I85PREP
⌘ K
GCP ML Engineer study materials
GCP ML Engineer study resources

GCP ML Engineer (Google Cloud Certifications) practice questions

One question at a time. Build your GCP ML Engineer understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Architect low-code AI solutions

A team has a labeled CSV of customer records and wants to train a classification model from the Google Cloud console without writing any training code or packaging a container. Which Vertex AI capability should they use?

  1. Vertex AI Pipelines built from prebuilt Kubeflow components
  2. AutoML training on a tabular dataset Correct answer.
  3. Custom training with a user-provided container image
  4. A Vertex AI Workbench notebook running scikit-learn locally

Explanation

Vertex AI AutoML performs data splitting, architecture search, and training entirely from the console, so the user supplies labeled data and no training code. It is the managed option designed for teams that want a model without writing ML code.

Question 2Architect low-code AI solutions

Which Google Cloud product is designed to extract structured fields such as vendor name and total amount from scanned invoices and receipts?

  1. Video Intelligence API
  2. Document AI Correct answer.
  3. Speech-to-Text API
  4. Cloud Translation API

Explanation

Document AI offers specialized processors, including invoice and receipt parsers, that return structured key-value fields rather than raw text. The other products work on different input modalities and do not perform document field extraction.

Question 3Architect low-code AI solutions

Which service lets analysts train and serve machine learning models by writing SQL statements against data that already resides in a data warehouse?

  1. BigQuery ML Correct answer.
  2. Cloud Dataproc
  3. Vertex AI Pipelines
  4. Cloud Dataflow

Explanation

BigQuery ML lets you create, train, evaluate, and predict with models using SQL, so the data stays inside BigQuery and no separate ML code is required. The other options are orchestration or data-processing services that require additional work to train a model.

Question 4Architect low-code AI solutions

Which pre-trained API should be used to turn call-center audio recordings into searchable text?

  1. Cloud Translation API
  2. Dialogflow CX
  3. Speech-to-Text API Correct answer.
  4. Vision API

Explanation

Speech-to-Text converts audio into text and supports features such as speaker diarization and model adaptation for telephony audio. It is the only listed product whose input is recorded audio.

Question 5Architect low-code AI solutions

What is the primary purpose of Vertex AI Model Garden?

  1. To configure VPC Service Controls perimeters around projects
  2. To schedule recurring notebook executions on a managed runtime
  3. To find and deploy pre-trained and foundation models Correct answer.
  4. To store, version, and label the datasets used for AutoML training

Explanation

Model Garden is a catalog where you can discover, evaluate, deploy, and sometimes tune foundation and pre-trained models. It is a model discovery surface, not a dataset store, scheduler, or security control plane.

Question 6Architect low-code AI solutions

Which Vertex AI resource provides a managed JupyterLab environment for interactive data exploration and prototyping?

  1. Vertex AI Endpoints
  2. Vertex AI Model Registry
  3. Vertex AI Workbench Correct answer.
  4. Vertex AI Pipelines

Explanation

Vertex AI Workbench provides managed JupyterLab notebook instances that are integrated with other Vertex AI resources. Endpoints serve predictions, the registry catalogs models, and pipelines orchestrate repeatable workflows.

Question 7Architect low-code AI solutions

Which product is intended for building multi-turn conversational agents using intents, flows, and pages with minimal custom coding?

  1. Cloud Vision API
  2. Document AI
  3. Vertex AI Batch Prediction
  4. Dialogflow CX Correct answer.

Explanation

Dialogflow CX models conversations as flows and pages with intents, and includes built-in natural language understanding so developers write little or no custom ML code. The other products address document parsing, image analysis, and offline scoring.

Question 8Architect low-code AI solutions

Within Vertex AI, what is Generative AI Studio used for?

  1. Labeling image datasets before an AutoML training job
  2. Designing prompts and tuning foundation models Correct answer.
  3. Scheduling retraining pipelines on a cron cadence
  4. Monitoring prediction drift on a deployed endpoint

Explanation

Generative AI Studio is a console for writing and comparing prompts, adjusting model parameters, and creating tuned versions of foundation models with little or no code. Monitoring, scheduling, and labeling are handled by other Vertex AI features.

