Professional-Machine-Learning-Engineer Google Professional Machine Learning Engineer Free Practice Exam Questions (2026 Updated)

According to the official exam guide 1 , one of the skills assessed in the exam is to “track the lineage of pipeline artifacts”.  Vertex AI Experiments 2  is a service that allows you to track and compare the results of your model training runs. Vertex AI Experiments automatically logs metadata such as hyperparameters, metrics, and artifacts for each training run. You can use Vertex AI Experiments to train your custom model using TensorFlow, PyTorch, XGBoost, or scikit-learn.  Vertex AI Model Registry 3  is a service that allows you to manage your trained models in a central location. You can use Vertex AI Model Registry to register your model, add labels and descriptions, and view the model’s lineage graph. The lineage graph shows the artifacts and executions that are part of the model’s creation, such as the dataset, the training pipeline, and the evaluation metrics. The other options are not relevant or optimal for this scenario.  References :

Professional ML Engineer Exam Guide

Vertex AI Experiments

Vertex AI Model Registry

Google Professional Machine Learning Certification Exam 2023

Latest Google Professional Machine Learning Engineer Actual Free Exam Questions

Option A is incorrect because it does not separate the training dataset into two tables based on the features, which is necessary to train the two models separately and accurately.

Option B is incorrect because it does not separate the training dataset into two tables based on the features, and because it uses the same monitoring-config-from parameter for both models, which would not account for the different feature selections.

Option C is incorrect because it deploys the models to two separate endpoints, which would increase the management effort and complexity of the pipeline.

Option D is correct because it separates the training dataset into two tables based on the features, which would enable the two models to be trained separately and accurately. It also deploys both models to the same endpoint, which would simplify the pipeline and reduce the management effort. It also submits a Vertex Al Model Monitoring job with a monitoring-config-from parameter that accounts for the model IDs and training datasets, which would enable the skew detection to work properly for each model.

According to the web search results, Vertex AI Pipelines is a serverles s orchestrator for running ML pipelines, using either the KFP SDK or TFX 1 .  Vertex AI Pipelines provides a set of prebuilt components that can be used to perform common ML tasks, such as t raining, evaluation, deployment, and more 2 .  Vertex AI ModelEvaluationOp and ModelDeployOp are two such components that can be used to evaluate and deploy a model to an endpoint for online inference 3 . However, Vertex AI Pipelines does not provide a prebuilt component for hyperparameter tuning.  Therefore, to have control ove r how the model parameters are tuned, you need to use a custom component that calls the Vertex AI HyperparameterTuningJob service 4 . Therefore, option A is the best way to decide which Google Cloud pipeline components to use for the given use case, as it includes a custom component for hyperparameter tuning, and prebuilt components for model evaluation and deployment. The other options are not relevant or optimal for this scenario.  References :

Vertex AI Pipelines

Google Cloud Pipeline Components

Vertex AI ModelEvaluationOp and ModelDeployOp

Vertex AI HyperparameterTuningJob

Google Professional Machine Learning Certification Exam 2023

Latest Google Professional Machine Learning Engineer Actual Free Exam Questions

 This answer is correct because it uses Firebase, a platform that provides a scalable and reliable notification system for mobile and web applications. Firebase Cloud Messaging (FCM) allows you to send messages and notifications to users across different devices and platforms. By registering each user with a user ID on the FCM server, you can target specific users based on their account balance predictions and send them personalized notifications when their balance is likely to drop below the $25 threshold. This way, you can provide a useful and timely feature for your customers and increase their engagement and retention.  References:

[Firebase Cloud Messaging]

[Firebase Cloud Messaging: Send messages to specific devices]

Vertex AI Feature Store is a service that allows you to store and manage your ML features on Google Cloud. You can use Vertex AI Feature Store to store features such as parcels delivered and truck locations over time, and retrieve them with low latency for online prediction. Online prediction is a type of prediction that provides low-latency responses to individual or small batches of input data. You can also use Vertex AI Feature Store to retrieve historical data at a specific point in time for model training. Model training is a process of learning the parameters of a ML model from data. By using Vertex AI Feature Store, you can store the features with minimal effort, and avoid the complexity of managing your own data storage and serving system.  References :

