Vertex AI Workbench is a managed notebook environment that lets you build, train, and iterate machine learning models without worrying about server setup or maintenance. For the Google Professional Machine Learning Engineer exam, you need to understand how this tool serves as your central command station for model development and how it integrates with Vertex AI Training to scale up your work. Mastery of this concept is critical because the exam tests your ability to choose the right environment and workflow for different stages of model development.
Jump to a section
A simple way to picture Developing and Training Models with Vertex AI Workbench
When you decide to cook a special meal, you first gather your recipes and ingredients, which leads you to set up your kitchen with the right tools and workspace. This is like preparing to build an ML model using Vertex AI Workbench.
Imagine you are a personal chef who has been asked to create a new signature dish. You don't just start cooking randomly. First, you go to your well-organised kitchen (the Workbench environment) where all your knives, pans, and utensils are already in place. You pull out your recipe notebook (Jupyter Notebook) and start writing down the steps: chop the vegetables (prepare your data), mix the spices (engineer features), and decide the cooking time (choose your model hyperparameters).
Now, cooking a complex dish might take hours on the stove. But you have a professional kitchen assistant (Vertex AI Training) who can take your recipe and cook it in multiple ovens at once (distributed training). You start the process on your own stove for a small test batch (local training on the Workbench), and when it works, you hand the recipe to your assistant to scale it up to cook for a hundred guests (training at scale in the cloud).
The key difference is that, unlike a home cook who might have only one oven, the professional kitchen lets you experiment with different cooking methods simultaneously (experiment tracking) and automatically cleans up any spilled ingredients (managed infrastructure). Your recipe notebook remains the single source of truth, and you can tweak it any time without starting from scratch. That is precisely how Vertex AI Workbench gives you a familiar notebook environment while handing the heavy lifting of training to powerful cloud resources.
Vertex AI Workbench is a fully managed, cloud-based service that provides you with a ready-to-use Jupyter Notebook environment. A Jupyter Notebook is an interactive document that lets you write and run code, see the results immediately, and add explanatory text all in one place. Think of it as a digital lab notebook where every experiment is recorded, reproducible, and shareable.
Before Vertex AI Workbench existed, data scientists had to set up their own environments manually. They would install Python, Jupyter, and all the necessary libraries on their local machines. This caused many problems: different team members had different versions, it was hard to share work, and training large models would crash a laptop because of limited memory. Vertex AI Workbench solves this by providing a pre-configured, consistent environment that lives entirely in Google Cloud. You access it through your browser, so your local computer's power does not matter.
One of the most important features of Vertex AI Workbench is its integration with Vertex AI Training. When you have developed a model in your notebook and you need to train it on a larger dataset or with more computational power, you do not have to copy code around. Instead, you can package your training code into a Python script and submit it as a training job directly from the Workbench. This job runs on Vertex AI Training, which automatically spins up virtual machines (VMs) with the required CPUs, GPUs, or TPUs to perform the training. You only pay for the time the machines are running, and once training finishes, the machines are shut down automatically.
Vertex AI Workbench also includes built-in version control through integration with Cloud Source Repositories and git. This means every change you make to your notebook can be tracked, reviewed, and rolled back if needed. For the exam, you should remember that Workbench notebooks can be customised by choosing different machine types (like n1-standard-4 or n1-highmem-8) to match the computational needs of your work.
Another key concept is the kernels offered in Workbench. A kernel is the computational engine that executes the code in your notebook. Vertex AI Workbench supports different kernels, including Python 3, R, and TensorFlow. You can also install custom packages to suit your project needs. The environment is persistent: if you close the notebook and come back later, all your work and installed packages are still there, unless you choose to reset the environment.
When developing models, you typically follow this workflow inside Workbench:
Explore and clean your data using libraries like pandas and numpy.
Visualise data to find patterns using matplotlib or seaborn.
Engineer features by creating new columns or transforming existing ones.
Split your data into training, validation, and test sets.