Question 9Architect low-code AI solutions

A retailer needs to sort product photos into 40 categories that are specific to its own catalog, and it has roughly 5,000 labeled images. Which approach is most appropriate?

  1. Run the Document AI OCR processor over each product photo
  2. Add a custom phrase set to the Cloud Vision API before calling it
  3. Call the pre-trained Cloud Vision API label detection method on each photo
  4. Train an AutoML image classification model on the labeled images Correct answer.

Explanation

AutoML image classification learns the company's own category definitions from labeled examples, which is required when labels fall outside any fixed taxonomy. Pre-trained Vision labels come from Google's general taxonomy and cannot be redefined.

Question 10Architect low-code AI solutions

An analyst has a terabyte-scale transactions table in BigQuery and wants a quick logistic regression baseline for churn without moving the data elsewhere. What is the most direct approach?

  1. Copy the table into a Vertex AI Workbench notebook and train with scikit-learn
  2. Build a Vertex AI Pipelines DAG that queries BigQuery and trains a model
  3. Create the model with a BigQuery ML CREATE MODEL statement Correct answer.
  4. Export the table to Cloud Storage and run a Vertex AI custom training job

Explanation

BigQuery ML trains supported model types directly on warehouse tables with SQL, so no data movement or custom code is needed for a baseline. The other paths either move the data out or add orchestration overhead.

Question 11Architect low-code AI solutions

Which Vertex AI component is used to define, schedule, and monitor a repeatable multi-step ML workflow that includes data preparation, training, and evaluation?

  1. Vertex AI Feature Store
  2. Vertex AI Endpoints
  3. Vertex AI Model Registry
  4. Vertex AI Pipelines Correct answer.

Explanation

Vertex AI Pipelines runs ML workflows expressed as a directed acyclic graph of components and supports scheduled and monitored executions. Feature storage, model cataloging, and prediction serving are separate services.

Question 12Architect low-code AI solutions

When a dataset is submitted to Vertex AI AutoML and no data split column or key is specified, what does AutoML do?

  1. It rejects the dataset until an explicit split column is provided
  2. It randomly divides the rows into training, validation, and test sets Correct answer.
  3. It trains on every row and therefore reports no evaluation metrics
  4. It treats the last column of the table as the test set

Explanation

Without a user-supplied split, AutoML performs a random split of the rows into training, validation, and test portions. This keeps a held-out test set available for reporting evaluation metrics.

Question 13Architect low-code AI solutions

A team must score 20 million stored records every night, and results are only needed by the next morning. Which Vertex AI prediction mode is the most cost-effective fit?

  1. Real-time prediction invoked from a Cloud Function per record
  2. Per-record streaming prediction through a Dataflow pipeline
  3. Online prediction from an always-on endpoint with autoscaling
  4. Batch prediction writing output to Cloud Storage or BigQuery Correct answer.

Explanation

Batch prediction processes a large table or file in a single job and writes the results to Cloud Storage or BigQuery, avoiding the cost of a continuously available endpoint. Because there is no latency requirement, offline scoring is the appropriate mode.

Question 14Architect low-code AI solutions

Which Vertex AI capability keeps a versioned catalog of trained models together with their lineage and lets you deploy a specific version to an endpoint?

  1. Vertex AI Model Registry Correct answer.
  2. Vertex AI Experiments
  3. Vertex AI Feature Store
  4. Vertex AI Workbench

Explanation

The Model Registry stores model versions with metadata and lineage and is the entry point for deploying a chosen version to an endpoint. Experiments track runs and metrics, which is a different concern.

Question 15Architect low-code AI solutions

A finance team receives thousands of PDF invoices each month and needs vendor name, invoice date, and total amount extracted as structured fields with minimal coding. Which option fits best?