Vertex AI Feature Store documentation

Preparing for Google Cloud Certification: Machine Learning Engineer Professional Certificate

Vertex AI is a unified platform for building and managing machine learning solutions on Google Cloud. It provides various services and tools for different stages of the machine learning lifecycle, such as data preparation, model training, deployment, monitoring, and experimentation. Vertex AI Workbench is an integrated development environment (IDE) that allows you to create and run Jupyter notebooks on Google Cloud. You can use Vertex AI Workbench to develop your ML model in Python, using libraries such as TensorFlow, PyTorch, scikit-learn, etc. You can also use the Vertex SDK, which is a Python client library for Vertex AI, to track artifacts and compare models during experimentation. You can use the  aiplatform.init  function to initialize the Vertex SDK with the name of your experiment. You can use the  aiplatform.start_run  and  aiplatform.end_run  functions to create and close an experiment run. You can use the  aiplatform.log_params  and  aiplatform.log_metrics  functions to log the parameters and metrics for each experiment run. You can also use the  aiplatform.log_datasets  and  aiplatform.log_model  functions to attach the dataset and model artifacts as inputs and outputs to each experiment run. These functions allow you to record and store the metadata and artifacts of your experiments, and compare them using the Vertex AI Experiments UI. After a successful experiment, you can create a Vertex AI pipeline, which is a way to automate and orchestrate your ML workflows. You can use the  aiplatform.PipelineJob  class to create a pipeline job, and specify the components and dependencies of your pipeline. You can also use the  aiplatform.CustomContainerTrainingJob  class to create a custom container training job, and use the  run  method to run the job as a pipeline component. You can use the  aiplatform.Model.deploy  method to deploy your model as a pipeline component. You can also use the  aiplatform.Model.monitor  method to monitor your model as a pipeline component. By creating a Vertex AI pipeline, you can rapidly and easily transition successful experiments to production, and reuse and share your ML workflows. This solution requires minimal changes to your code, and leverages the Vertex AI services and tools to streamline your ML development process.  References : The answer can be verified from official Google Cloud documentation and resources related to Vertex AI, Vertex AI Workbench, Vertex SDK, and Vertex AI pipelines.

Vertex AI | Google Cloud

Vertex AI Workbench | Google Cloud

Vertex SDK for Python | Google Cloud

Vertex AI pipelines | Google Cloud

Cost-effectiveness: User-managed notebooks in Vertex AI Workbench allow you to leverage pre-configured virtual machines with reasonable resource allocation, keeping costs lower compared to options involving managed notebooks or Dataproc clusters.

Development flexibility: User-managed notebooks offer full control over the environment, allowing you to install additional libraries or dependencies needed for your specific EDA, preprocessing, and model training tasks. This flexibility is crucial while experimenting with different algorithms.

BigQuery integration: The %%bigquery magic commands provide seamless integration with BigQuery within the Jupyter Notebook environment. This enables efficient querying and exploration of customer transaction data stored in BigQuery directly from the notebook, streamlining the workflow.

Other options and why they are not the best fit:

B. Managed notebook: While managed notebooks offer an easier setup, they might have limited customization options, potentially hindering your ability to install specific libraries or tools.

C. Dataproc Hub: Dataproc Hub focuses on running large-scale distributed workloads, and it might be overkill for your scenario involving exploratory analysis and experimentation with different algorithms. Additionally, it could incur higher costs compared to a user-managed notebook.

D. Dataproc cluster with spark-bigquery-connector: Similar to option C, using a Dataproc cluster with the spark-bigquery-connector would be more complex and potentially more expensive than using %%bigquery magic commands within a user-managed notebook for accessing BigQuery data.