Select a machine learning algorithm and train it locally in the notebook to test the pipeline.
Evaluate the model's performance using metrics like accuracy or mean squared error.
Once you are satisfied, package the code and submit a Vertex AI Training job for full-scale training.
Vertex AI Workbench also supports collaboration. You can share notebooks with other team members, and each person runs their own copy of the environment. This is helpful when multiple data scientists are working on different aspects of the same project.
Why does this matter for the PMLE exam? The exam will ask you to decide when to use Vertex AI Workbench versus other tools like Colab or local Jupyter. The correct answer is usually that Workbench is best when you need a persistent, secure, and scalable environment that integrates with other GCP services. It is also the recommended choice for enterprise teams because of its managed infrastructure, built-in security, and audit logging through Cloud Logging.
Create a Vertex AI Workbench instance
Open the Google Cloud Console and navigate to Vertex AI Workbench. Create a new notebook instance, choosing a name, region, and machine type. This sets up a persistent VM with Jupyter pre-installed. This step is essential because it provides you with a consistent, ready-to-use development environment that you can access from any browser.
Load and explore your data
Inside the notebook, write Python code to load data from BigQuery or Cloud Storage. Use libraries like pandas to clean and explore the data. This step is crucial because understanding your data forms the foundation of any machine learning project, and Workbench's integration makes data access straightforward.
Engineer features and split data
Create new features that might improve model performance, such as date-based features or aggregations. Split the dataset into training, validation, and test sets. Feature engineering directly impacts model accuracy, and proper splitting prevents data leakage.
Train a baseline model locally
Use the notebook to train a simple model, such as linear regression or logistic regression, on the small dataset. Evaluate its performance to establish a baseline. This step validates that your data pipeline works and gives you a benchmark before scaling up.
Package code and submit a Vertex AI Training job
Refactor your training code into a Python script and place it in a Cloud Storage bucket. Use the Vertex AI SDK to define a training job with custom machine requirements (e.g., GPUs). Submit the job and monitor progress. This step demonstrates how to move from development to scalable training, which is a key exam concept.
Compare experiments and save the final model
After training completes, retrieve the metrics from Cloud Logging or Vertex AI Experiments. If satisfied, register the model in Vertex AI Model Registry. This step ensures you can track iterations and deploy the best version later.
Consider the scenario at a retail company called ShopEase. The company wants to build a model that predicts customer churn — whether a customer is likely to stop buying from them within the next month. A senior machine learning engineer, Priya, is tasked with developing this model.
Priya starts by opening Vertex AI Workbench from the Google Cloud Console. She creates a new notebook instance and chooses a standard environment with a Python 3 kernel. The first thing she does is load the customer data from BigQuery, Google's data warehouse. She writes Python code in the notebook to query the database and pull the last six months of customer transaction data into a pandas DataFrame.
Next, she explores the data: checking for missing values, visualising the distribution of purchase amounts, and examining how often customers visit the website. She creates new features like 'days since last purchase' and 'average order value'. She splits the data into a training set (80%) and a test set (20%) to ensure she can evaluate her model honestly.
Priya then trains a simple logistic regression model directly in the notebook. The training finishes in about 30 seconds because the dataset is small. She evaluates the model using a confusion matrix and gets a baseline accuracy of 75%. She knows she can do better. She decides to try a gradient boosting model, which is more complex and computationally intensive.
Because her Workbench instance has only 4 vCPUs and 16 GB of RAM, training a gradient boosting model on the full dataset would take too long. Instead, she packages her training code into a Python script and defines a training job in Vertex AI Training. She specifies that the job should use a machine with 8 vCPUs and 64 GB of RAM. She starts the job from the Workbench interface and monitors its progress through the Cloud Console.
While the job runs, Priya can close her laptop and go to a meeting. When she returns, the training is complete. She pulls the evaluation metrics from Cloud Logging and sees that the gradient boosting model has an accuracy of 85%. She then saves the trained model to Vertex AI Model Registry, making it ready for deployment.