  1. A Document AI OCR processor, followed by custom regular-expression parsing
  2. A Document AI invoice parser processor Correct answer.
  3. Cloud Vision API text detection on each page of every invoice
  4. An AutoML text classification model trained on the invoice text

Explanation

The Document AI invoice parser is a pre-trained processor that returns labeled fields such as vendor, date, and amount directly. OCR alone yields raw text, which would require custom parsing code to produce those fields.

Question 16Architect low-code AI solutions

Which Vertex AI service stores feature values and serves them at low latency for both training and online prediction, helping avoid training-serving skew?

  1. Vertex AI Model Registry
  2. Vertex AI Pipelines
  3. Vertex AI Feature Store Correct answer.
  4. BigQuery ML

Explanation

Feature Store provides a central repository of feature values that can be read in bulk for training and in near real time for online serving. Using one definition for both paths is what prevents training-serving skew.

Question 17Architect low-code AI solutions

A Gemini-powered assistant must answer questions about a company's internal policy documents and cite the source passages. Which approach best reduces unsupported answers?

  1. Ground responses in a Vertex AI Search data store built from the policy documents Correct answer.
  2. Increase the temperature setting so the model explores more possible answers
  3. Fine-tune the model on a large public corpus so it learns the domain vocabulary
  4. Add a system instruction telling the model to answer only from its memory

Explanation

Grounding retrieves relevant passages from an indexed corpus and supplies them to the model at request time, so answers are tied to actual documents and can cite them. This is the standard way to reduce hallucination for private, changing content.

Question 18Architect low-code AI solutions

In Vertex AI AutoML, which portion of the data is used to report the model's final evaluation metrics?

  1. The validation split, which is used during model selection
  2. The test split, which is held out from training Correct answer.
  3. The training split, measured after data augmentation
  4. The complete dataset, including all three splits

Explanation

AutoML reserves a held-out test split and computes the reported evaluation metrics on it, which gives an unbiased estimate of generalization. The validation split guides model selection during training rather than producing final metrics.

Question 19Architect low-code AI solutions

A translation workflow must always render the company's product names using approved equivalents in each target language. Which Cloud Translation feature supports this requirement?

  1. Batch translation of document files
  2. Automatic source language detection
  3. Turning on HTML translation mode for the request
  4. A glossary of term pairs attached to the translation request Correct answer.

Explanation

A glossary supplies mandatory term mappings that the translation service applies consistently, which is how product names are pinned to approved equivalents. Detection, batching, and markup handling do not control terminology.

Question 20Architect low-code AI solutions

Which pre-trained API can identify labels, detect shot changes, and flag explicit content in uploaded video files without any custom training?

  1. AutoML Video Classification
  2. Video Intelligence API Correct answer.
  3. Cloud Vision API
  4. Speech-to-Text API

Explanation

The Video Intelligence API offers pre-trained annotation for videos, including label detection, shot detection, and explicit content detection, with no training required. AutoML Video would instead require a labeled custom dataset.

Question 21Architect low-code AI solutions

A team needs a custom feature-engineering step and a model architecture that AutoML does not offer for a 500 GB BigQuery dataset. Which path is most appropriate?

  1. Raise the AutoML model's classification threshold after training
  2. Change the AutoML optimization objective to a different metric
  3. Run custom training on Vertex AI using a prebuilt or custom container Correct answer.
  4. Retrain the existing AutoML model with a longer training budget

Explanation

When the required preprocessing and architecture are outside AutoML's supported options, custom training on Vertex AI gives full control over code and container while still using managed infrastructure. Adjusting AutoML settings cannot add unsupported architectures.

Question 22Architect low-code AI solutions

A company needs a Gemini model to return answers in a strict JSON structure, and the underlying knowledge base changes every day. Which approach is most appropriate?

  1. Enforce the output schema through prompting and retrieve context from a daily-refreshed index Correct answer.
  2. Supervised fine-tuning on the full knowledge base, rebuilt and retrained each day
  3. Raise the temperature and add more few-shot examples that show the schema
  4. Pre-train a domain-specific foundation model from scratch on the corpus

Explanation

Output format is best handled through prompt instructions and response-schema constraints, while changing knowledge is best supplied at request time through retrieval rather than baked into weights. Combining the two keeps answers current without daily retraining.

