Vertex AI Model Monitoring is one of those services that the Professional Cloud Architect exam likes to test in a very specific way. The questions are not really about how to configure the service. They are about whether you can tell drift from skew, and whether you understand the difference between feature monitoring and feature attribution monitoring. If you walk into the exam with those four ideas crisp in your head, the model monitoring questions become almost mechanical. If you do not, they look like a wall of similar phrases.
I want to walk through how I think about Vertex AI Model Monitoring for the Professional Cloud Architect exam, starting with what the service does, then breaking down the two detection types, and then closing with the drift versus skew distinction that almost always shows up in the exam questions.
Even with retraining, model degradation is inevitable. The world changes, and a model that was accurate last quarter will eventually make worse predictions on this quarter's data. Vertex AI Model Monitoring exists to track the quality of predictions over time on real production data, so the team running the model can catch degradation as it happens rather than discovering it from a dashboard six weeks later.
The service is most often associated with online prediction on Vertex AI Endpoints, and that is also where you will see it in exam questions. It can also be used on batch prediction, but the exam framing is almost always about a model serving live traffic from an Endpoint. One small piece of trivia worth knowing is that Model Monitoring used to be called Continuous Evaluation back when Vertex AI was still called AI Platform. The capability is the same.
Vertex AI Model Monitoring offers two detection types, and the exam expects you to know both. The first is feature monitoring, which looks at the input features going into the model. The second is feature attribution monitoring, which looks at how much each feature is actually contributing to the model's predictions. Both can be configured for either drift or skew, which is why this corner of the exam can feel like a four-cell grid that has to be reasoned through quickly.
Feature monitoring is the more intuitive of the two. It watches the statistical distribution of the model's inputs. If the model was trained on a feature that was roughly normally distributed around a value of 50, and suddenly the production data is showing values clustered around 80, that is a signal that something has changed in the inputs. The model itself may still be working as designed, but it is now being asked to make predictions on data that does not look like its training data.
Feature attribution monitoring is more subtle. Instead of looking at the inputs, it looks at the relative contribution of each feature to the prediction at the level of the individual prediction, and then aggregates those contributions. The mental model I use is that this is measuring how the model is actually behaving, not what it is seeing. Two features can have stable distributions, but if the model is now leaning much harder on one of them than it used to, that is a meaningful signal that the model's decision-making has shifted.
It is worth being careful here. Feature attribution drift is not the same as feature importance, which is a structural property of the trained model and does not change at inference time. It is also not the same as concept drift, which is a change in the real-world relationship between inputs and outputs. Feature attribution drift sits in between. It is asking how much each feature is contributing to predictions in production, and whether that has changed.
This is the part that trips people up, because the words sound interchangeable but they mean something specific in Vertex AI Model Monitoring. Drift is always about comparing production to itself over time. Skew is always about comparing production to training.
Feature drift is when the statistical distribution of a feature in production changes compared to itself over time. The classic example is a fraud detection model where desktop traffic was 80 percent of inputs at launch and mobile was 20 percent, and a few years later that has flipped to 40 percent desktop and 60 percent mobile. The fraud rate has stayed at around 2.1 percent. The average transaction values per channel have stayed stable. But the input mix has shifted enough that the model is no longer seeing the same kind of data it used to see in production. This is feature drift, and Vertex AI Model Monitoring detects it by comparing recent production windows against earlier production windows.
Feature skew is when a feature's distribution in production is different from its distribution in the training data. This is a comparison across environments rather than across time. The model was trained on a particular slice of data, and the data flowing into the Endpoint does not match. Skew is also called training-serving skew, where serving just means production. If a question on the exam mentions training-serving skew, it is asking about feature skew.
Feature attribution drift is the same time-based comparison, but applied to attribution values rather than feature distributions. The predictive contributions of features in production change compared to themselves over time. Login distance might have been the dominant factor in fraud predictions at launch, with a SHAP value around 0.3, while transaction speed sat lower at 0.15. Six months later, transaction speed could be the dominant factor at 0.4 and login distance might have dropped to 0.25. The features themselves might still have similar distributions, but the model is now relying on them differently. That is attribution drift.
Feature attribution skew is the cross-environment version of the same idea. The predictive contributions of features in production are different from their contributions during training. The model was supposedly built to lean on certain features in particular ways, and in production it is leaning on them differently from day one.
The fastest way I have found to keep the four detection variants straight is a two-by-two. One axis is what is being monitored: the feature distributions or the feature attributions. The other axis is the comparison: production against itself over time, or production against training. Drift is always production against itself. Skew is always production against training.
When a Professional Cloud Architect exam question describes a model that has been live for months and is now starting to behave differently from how it behaved at launch, the right answer involves drift, because the comparison is across time within production. When a question describes a model that has been deployed and the team is worried that the production data does not match the data the model was trained on, the right answer involves skew. When the question is about input distributions, it is feature drift or feature skew. When the question is about how much each feature is contributing to the prediction, it is feature attribution drift or feature attribution skew.
Once that grid is locked in, the rest of the model monitoring material on the Professional Cloud Architect exam becomes much easier to navigate, because the service itself is fairly thin on its own. The hard part is the vocabulary, and Vertex AI Model Monitoring is one of the places where Google is precise about words.
If you want a deeper walk through Vertex AI Model Monitoring alongside the rest of the ML and AI material that shows up on this certification, my full course is at https://gcpstudyhub.com/courses/professional-cloud-architect.