The best option for deploying a scikit-learn classification model to production is to deploy an online Vertex AI prediction endpoint and set the max replica count to 100. This option allows you to leverage the power and scalability of Google Cloud to serve requests 24/7 and handle millions of requests per second. Vertex AI is a unified platform for building and deploying machine learning solutions on Google Cloud. Vertex AI can deploy a trained scikit-learn model to an online prediction endpoint, which can provide low-latency predictions for individual instances. An online prediction endpoint consists of one or more replicas, which are copies of the model that run on virtual machines. The max replica count is a parameter that determines the maximum number of replicas that can be created for the endpoint. By setting the max replica count to 100, you can enable the endpoint to scale up to 100 replicas when the traffic increases, and scale down to zero replicas when the traffic decreases. This can help minimize the cost of deployment, as you only pay for the resources that you use.  Moreover, you can use the autoscaling algorithm option to optimize the scaling behavior of the endpoint based on the latency and utilization metrics 1 .

The other options are not as good as option B, for the following reasons:

Option A: Deploying an online Vertex AI prediction endpoint and setting the max replica count to 1 would not be able to serve requests 24/7 and handle millions of requests per second. Setting the max replica count to 1 would limit the endpoint to only one replica, which can cause performance issues and service disruptions when the traffic increases.  Moreover, se tting the max replica count to 1 would prevent the endpoint from scaling down to zero replicas when the traffic decreases, which can increase the cost of deployment, as you pay for the resources that you do not use 1 .

Option C: Deploying an online Vertex AI prediction endpoint with one GPU per replica and setting the max replica count to 1 would not be able to serve requests 24/7 and handle millions of requests per second, and would increase the cost of deployment. Adding a GPU to each replica would increase the computational power of the endpoint, but it would also in crease the cost of deployment, as GPUs are more expensive than CPUs.  Moreover, setting the max replica count to 1 would limit the endpoint to only one replica, which can cause performance issues and service disruptions whe n the traffic increases, and prevent the endpoint from scaling down to zero replicas when the traffic decreases 1 .  Furthermore, scikit-learn models do not benefit from GPUs, as scikit-learn is not optimized for GPU acceleration 2 .

Option D: Deploying an online Vertex AI prediction endpoint with one GPU per replica and setting the max replica count to 100 would be able to serve requests 24/7 and handle millions of requests per second, but it would increase the cost of deployment. Adding a GPU to each replica would increase the computational power of the endpoint, but it would also increase the cost of deployment, as GPUs are more expensive than CPUs. Setting the max replica count to 100 would enable the endpoint to scale up to 100 replicas when the traffic increases, and scale down to zero replicas when the traffic decreases, which can help minimize the cost of deployment.  However , scikit-learn models do not benefit from GPUs, as scikit-learn is not optimized for GPU acceleration 2 . Therefore, using GPUs for scikit-learn models would be unnecessary and wasteful.

Z-score normalization is a technique that transforms the values of a numeric variable into standardized units, such that the mean is zero and the standard deviation is one. Z-score normalization can help to compare variables with different scales and ranges, and to reduce the effect of outliers and skewness. The formula for z-score normalization is:

z = (x - mu) / sigma

where x is the original value, mu is the mean of the variable, and sigma is the standard deviation of the variable.

Dataflow is a service that allows you to create and run data processing pipelines on Google Cloud. You can use Dataflow to preprocess raw data prior to model training and prediction, such as applying z-score normalization on data stored in BigQuery. However, using Dataflow for this task may not be the most efficient option, as it involves reading and writing data from and to BigQuery, which can be time-consuming and costly. Moreover, using Dataflow requires manual intervention to update the pipeline whenever new training data is added.

A more efficient way to perform z-score normalization on data stored in BigQuery is to translate the normalization algorithm into SQL and use it with BigQuery. BigQuery is a service that allows you to analyze large-scale and complex data using SQL queries. You can use BigQuery to perform z-score normalization on your data using SQL functions such as AVG(), STDDEV_POP(), and OVER(). For example, the following SQL query can normalize the values of a column called temperature in a table called weather:

SELECT (temperature - AVG(temperature) OVER ()) / STDDEV_POP(temperature) OVER () AS normalized_temperature FROM weather;

By using SQL to perform z-score normalization on BigQuery, you can make the process more efficient by minimizing computation time and manual intervention. You can also leverage the scalability and performance of BigQuery to handle large and complex datasets. Therefore, translating the normalization algorithm into SQL for use with BigQuery is the best option for this use case.