The next steps involve iterating: Priya tries different hyperparameters, adds more features, and re-trains. Each time, she tracks her experiments using Vertex AI Experiments, which is also accessible from the Workbench. She can compare the performance of different runs side by side. At the end of the week, she has a final model that predicts churn with 90% accuracy.
What did Priya actually do? She used Vertex AI Workbench as her development cockpit, performed data exploration and initial lightweight training inside the notebook, and delegated heavy training to Vertex AI Training. She never installed any software on her local machine, never worried about running out of memory, and never had to configure servers. The entire workflow was managed by Google Cloud.
For the PMLE exam, you need to recognise that this scenario — using Workbench for development and then scaling training — is the recommended best practice. Questions might present a similar scenario and ask you which service to use for a given step in the pipeline.
The PMLE exam tests your understanding of Vertex AI Workbench and training services in several specific ways. You need to know not just the features, but also when to use each tool and how they fit together.
Here are the exact exam topics you must master for objective 2.2:
When to use Vertex AI Workbench vs. Colab vs. local Jupyter. The correct answer pattern is: Workbench is for enterprise environments requiring persistent, reproducible, and secure notebooks with integration to other GCP services. Colab is for quick prototyping with free GPUs. Local Jupyter is for offline development with no cloud dependency.
How to submit a training job from a notebook. The exam loves to test the sequence: you first write and test the training code in the notebook, then package it into a Python module, and then use the Vertex AI SDK to submit the job.
The difference between custom training (where you provide your own code) and AutoML training (where Google's systems train the model automatically). Workbench is typically used for custom training where you want full control over the code.
Understanding machine types and accelerators. They may ask which machine type to choose for a given workload, for example, using a machine with GPUs for deep learning and high-memory machines for large datasets.
Experiment tracking with Vertex AI Experiments. They will test that you can log parameters and metrics in your notebook and compare runs.
Common traps in the exam include:
Believing that Vertex AI Workbench is only for Python. It supports R and other kernels, so do not assume Python is the only option.
Thinking that all training must happen on the Workbench instance itself. The correct approach is to use Workbench for development and scale training with Vertex AI Training.
Confusing Vertex AI Workbench with Vertex AI Notebooks (older name). The current name is Vertex AI Workbench, and the exam uses this terminology.
Ignoring the cost implications. They might ask which option is most cost-effective for a one-off training job — the answer is often Vertex AI Training because you only pay for the compute time used, whereas a Workbench instance runs continuously.
Key definitions to memorise:
Managed notebook: a notebook environment that Google maintains for you, applying security patches and updates automatically.
Custom training: training a model using your own code and algorithms, as opposed to AutoML.
Hyperparameter tuning: the process of searching for the best set of hyperparameters (like learning rate) for your model. Vertex AI supports this as a service where you define a search space and it runs multiple training jobs automatically.
Distributed training: splitting the training workload across multiple machines to make it faster. Vertex AI Training supports this for large models.
The exam often includes scenario-based questions where you are given a description of a data scientist's tasks, and you must decide the correct sequence of steps using Vertex AI Workbench and Training. The correct answer pattern almost always involves using Workbench for initial exploration and lightweight training, then submitting a job to Vertex AI Training for the heavy lifting.
Vertex AI Workbench is a managed Jupyter notebook environment that integrates seamlessly with other GCP services for end-to-end machine learning development.
Use Vertex AI Workbench for model development and lightweight training, but delegate large-scale training to Vertex AI Training for better cost and performance.
Workbench environments are persistent and customisable; you can choose machine types and kernels to match your workload requirements.
Experiment tracking using Vertex AI Experiments should be used inside your notebook to log and compare different model runs and hyperparameters.
Version control your notebooks using Cloud Source Repositories or git to ensure reproducibility and collaboration within a team.
The exam will test your ability to choose between Workbench, Colab, and local Jupyter based on security, persistence, and integration needs.
Custom training in Vertex AI Training requires packaging your code as a Python module and using the Vertex AI SDK to submit the job from within Workbench.
