving thousands or millions of daily predictions, those savings add up quickly. On smaller or low-volume endpoints, however, an aggressive reduction hurts monitoring accuracy. If an endpoint handles only a few hundred predictions per day and samples at 10%, the monitoring job may see only a few dozen data points, which is rarely enough to detect subtle distribution changes or separate real drift from random statistical noise. That weakness grows when traffic has seasonal patterns or when drift unfolds gradually, increasing the risk of missed signals or false alarms from normal variation.
Monitoring frequency controls when Agent Platform Model Monitoring runs its analysis on the collected samples. You configure the schedule using standard five-field cron notation, where an asterisk (*) represents every unit for that field:
For example, the schedule 0 9 * * 1 runs at minute 0 of hour 9 (9 a.m.), across every day of the month and every month, restricted to day 1 of the week, so the monitoring job executes at 9 a.m. every Monday.
Sampling rate and monitoring frequency work together to govern three outcomes: detection speed, detection reliability, and cost. Higher frequency catches shifts sooner, and higher sampling improves both reliability and the chance of catching changes by covering more traffic. At the same time, higher sampling increases storage and processing costs, while higher frequency increases compute spent running analysis jobs.
In Professional Machine Learning Engineer scenarios, we would match these two settings to traffic volume, rate of change, and risk tolerance:
Skew detection carries a strict prerequisite: you must supply a valid training data baseline. Because training-serving skew compares production inputs or attributions against training, the monitoring system cannot measure deviation without the original training distribution. The baseline must be the exact training dataset used to train the deployed model rather than a different dataset from the same domain.
Configuring skew detection follows a four-step sequence:
Drift detection does not require uploading baseline training data. Because drift compares production data against itself over time, the monitoring service builds its baseline directly from production traffic.
Our Professional Machine Learning Engineer course covers Agent Platform Model Monitoring alongside endpoint deployment and model retraining workflows, with practice questions that drill these distinctions.