A fairness metric is a way to measure how well a machine learning model treats different groups of customers, such as by sex or age. A common fairness metric is accuracy, which is the proportion of correct predictions among all predictions. Accuracy across the sensitive features means calculating the accuracy for each group separately, and then comparing them. For example, if the model has 90% accuracy for male customers and 80% accuracy for female customers, there is a 10% accuracy gap that indicates potential bias against female customers.

To investigate and mitigate potential bias, it is important to define a fairness metric and evaluate it on a test set. A test set is a subset of the data that is not used for training the model, but only for evaluating its performance. By joining the test set predictions with the sensitive features, you can calculate the fairness metric and see if it meets your requirements. For example, you may require that the accuracy gap between any two groups is less than 5%. If the fairness metric does not meet your requirements, you may need to adjust the model or the data to reduce bias.

Option A is not the best answer because excluding the sensitive features and any meaningfully correlated features may not eliminate bias. For example, if salary is correlated with sex, and salary is also a predictive feature for the target variable, excluding both features may reduce the model accuracy and still leave some residual bias. Moreover, excluding features based on correlation may not capture the complex interactions and dependencies among the features that may affect bias.

Option B is not the best answer because using the global attribution values for each feature of the model may not reflect the individual-level impact of the features on the predictions. Global attribution values are calculated by averaging the attribution values across all the data points, and they indicate how important each feature is for the overall model performance. However, they do not show how each feature affects each customer’s prediction, which may vary depending on the values of the other features. For example, sex may have a low global attribution value, but it may have a high impact on some customers’ predictions, especially if it interacts with other features such as salary or age.

Option C is not the best answer because discarding the model and training the model again without a feature based on a single customer’s attribution value may not be a robust or scalable way to mitigate bias. Attribution values are calculated by measuring how much each feature contributes to the prediction for a given data point, and they indicate how sensitive the prediction is to the feature value. However, they do not show how the feature affects the overall fairness metric or the model accuracy. For example, sex may have a high attribution value for a customer, but it may not affect the accuracy gap between the groups. Moreover, discarding and retraining the model based on a single customer’s attribution value may not be feasible if there are many customers with high attribution values for different features.

 The best option for monitoring the model to determine when retraining is necessary is to schedule a weekly query in BigQuery to compute the success metric. This option has the following advantages:

It allows the model performance to be evaluated regularly, based on the actual outcome of the recommendations. By computing the success metric, which is the percentage of articles that are opened within two days and read for at least one minute, you can measure how well the model is achieving its objective and compare it with the acceptable baseline.

It leverages the scalability and efficiency of BigQuery, which is a serverless, fully managed, and highly scalable data warehouse that can run complex queries over petabytes of data in seconds. By using BigQuery, you can access and analyze all the information needed to compute the success metric, such as the newsletter publication date, the article opening date, and the user reading time, without worrying about the infrastructure or the cost.

It simplifies the model monitoring and retraining workflow, as the weekly query can be scheduled and executed automatically using BigQuery’s built-in scheduling feature. You can also set up alerts or notifications to inform you when the success metric falls below the acceptable baseline, and trigger the model retraining process accordingly.

The other options are less optimal for the following reasons:

Option A: Using Vertex AI Model Monitoring to detect skew of the input features with a sample rate of 100% and a monitoring frequency of two days introduces additional complexity and overhead. This option requires setting up and managing a Vertex AI Model Monitoring service, which is a managed service that provides various tools and features for machine learning, such as training, tuning, serving, and monitoring. However, using Vertex AI Model Monitoring to detect skew of the input features may not reflect the actual performance of the model, as skew is the discrepancy between the distributions of the features in the training dataset and the serving data, which may not affect the outcome of the recommendations. Moreover, using a sample rate of 100% and a monitoring frequency of two days may incur unnecessary cost and latency, as it requires analyzing all the input features every two days, which may not be needed for the model monitoring.