Distributed training and hyperparameter tuning are supported by Vertex AI Training, and you can invoke them from your Workbench notebook.
These come up on the exam all the time. Here's how to tell them apart.
Vertex AI Workbench
Persistent environment with custom machine types and kernels
Integrated with GCP services like BigQuery and Cloud Storage
Enterprise-grade security, audit logging, and IAM roles
Google Colab
Free, with limited compute resources
No persistent storage and resets on disconnect
Basic sharing via links, no enterprise security
Custom Training
You write your own training code and algorithms
Full control over model architecture and hyperparameters
Works with Vertex AI Workbench for development
AutoML Training
Google’s systems automatically train the best model
Limited control, but requires no ML expertise
Accessed through the Vertex AI dashboard, not notebooks
Vertex AI Training
Runs on scalable, managed infrastructure
Supports distributed training across multiple machines
Cost-efficient for large jobs; only pay for compute time
Vertex AI Workbench (local training)
Runs on the single VM of the notebook instance
Limited to the instance’s CPU/GPU resources
Ideal for prototyping and small datasets
Mistake
Vertex AI Workbench is just a hosted Jupyter notebook with no additional features.
Correct
Vertex AI Workbench includes deep integration with other GCP services like BigQuery, Cloud Storage, and Vertex AI Training, along with built-in version control, environment management, and experiment tracking. It is far more than a plain Jupyter instance.
Beginners often think of 'notebook' as a simple code editor, not realising the ecosystem of integrations that make Workbench a full development platform.
Mistake
You must train your model inside the Workbench notebook itself, otherwise you cannot use it.
Correct
Workbench is designed for development and experimentation, but for large-scale training you should submit a job to Vertex AI Training. The notebook is the control centre, not the production engine.
Because a notebook is the first place you write and test code, it is natural to assume all training happens there. The exam tests that you know when to offload to a separate training service.
Mistake
Vertex AI Workbench automatically saves your work without any extra steps, so you never lose data.
Correct
Workbench provides persistent storage, but you must manage your own backups and version control. If you accidentally delete the instance, you could lose your notebooks and data unless they are saved to Cloud Storage or a git repository.
The convenience of managed services can lead people to assume they do not need to follow best practices like regular saves and version control, which is a dangerous assumption in an exam context.
Mistake
Vertex AI Workbench is only for machine learning engineers and cannot be used for simple data analysis.
Correct
Workbench can be used for any data analysis task that benefits from a notebook interface, including data cleaning, statistical analysis, and visualisation, even if you never build a model.
The name 'Machine Learning' in the service misleadingly suggests it is only for ML, but it is a general-purpose analytical tool.
Reveal each answer, then mark whether you got it right. Score 60%+ to unlock the next chapter.
No. Colab is a free, lightweight notebook service designed for quick prototyping and education. Vertex AI Workbench is an enterprise-grade, fully managed notebook environment with persistent storage, team collaboration, and deep integration with GCP services.
Yes. Vertex AI Workbench supports multiple kernels, including Python, R, and TensorFlow. You can install additional packages as needed.
Yes, if you are not actively using it, you should stop the instance to avoid paying for idle compute resources. The environment is persistent, so your files remain when you stop and start the instance.
You package your training code into a Python module, upload it to Cloud Storage, and then use the Vertex AI SDK in your notebook to define and submit a training job. You can specify the machine type, number of replicas, and accelerator type.
Custom training uses your own code and algorithms, giving you full control. AutoML automatically searches for the best model architecture and hyperparameters based on your data. Vertex AI Workbench is typically used with custom training.
Yes. You can share the notebook instance with other users through IAM permissions. Each collaborator can run their own kernel and edit the notebook simultaneously, but changes are not merged automatically. It is recommended to use version control with git for true collaboration.
You've finished Developing and Training Models with Vertex AI Workbench. Continue through the PMLE study guide to build a complete picture of the exam.
Done with this chapter?