Vertex AI Study Notes

Vertex AI is Google Cloud's machine learning platform for building, training, deploying, and managing ML models, including access to foundation models through Model Garden. It brings together datasets, training jobs, pipelines, a feature store, and prediction endpoints in one service. For the ACE exam it matters mainly as a managed service to provision and integrate correctly.

Key facts

• Vertex AI supports both AutoML, which trains models from your data with little code, and custom training, where you supply your own training code in a container or a supported framework environment.
• A trained model is deployed to an endpoint for online prediction, giving applications a stable API for low-latency inference, while batch prediction runs large scoring jobs asynchronously.
• Vertex AI Pipelines, based on Kubeflow Pipelines and TFX concepts, orchestrate multi-step ML workflows such as preprocessing, training, evaluation, and deployment in a repeatable, tracked way.
• The Feature Store provides a managed repository for ML features so training and serving use the same feature definitions and values, reducing training-serving skew.
• Model Garden gives access to Google and third-party foundation models, including Gemini models, which applications can call through the Vertex AI API without hosting the model yourself.
• Datasets, models, endpoints, and jobs are regional resources stored with data in the region you choose, and access to them is governed by IAM like other Google Cloud resources.
• Vertex AI integrates with BigQuery, Cloud Storage, and Dataflow for data preparation, and experiment tracking and metadata help teams compare runs and reproduce results.
• Workbench provides managed Jupyter notebook environments attached to Vertex AI for exploration and development, with access to accelerators such as GPUs where configured.

What the exam expects you to know

• Choosing Vertex AI as the managed platform answer when a scenario wants to train and deploy a model without managing ML infrastructure.
• Distinguishing online prediction through an endpoint from batch prediction for large, latency-tolerant scoring jobs.
• Knowing AutoML suits teams without deep ML coding experience, while custom training suits teams bringing existing TensorFlow, PyTorch, or scikit-learn code.
• Placing training data in Cloud Storage or BigQuery and granting the Vertex AI service account access, as the usual data-path setup in scenarios.
• Recognizing Model Garden and the Vertex AI API as the managed route to call Gemini and other foundation models from an application.
• Applying IAM and VPC Service Controls concepts to ML resources the same way as to any other Google Cloud service in a security question.

Vertex AI vs BigQuery ML: when to use which

Choose BigQuery ML when the data already lives in BigQuery and the team wants to create and run standard models entirely with SQL, without a separate ML platform. Choose Vertex AI when you need custom model code, AutoML for images, text, or tabular data, managed endpoints, pipelines, a feature store, or foundation models from Model Garden. The criteria are skill set and scope: BigQuery ML optimizes for SQL-first simplicity on warehouse data, while Vertex AI is the full platform for the complete ML lifecycle. See also the BigQuery ML notes.

Useful gcloud commands

gcloud ai models list --region=us-central1
gcloud ai endpoints list --region=us-central1
gcloud ai custom-jobs list --region=us-central1

Common exam traps

• Assuming deploying a model is the same as registering it: a model must be deployed to an endpoint before it can serve online predictions, and undeploying stops serving but keeps the model.
• Expecting endpoint traffic to be free of cold scaling behavior: endpoints allocate machine resources you choose and pay for while deployed, unlike scale-to-zero serverless services.
• Storing training data only on a local notebook disk: managed training reads from Cloud Storage or BigQuery, so data must be placed where the service account can reach it.
• Confusing Vertex AI with a general data warehouse: analytics queries belong in BigQuery; Vertex AI consumes prepared data for training and prediction.

Frequently asked questions

What is the difference between AutoML and custom training?

AutoML automates model selection and tuning: you provide labeled data and Vertex AI builds a model with minimal code, which suits tabular, image, text, and video use cases for teams without ML specialists. Custom training runs your own code and framework, giving full control over architecture and preprocessing at the cost of more engineering work.

How does an application get predictions from a Vertex AI model?

For real-time needs you deploy the model to an endpoint and call that endpoint's API with input instances, receiving predictions in the response. For large volumes with relaxed timing you run a batch prediction job that reads inputs from Cloud Storage or BigQuery and writes results back to storage or a table.

What is Vertex AI Feature Store used for?

Feature Store is a central, managed repository where teams define and store ML features once and reuse them for both training and online serving. Serving features from the same source used in training prevents mismatches between the two, and point-in-time lookups help build correct training datasets without leaking future data.

Study aid disclaimer: Cloud Cert Notes is an independent educational study aid. It is not affiliated with, endorsed by, or sponsored by Google, Microsoft, Amazon, Oracle, or Alibaba. Google Cloud, Associate Cloud Engineer, and Cloud Digital Leader are trademarks of Google LLC; all other certification names and product names are trademarks of their respective owners. Always confirm service details against the official provider documentation before sitting an exam.