Option B: Scheduling a cron job in Cloud Tasks to retrain the model every week before the newsletter is created introduces additional cost and risk. This option requires creating and running a cron job in Cloud Tasks, which is a fully managed service that allows you to schedule and execute tasks that are invoked by HTTP requests. However, using Cloud Tasks to retrain the model every week may not be optimal, as it may retrain the model more often than necessary, wasting compute resources and cost. Moreover, using Cloud Tasks to retrain the model before the newsletter is created may introduce risk, as it may deploy a new model version that has not been tested or validated, potentially affecting the quality of the recommendations.

Option D: Scheduling a daily Dataflow job in Cloud Composer to compute the success metric introduces additional complexity and cost. This option requires creating and running a Dataflow job in Cloud Composer, which is a fully managed service that runs Apache Airflow pipelines for workflow orchestration. Dataflow is a fully managed service that runs Apache Beam pipelines for data processing and transformation. However, using Dataflow and Cloud Composer to compute the success metric may not be necessary, as it may add more steps and overhead to the model monitoring process. Moreover, using Dataflow and Cloud Composer to compute the success metric daily may not be optimal, as it may compute the success metric more often than needed, consuming more compute resources and cost.

Option A is incorrect because migrating your model to TensorFlow, and training it using Vertex AI Training, is not the easiest way to improve the model’s training time. TensorFlow is a framework that allows you to create and train ML models using Python or other languages. Vertex AI Training is a service that allows you to train and optimize ML models using built-in algorithms or custom containers. However, this option requires significant code changes, as TensorFlow and scikit-learn have different APIs and functionalities. Moreover, this option does not leverage the parallelism or the scalability of the cloud, as it only uses a single instance.

Option B is incorrect because training your model in a distributed mode using multiple Compute Engine VMs, is not the most convenient way to improve the model’s training time. Compute Engine is a service that allows you to create and manage virtual machines that run on Google Cloud. You can use Compute Engine to run your scikit-learn model in a distributed mode, by using libraries such as Dask or Joblib. However, this option requires more effort and resources than option D, as it involves creating and configuring the VMs, installing and maintaining the libraries, and writing and running the distributed code.

Option C is incorrect because training your model with DLVM images on Vertex AI, and ensuring that your code utilizes NumPy and SciPy internal methods whenever possible, is not the most effective way to improve the model’s training time.  DLVM (Deep Learning Virtual Machine) images are preconfigured VM images that include popular ML frameworks and tools, such as TensorFlow, PyTorch, or scikit-learn 1 . You can use DLVM images on Vertex AI to train your scikit-learn model, by using a custom container. NumPy and SciPy are libraries that provide numerical and scientific computing functionalities for Python.  You can use NumPy and SciPy internal methods to optimize your scikit-learn code, as they are faster and more efficient than pure Python code 2 . However, this option does not leverage the parallelism or the scalability of the cloud, as it only uses a single instance.  Moreover, thi s option may not have a significant impact on the training time, as scikit-learn already relies on NumPy and SciPy for most of its operations 3 .

Option D is correct because training your model using Vertex AI Training with GPUs, is the best way to improve the model’s training time.  A GPU (Graphics Processing Unit) is a hardware acceler ator that can perform parallel computations faster than a CPU (Central Processing Unit) 4 . Vertex AI Training is a service that allows you to train and optimize ML models using built-in algorithms or custom containers.  You can use Vertex AI Training with GPUs to train your scikit-learn model, by using a custom container and specify ing the accelerator type and count 5 . By using Vertex AI Training with GPUs, you can leverage the parallelism and the scalability of the cloud, and speed up the training process significantly, without changing your code.

 Vertex AI Experiments is a service that allows you to track, compare, and optimize your ML experiments on Google Cloud. You can use Vertex AI Experiments to log metrics and parameters from your TensorFlow models, and then visualize them in Vertex AI TensorBoard. Vertex AI TensorBoard is a managed service that provides a web interface for viewing and debugging your ML experiments. You can use Vertex AI TensorBoard to compare different runs, inspect model graphs, analyze scalars, histograms, images, and more. By using Vertex AI Experiments and Vertex AI TensorBoard, you can simplify your ML experiment tracking and visualization workflow, and avoid the overhead of setting up and maintaining your own Cloud Functions, Cloud Storage buckets, or VMs.  References :

[Vertex AI Experiments documentation]

[Vertex AI TensorBoard documentation]