Question 23Architect low-code AI solutions

A BigQuery ML model performs well, and the team now needs low-latency online predictions from an application that cannot call BigQuery on every request. What is the recommended low-code path?

  1. Export the model to Vertex AI Model Registry and deploy it to an endpoint Correct answer.
  2. Run a BigQuery query through the API on every prediction request
  3. Export the artifacts to Cloud Storage and self-host them on GKE
  4. Copy the training data into a Vertex AI dataset and retrain with AutoML

Explanation

BigQuery ML models can be exported and registered in Vertex AI Model Registry, then deployed to an endpoint for low-latency online serving without rewriting the model. This preserves the trained model while moving serving to the appropriate runtime.

Question 24Architect low-code AI solutions

A fraud model trained on a heavily imbalanced dataset must catch at least 90% of fraud cases while keeping false positives manageable. Which action is most appropriate at the low-code level?

  1. Increase the training budget so AutoML searches more architectures
  2. Tune the score threshold using the reported precision-recall metrics Correct answer.
  3. Remove majority-class rows until the classes are exactly balanced
  4. Retrain repeatedly with different random seeds until recall improves

Explanation

For imbalanced problems the decision threshold controls the precision-recall trade-off, and AutoML evaluation output shows how metrics move as the threshold changes. Selecting a threshold that meets the recall target is the direct, low-code way to satisfy the requirement.

Question 25Architect low-code AI solutions

A team wants an assistant that answers HR policy questions from a document repository and can also look up an employee's leave balance in a backend API, with minimal custom code. Which Google Cloud approach fits best?

  1. An agent built with Vertex AI Agent Builder using a grounded data store and an API tool Correct answer.
  2. An AutoML text classification model plus a Cloud Function that calls the API
  3. A BigQuery ML remote model that calls a foundation model endpoint
  4. A Dialogflow CX agent whose intents return only static, pre-written responses

Explanation

Agent Builder supports grounding an agent in a document data store for retrieval-based answers and attaching tools or extensions so the agent can call backend systems for live data. That covers both the knowledge question and the action in one low-code configuration.

Question 26Architect low-code AI solutions

A data science team wants to browse a catalog inside Vertex AI, try foundation and task-specific models interactively, and deploy one to an endpoint without training anything itself. Which Vertex AI resource serves this purpose?

  1. Model Garden, a searchable catalog of foundation and task-specific models Correct answer.
  2. Vertex AI Pipelines, an orchestrator for repeatable training workflows
  3. Vertex AI Experiments, a tracker for training runs and their metrics
  4. Vertex AI Workbench, a managed notebook environment for exploration

Explanation

Model Garden is the catalog where foundation models, partner models, and task-specific models can be discovered, tried in a console, and deployed to an endpoint. It requires no model training by the customer, so it fits a team that wants to consume a model rather than build one.

Question 27Architect low-code AI solutions

An operations manager who does not program needs an internal mobile app that displays open work orders and lets staff update their status, with the records stored in an existing BigQuery table. Which approach delivers the app with the least custom code?

  1. Write a Cloud Run service that wraps the BigQuery table in a REST API and build a client against it
  2. Schedule a BigQuery query that emails a spreadsheet of work orders each hour
  3. Build a Looker Studio dashboard with filter controls for the work order fields
  4. Create an AppSheet app that uses the BigQuery table directly as its data source Correct answer.

Explanation

AppSheet is a no-code application platform that can read and write an external data source such as BigQuery, producing a usable mobile app without programming. Looker Studio and scheduled queries only present data read-only, and a custom API requires engineering effort.

Question 28Architect low-code AI solutions

A support organization wants to turn recorded call-center audio into searchable text within days, using a pre-trained service rather than training its own acoustic model. Which Google Cloud API should it call?