Preparing for Google Cloud Certification: Machine Learning Engineer Professional Certificate

 The best option for splitting the data between the training, validation, and test sets, using a managed tabular dataset in Vertex AI that contains sales data from three different stores, is to use Vertex AI default data split. This option allows you to leverage the power and simplicity of Vertex AI to automatically and randomly split your data into the three sets by percentage. Vertex AI is a unified platform for building and deploying machine learning solutions on Google Cloud. Vertex AI can support various types of models, such as linear regression, logistic regression, k-means clustering, matrix factorization, and deep neural networks. Vertex AI can also provide various tools and services for data analysis, model development, model deployment, model monitoring, and model governance. A default data split is a data split method that is provided by Vertex AI, and does not require any user input or configuration. A default data split can help you split your data into the training, validation, and test sets by using a random sampling method, and assign a fixed percentage of the data to each set. A default data split can help you simplify the data split process, and works well in most cases. A training set is a subset of the data that is used to train the model, and adjust the model parameters. A training set can help you learn the relationship between the input features and the target variable, and optimize the model performance. A validation set is a subset of the data that is used to validate the model, and tune the model hyperparameters. A validation set can help you evaluate the model performance on unseen data, and avoid overfitting or underfitting. A test set is a subset of the data that is used to test the model, and provide the final evaluation metrics. A test set can help you assess the model performance on new data, and measure the generalization ability of the model.  By using Vertex AI default data split, you can split your data into the training, validation, and test sets by using a random sampling method, and assign the following percentages of the data to each set 1 :

The other options are not as good as option B, for the following reasons:

Option A: Using Vertex AI manual split, using the store name feature to assign one store for each set would not allow you to split your data into representative and balanced sets, and could cause errors or poor performance. A manual split is a data split method that allows you to control how your data is split into sets, by using the ml_use label or the data filter expression. A manual split can help you customize the data split logic, and handle complex or non-standard data formats. A store name feature is a feature that indicates the name of the store where the sales data was collected. A store name feature can help you identify the source of the data, and group the data by store. However, using Vertex AI manual split, using the store name feature to assign one store for each set would not allow you to split your data into representative and balanced sets, and could cause errors or poor performance. You would need to write code, create and configure the ml_use label or the data filter expression, and assign one store for each set.  Moreover, this option would not ensure that the data in each set has the same distribution and characteristics as the data in the whole dataset, which could prevent you from learning the general pattern o f the data, and cause bias or variance in the model 2 .

Option C: Using Vertex AI chronological split and specifying the sales timestamp feature as the time variable would not allow you to split your data into representative and balanced sets, and could cause errors or poor performance. A chronological split is a data split method that allows you to split your data into sets based on the order of the data. A chronological split can help you preserve the temporal dependency and sequence of the data, and avoid data leakage. A sales timestamp feature is a feature that indicates the date and time when the sales data was collected. A sales timestamp feature can help you track the changes and trends of the data over time, and capture the seasonality and cyclicality of the data. However, using Vertex AI chronological split and specifying the sales timestamp feature as the time variable would not allow you to split your data into representative and balanced sets, and could cause errors or poor performance. You would need to write code, create and configure the time variable, and split the data by the order of the time variable.  Moreover, this option would not ensure that the data in each set has the same distribution and characteristics as the data in the whole dataset, which could prevent you from learning the general pattern of the data, and cause bias or variance in the model 3 .

Option D: Using Vertex AI random split, assigning 70% of the rows to the training set, 10% to the validation set, and 20% to the test set would not allow you to use the default data split method that is provided by Vertex AI, and could increase the complexity and cost of the data split process. A random split is a data split method that allows you to split your data into sets by using a random sampling method, and assign a custom percentage of the data to each set. A random split can help you split your data into representative and balanced sets, and avoid data leakage. However, using Vertex AI random split, assigning 70% of the rows to the training set, 10% to the validation set, and 20% to the test set would not allow you to use the default data split method that is provided by Vertex AI, and could increase the complexity and cost of the data split process. You would need to write code, create and configure the random split method, and assign the custom percentages to each set.  Moreover, this option would not use the default data split method that is provided by Vertex AI, which can simplify the data split process, and works well in most cases 1 .