  1. The Text-to-Speech API, which renders written text as spoken audio
  2. The Cloud Translation API, which converts text between languages
  3. The Speech-to-Text API, which transcribes audio into text with pre-trained models Correct answer.
  4. The Natural Language API, which returns entities and sentiment for text input

Explanation

Speech-to-Text accepts audio such as call recordings and returns a transcript using pre-trained models, so no custom training is needed. It is the only API listed that consumes audio and produces text.

Question 29Architect low-code AI solutions

A retailer stores three years of daily unit sales per store and SKU in BigQuery and needs weekly forecasts for the next quarter. The forecasts must account for seasonality, promotions, and holidays, and the team will not write training code. Which Vertex AI option fits best?

  1. A Vertex AI Pipelines job running a hand-written forecasting component
  2. Vertex AI Forecasting, which trains one model across many related time series Correct answer.
  3. AutoML tabular regression on a table whose rows are individual store-days
  4. BigQuery ML k-means clustering on the historical sales amounts

Explanation

Vertex AI Forecasting is purpose-built for time series data, learning across many related series and handling seasonality and holiday effects with no code. A plain regression or clustering approach would not model the temporal structure the forecast depends on.

Question 30Architect low-code AI solutions

A model has been serving on a Vertex AI endpoint for several months, and the team suspects that incoming requests now differ from the training data. They want automatic alerts when input feature distributions move away from the training baseline, without building their own tooling. What should they enable?

  1. Vertex AI Model Monitoring on the endpoint with drift thresholds configured Correct answer.
  2. A BigQuery scheduled query that recomputes feature averages each morning
  3. Cloud Logging alerts triggered by an elevated prediction error rate
  4. A nightly Vertex AI Pipelines run that retrains the model from scratch

Explanation

Model Monitoring compares the feature distributions of prediction requests against a baseline captured from training data and raises alerts when drift or skew exceeds configured thresholds. It provides this detection without custom tooling.

Question 31Architect low-code AI solutions

When a Vertex AI AutoML tabular model is trained, how are the input feature columns handled?

  1. Any column containing a missing value is dropped from the training data
  2. Built-in transformations such as normalization and one-hot encoding are applied automatically Correct answer.
  3. Only linear algorithms are considered so that each transformation stays interpretable
  4. Every column must be pre-encoded into numeric values before the dataset is submitted

Explanation

AutoML tabular performs feature transformation, including scaling numeric columns and encoding categorical columns, as part of training and then searches over candidate model architectures. Users therefore supply the raw columns rather than hand-engineering the encodings.

Question 32Architect low-code AI solutions

An inspector takes a photo of each finished part and needs a short written verdict back within seconds. The team has no labeled training data and wants to prototype with a single prompt. Which approach should they use?

  1. Assemble the photos into a video and call the Video Intelligence API for labels
  2. Send each photo together with a text instruction to a Gemini multimodal model Correct answer.
  3. Collect labels for the photos and train a separate AutoML image model per part type
  4. Call logo detection on each photo and map detected logos to verdict categories

Explanation

Gemini multimodal models accept an image and a text instruction in the same request and return generated text, so a working prototype needs only a prompt and no labeled dataset. Training a custom model, by contrast, requires a labeled image set.

Question 33Architect low-code AI solutions

A legal team uses a Document AI processor to extract fields from contracts. Fields whose confidence score falls below a threshold must be checked by a person before the values feed downstream systems. Which setup accomplishes this with the least custom code?

  1. A Cloud Run service that resubmits the document until every field confidence rises above the threshold
  2. A BigQuery scheduled query that selects low-confidence rows and emails the output to the team
  3. An AutoML tabular model trained on past confidence scores to predict which documents will fail
  4. A Document AI processor whose low-confidence documents are routed to reviewers through human-in-the-loop review Correct answer.

Explanation

Document AI integrates with a human-in-the-loop review workflow that sends documents with low-confidence fields to a reviewer queue and returns the corrected values. This provides the review step without building custom review tooling.

Question 34Architect low-code AI solutions

A support team has about 150 labeled tickets, and its set of issue categories is revised every month as products change. They want a low-code classifier that stays accurate as the taxonomy evolves. Which approach is most appropriate?

  1. Use the Cloud Translation API to map each incoming ticket onto the nearest category
  2. Train a BigQuery ML logistic regression model on the ticket text and refresh it monthly
  3. Train a Vertex AI AutoML text classification model and retrain it from scratch each month
  4. Prompt a Gemini model with the current category definitions and a few examples, revising the prompt when categories change Correct answer.

Explanation

When labels change frequently and the labeled set is small, prompt-based classification with Gemini keeps the category rules in the prompt, so a taxonomy change is a prompt edit rather than a new training cycle. It also avoids the cost and lag of retraining a supervised model every month.

Question 35Collaborate within and across teams to manage data and models

A data scientist on a shared Vertex AI project needs to submit training jobs and create endpoints, but must not be able to modify project IAM policies. Which predefined role should be granted?

  1. roles/aiplatform.user Correct answer.
  2. roles/iam.securityAdmin
  3. roles/aiplatform.admin
  4. roles/editor

Explanation

Vertex AI User grants the ability to create and use Vertex AI resources such as training jobs, models and endpoints without granting permission to change IAM policy bindings. Administrative and basic roles carry broader permissions than this task requires.

Question 36Collaborate within and across teams to manage data and models

Team A owns a BigQuery dataset containing PII columns, and Team B needs to read only three non-PII columns without receiving any access to the underlying tables. Which approach meets the requirement?

  1. Grant Team B roles/bigquery.dataViewer on the source dataset and rely on them to filter columns in queries.
  2. Grant Team B roles/bigquery.dataEditor on the source dataset so they can define their own restricted views.
  3. Export the three columns nightly to a Cloud Storage bucket and grant Team B object viewer on the bucket.
  4. Create an authorized view in a second dataset that selects the three columns, then grant Team B access to that second dataset. Correct answer.

Explanation

An authorized view lets the view's owner grant consumers access to query results without granting access to the source tables. Granting Team B access to the dataset that holds the view exposes only the projected columns.

Question 37Collaborate within and across teams to manage data and models

A team publishes a new model build every week. Consumers should always serve the most recently approved build without changing their code. What should the team do in Vertex AI Model Registry?

  1. Register each build as a new version of the same parent model and move a production alias to the approved version. Correct answer.
  2. Write each build to a new Cloud Storage bucket and tell consumers the bucket name at release time.
  3. Register each build as a separate top-level model whose display name includes the build number.
  4. Create a separate endpoint per build and ask consumers to update the endpoint ID at each release.

Explanation

Model Registry groups versions under a parent model resource, and version aliases let consumers resolve a stable reference that the owning team repoints after approval. This keeps consumer code unchanged across releases.

Question 38Collaborate within and across teams to manage data and models

A Vertex AI pipeline must read a BigQuery dataset owned by the analytics team and write artifacts into a Cloud Storage bucket owned by the platform team. What is the recommended way to grant access?

  1. Create a dedicated service account, grant it the required dataset and bucket roles, and run the pipeline as it. Correct answer.
  2. Run the pipeline as the developer's user account, which already holds the required access.
  3. Copy the source data into the pipeline project so no cross-team permissions are needed.
  4. Grant the pipeline's service account roles/owner on the project so permission errors cannot occur.

Explanation

Workloads should run under a dedicated service account with least-privilege roles scoped to the specific dataset and bucket. This keeps the job independent of any individual's account and survives personnel changes.

Question 39Collaborate within and across teams to manage data and models

Two separate companies want to share a BigQuery dataset without copying data, with subscriber management and usage visibility on the provider side. Which capability fits best?

  1. Sharing a signed Cloud Storage extract that consumers load into their own datasets
  2. Granting each subscriber group roles/bigquery.dataViewer directly on the source dataset
  3. Publishing the dataset as a listing in an Analytics Hub data exchange and subscribing from the consumer project Correct answer.
  4. Scheduling a BigQuery Data Transfer Service job that copies tables into the consumer project nightly

Explanation

Analytics Hub data exchanges publish datasets as listings that subscribers attach from their own projects without copying data, while the provider manages subscriber access and can observe usage. Direct grants lack the listing-level control and subscriber management model.