Option A is incorrect because using AI Platform to run distributed training jobs with checkpoints does not reduce the compute costs, but rather increases them by using more resources and storing the checkpoints.

Option B is incorrect because using AI Platform to run distributed training jobs without checkpoints may reduce the compute costs, but it also risks losing the progress of the training if the job fails or is interrupted.

Option C is correct because migrating to training with Kubeflow on Google Kubernetes Engine, and using preemptible VMs with checkpoints can reduce the compute costs significantly by using cheaper and more scalable resources, while also preserving the state of the training with checkpoints.

Option D is incorrect because using preemptible VMs without checkpoints may reduce the compute costs, but it also risks losing the training progress if the VMs are preempted.

Auto scaling is a feature that allows you to automatically adjust the number of prediction nodes based on the traffic and load of your deployed model 1 .  However, auto scaling depends on the CPU utilization of your prediction nodes, which is the percentage of CPU resources used by your model server 1 .  If your CPU utilization is low, even during periods of high load, it means that your model server is not fully utilizing the available CPU resources, and thus au to scaling will not trigger more replicas 2 .

On e possible reason for low CPU utilization is that your model server is using a single worker process to handle prediction requests 3 .  A worker process is a subprocess that runs your model code and handles prediction requests 3 .  If you have only one worker process, it can only handle one request at a time, which can lead to dropped requests when the traffic is high 3 .  To increase the CPU utilization and the throughput of your model server, you can increase the number of worker processes, which will allow your model server to handle multiple requests in parallel 3 .

To increase the number of workers i n your model server, you need to modify your custom model server code and use the  --workers  flag to specify the number of worker processes you want to use 3 . For example, if you are using a Gunicorn server, you can use the following command to start your model server with four worker processes:

gunicorn --bind :$PORT --workers 4 --threads 1 --timeout 60 main:app

By increasing the number of workers in your model server, you can increase the CPU utilization of your prediction nodes, and thus enable auto scaling to scale beyond one replica.

The other options are not suitable for your scenario, because they either do not address the root cause of low CPU utilization, such as attaching a GPU or scheduling scaling, or they do not enable auto scaling, such as increasing the minReplicaCount, which is a fixed number of nodes that will always run regardless of the traffi c 1 .

 The Cloud DLP API is a service that allows users to inspect, classify, and de-identify sensitive data. It can be used to scan data in Cloud Storage, BigQuery, Cloud Datastore, and Cloud Pub/Sub. The best way to ensure that the PII is not accessible by unauthorized individuals is to use a quarantine bucket to store the data before scanning it with the DLP API. This way, the data is isolated from other applications and users until it is classified and moved to the appropriate bucket. The other options are not as secure or efficient, as they either expose the data to BigQuery before scanning, or scan the data after writing it to a non-sensitive bucket.  References:

Cloud DLP documentation

Scanning and classifying Cloud Storage files

In Vertex AI, the Gen AI Evaluation Service is specifically designed for this workflow. It allows you to perform " side-by-side " or automated evaluations of various Foundation Models (FMs) and large language models (LLMs) using your own datasets.

Efficiency: Using the Evaluation Service API (integrated with Model Garden) is the most efficient approach because it eliminates the need to manually write evaluation scripts or manage complex experiment tracking (Option D).

Internal Benchmarks: The prompt emphasizes using curated internal benchmark datasets . Options B and C are incorrect because they rely on open-source datasets or public leaderboards, which may not represent your company ' s specific use cases or data distribution.

According to the web search results, Reduction Server is a faster GPU all-reduce algorithm developed at Google that uses a dedica ted set of reducers to aggregate gradients from workers 1 2 .  Reducers are lightweight CPU VM instances that are significantly cheaper than GPU VMs 2 .  Therefore, the third worker pool should not have any accelerators, and should use a machine type that has high network bandwidth to optimize the communication between workers and reducers 2 .  TPUs are not supported by Reduction Server, so the first two worker pools should have GPUs and use a container image that contains the training code 1 2 .  The reduction-server container image is provided by Google and should be used for the third worker pool 2 .