Question 40Collaborate within and across teams to manage data and models

Analysts across several teams cannot tell which BigQuery tables contain churn features or how the columns are defined. Which practice best makes the data discoverable?

  1. Attach Dataplex and Data Catalog tags, descriptions and glossary terms to the tables and columns. Correct answer.
  2. Grant every team roles/bigquery.dataViewer on all datasets in the organization.
  3. Email a quarterly spreadsheet of table names and owners to every analytics team.
  4. Move all feature tables into one dataset named features and rely on table names for meaning.

Explanation

Dataplex and Data Catalog provide searchable metadata, tags and business glossary terms so teams can find and interpret datasets without out-of-band documentation. Metadata is attached to resources and stays current as part of the data lifecycle.

Question 41Collaborate within and across teams to manage data and models

An organization policy requires that a shared BigQuery dataset be encrypted with a customer-managed key that the owning team controls and can revoke. Which configuration satisfies this?

  1. Enable CMEK in each consumer project and let BigQuery re-encrypt the shared tables on read.
  2. Set the Cloud KMS key on the Cloud Storage export bucket rather than on the BigQuery dataset itself.
  3. Create a Cloud KMS key in the owner project, grant the BigQuery service agent use of it, and set it as the dataset default key. Correct answer.
  4. Use Google-managed encryption and rely on VPC Service Controls to block unauthorized reads of the dataset.

Explanation

BigQuery applies CMEK per dataset or table using a key from the project where the data resides, and the BigQuery service agent must hold encrypt and decrypt permission on that key. Revoking the key binding removes the service's ability to decrypt the data.

Question 42Collaborate within and across teams to manage data and models

Several teams need the same near-real-time features for online prediction. Which arrangement supports sharing while preserving least privilege?

  1. Features are exported hourly to CSV files in a shared bucket that each team loads into its own store.
  2. Each team builds and maintains its own copy of the feature pipeline to avoid coordination.
  3. A central Vertex AI Feature Store in a shared project publishes feature views, and consuming teams receive read access to those views. Correct answer.
  4. Every team receives roles/editor on the feature project so they can read and write features freely.

Explanation

A centrally governed feature store with published feature views lets many teams consume identical feature definitions while the owning team controls write access. Consumers receive read-only access scoped to the views they need.

Question 43Collaborate within and across teams to manage data and models

Two teams each own a model and want to serve both behind a single hostname while sending each model a share of the traffic. What should they do in Vertex AI?

  1. Merge the two models into one custom container image and deploy that image once.
  2. Deploy both models to the same endpoint and route traffic by percentage across the deployed model IDs. Correct answer.
  3. Deploy only one model and have its container call the other model's endpoint internally.
  4. Deploy both models to separate endpoints and place an external load balancer in front of both.

Explanation

A Vertex AI endpoint can host multiple deployed models, each with its own traffic split percentage, which allows shared serving with a single hostname and controlled rollout. Traffic allocation is adjusted per deployed model ID.

Question 44Collaborate within and across teams to manage data and models

Two engineers must co-develop notebooks in a shared Vertex AI environment without overwriting each other's runtime state. Which approach is recommended?

  1. Both engineers sign in to the same user account so the instance state is identical for both.
  2. Send the .ipynb file back and forth by email and merge the changes at the end of each sprint.
  3. Both engineers receive roles/aiplatform.admin and work concurrently inside one Workbench instance.
  4. Keep notebooks in a shared Git repository or bucket and give each engineer a separate Workbench instance. Correct answer.

Explanation

Workbench instances are single-user environments, so collaboration happens through version-controlled notebook files while each engineer runs an independent instance. This avoids runtime conflicts and preserves an auditable history of changes.

Question 45Collaborate within and across teams to manage data and models

Auditors ask which dataset version, code commit and hyperparameters produced a model that is currently deployed. Which mechanism provides an authoritative answer?

  1. Reading the model's display name, which encodes the training date and run number.
  2. Reviewing Cloud Audit Logs for the training job's start time and caller identity.
  3. Querying Vertex ML Metadata lineage to trace the model artifact back to its execution and input datasets. Correct answer.
  4. Inspecting the endpoint's request logs to infer which data was used during training.

Explanation

Vertex ML Metadata records artifacts, executions and their relationships, so a model artifact can be traced to the pipeline execution that produced it along with its input datasets and parameters. Audit logs record who called an API but not the data and configuration lineage.

Question 46Collaborate within and across teams to manage data and models

A team-owned Cloud Storage bucket must let another team upload new objects while preventing them from reading or deleting existing ones. Which grant achieves this?

  1. Grant roles/storage.objectAdmin on the bucket to the other team's group.
  2. Grant roles/storage.objectViewer on the bucket to the other team's group.
  3. Grant roles/storage.objectCreator on the bucket to the other team's group. Correct answer.
  4. Make the bucket public and require the other team to use signed URLs for uploads.

Explanation

Object Creator allows creating objects in a bucket without granting read, overwrite or delete rights on existing objects. Combined with uniform bucket-level access, bucket-level IAM controls all object operations consistently.

Question 47Collaborate within and across teams to manage data and models

A shared BigQuery table must remain directly queryable by many teams, but its salary column may only be read by the HR group. Which mechanism enforces this on the table itself?

  1. Granting roles/bigquery.dataViewer through IAM Conditions limited to business hours
  2. Applying policy tags from a Data Catalog taxonomy to the sensitive column and granting fine-grained reader only to HR Correct answer.
  3. Creating a separate view per consumer team and granting viewer on each view
  4. Renaming sensitive columns with a confidential prefix so consumers avoid querying them

Explanation

Policy tags provide column-level access control on the table itself, so users with dataset access can query permitted columns while the tagged column returns an access error for anyone lacking fine-grained reader on the tag. This keeps the table directly queryable without per-team copies.

Question 48Collaborate within and across teams to manage data and models

Data scientists want teammates to review and compare the metrics of their Vertex AI Experiments runs. What is required?

  1. Each scientist exports run metrics to a personal spreadsheet so results can be combined later.
  2. Runs can only be compared after each model has been deployed to a prediction endpoint.
  3. Experiments are private to each user and must be cloned into each teammate's own project before viewing.
  4. Experiments are project-scoped resources, so granting teammates an appropriate Vertex AI role on the project lets them view and compare runs. Correct answer.

Explanation

Vertex AI Experiments and their runs live in the project, and IAM controls who can read them. Granting a suitable Vertex AI role on the shared project lets colleagues open the same experiments and compare metrics and parameters.

Question 49Collaborate within and across teams to manage data and models

A model is handed to other teams for reuse, and they need authoritative statements of intended use, limitations and evaluation results. Where should that information live?

  1. In a model card attached to the model version in Vertex AI Model Registry. Correct answer.
  2. In the team wiki, linked from the deployment script.
  3. In the commit message of the training code in the source repository.
  4. As resource labels attached to the deployed endpoint.

Explanation

Model cards are the standardized place to record intended use, limitations and evaluation metrics alongside a registered model version, so consumers see the documentation with the model itself. Attaching it to the version keeps documentation aligned with the exact artifact in use.

Question 50Collaborate within and across teams to manage data and models

A platform team must let every analyst group in a folder of projects read BigQuery data, using least privilege and without granting per-project or per-dataset bindings. What should they do?

  1. Consolidate all datasets into one project and grant the role on that project only.
  2. Grant roles/bigquery.dataViewer to a group containing all analyst groups at the folder level so the projects inherit it. Correct answer.
  3. Grant roles/bigquery.dataViewer to individual analyst accounts at the organization level.
  4. Grant the role on each dataset in every project and re-apply the grants whenever a new project is created.

Explanation

IAM policies are inherited downward through the resource hierarchy, so a single folder-level binding reaches every current and future project in the folder. Binding to a group at the folder keeps administration centralized and avoids over-privileged org-level grants.