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

Which input can be imported directly to create a Vertex AI AutoML tabular dataset?

  1. A folder of unlabeled JPEG images in Cloud Storage
  2. A PDF file set stored in a Cloud Storage bucket
  3. A BigQuery table or a CSV file in Cloud Storage Correct answer.
  4. A Pub/Sub topic that streams JSON events

Explanation

AutoML tabular datasets accept structured, column-oriented data, which Vertex AI can read from a BigQuery table or from CSV files in Cloud Storage. The schema of the columns becomes the feature set used for training. Documents, raw images and message streams are handled by other dataset types or ingestion paths.

Question 2Architect low-code AI solutions

Which statement trains a machine learning model inside BigQuery ML?

  1. CREATE MODEL ... OPTIONS(...) AS SELECT ... Correct answer.
  2. CREATE VIEW ... AS SELECT ...
  3. CREATE TABLE ... AS SELECT ...
  4. CREATE FUNCTION ... AS '...'

Explanation

BigQuery ML trains models with a CREATE MODEL statement whose OPTIONS clause selects the model type and hyperparameters, and whose query supplies the training rows. The resulting model object is stored in a BigQuery dataset and can be used with ML functions. The other statements create tables, functions or views, none of which produce a model.

Question 3Architect low-code AI solutions

You have daily sales data per store and want BigQuery ML to produce a forecast that captures trend and seasonality. Which model type should you choose?

  1. KMEANS
  2. LINEAR_REG
  3. ARIMA_PLUS Correct answer.
  4. LOGISTIC_REG

Explanation

ARIMA_PLUS is the BigQuery ML time-series model designed for forecasting, automatically modeling trend, seasonality and holiday effects, and it supports ML.FORECAST for predictions with intervals. LINEAR_REG and LOGISTIC_REG are regression classifiers, and KMEANS is unsupervised clustering, so none of them model temporal structure natively.

Question 4Architect low-code AI solutions

You start a Vertex AI AutoML tabular training job without providing a split column or manual data split fractions. What happens?

  1. Rows are randomly split about 80/10/10 into training, validation and test Correct answer.
  2. The dataset is rejected because a split column is mandatory
  3. All rows are used for training and evaluation metrics come from the training data
  4. Rows are split evenly 50/50 into training and test sets

Explanation

When no split configuration is supplied, AutoML applies a default random split of roughly 80% training, 10% validation and 10% test. The validation slice is used during training for model selection and the test slice for the reported evaluation metrics. Supplying a split column or fractions overrides this default.

Question 5Architect low-code AI solutions

What is Vertex AI Studio primarily used for?

  1. Designing and testing prompts for generative models Correct answer.
  2. Managing human labeling tasks for images
  3. Serving custom containers on GKE
  4. Running distributed TensorFlow training jobs

Explanation

Vertex AI Studio gives a console workspace where you write and iterate on prompts, compare model responses, adjust parameters, and export sample code for the API call. It targets generative AI prototyping rather than training infrastructure, data labeling or cluster operations.

Question 6Architect low-code AI solutions

What does Vertex AI Model Garden provide?

  1. A catalog of foundation and task-specific models Correct answer.
  2. A managed service for training scikit-learn models
  3. A gallery of Dataflow pipeline templates
  4. A repository of public BigQuery datasets

Explanation

Model Garden is a curated catalog where you can browse foundation models and task-specific models from Google and partners, then deploy them to an endpoint, tune them, or open their model cards. It is a discovery and provisioning surface, not a training service, dataset repository or pipeline gallery.

Question 7Architect low-code AI solutions

You want BigQuery ML to search several values of an L2 regularization strength while training a model. Which addition to CREATE MODEL accomplishes this?

  1. Switching the model type to ARIMA_PLUS
  2. Adding more rows to the training table
  3. NUM_TRIALS = 20 plus an HPARAM_RANGE for the tuned option Correct answer.
  4. Raising MAX_ITERATIONS to 20

Explanation

BigQuery ML performs hyperparameter tuning when you set NUM_TRIALS to a value greater than one and declare the search space with HPARAM_RANGE, MIN or MAX around the hyperparameter. The service then runs multiple trials and keeps the best-performing configuration. Iteration counts, model type changes and extra rows do not create a search over hyperparameters.

Question 8Architect low-code AI solutions

What is the main purpose of Document AI?

  1. Transcribing speech to text in real time
  2. Extracting structured fields from documents Correct answer.
  3. Building multi-turn agents with flows and pages
  4. Scoring clickstream events for recommendations

Explanation

Document AI applies OCR plus specialized parsers to turn scanned or digital documents into structured fields such as invoice totals or form entries. Each processor targets a document type and returns typed entities rather than raw text alone. The other options describe speech, conversational and recommendation services.

Question 9Architect low-code AI solutions

A team must build a complex customer-service agent with many multi-turn branches and explicit conversation state. Why is Dialogflow CX the appropriate choice over a simpler intent-matching agent?

  1. It requires no training phrases at all
  2. It stores conversation state in BigQuery
  3. It models an agent as a state machine of flows and pages Correct answer.
  4. It supports voice channels only

Explanation

Dialogflow CX organizes an agent into flows and pages, with explicit state and transition routes, which suits long, branching conversations that must track where the user is. Intent matching still exists but is embedded in a structured dialogue design. This state model is what distinguishes it from flatter, single-flow designs.

Question 10Architect low-code AI solutions

You need a generative assistant that answers questions over your company's internal policy PDFs and cites the source pages. Which low-code Google Cloud approach fits best?

  1. AutoML Tables trained on query logs
  2. BigQuery ML with a linear regression model
  3. Vertex AI Vector Search with a hand-built index
  4. Vertex AI Search with the documents loaded into a data store Correct answer.

Explanation

Vertex AI Search ingests documents into a data store, indexes them, and can ground generative responses with citations, requiring little custom code. Vector Search provides embeddings-based retrieval but you must build and maintain the index and retrieval pipeline yourself. Regression, classification and clustering models do not perform document retrieval.

Question 11Architect low-code AI solutions

What is AppSheet?

  1. A no-code platform for building apps over existing data sources Correct answer.
  2. A registry for versioning container images
  3. A streaming ETL service for event data
  4. A managed notebook for Python development

Explanation

AppSheet lets business users create mobile and web applications from existing data such as spreadsheets, databases or APIs without writing application code. It is a low-code application layer, distinct from notebook, data pipeline and artifact storage services.

Question 12Architect low-code AI solutions

Which columns must a Vertex AI AutoML forecasting dataset contain to train a model?

  1. A text column and an entity annotation column
  2. A series ID column, a timestamp column and a target column Correct answer.
  3. Only a target column with numeric values
  4. An image URI column and a label column

Explanation

AutoML forecasting needs to know which entity each row belongs to, when the observation occurred, and the value to predict, so it requires a time series identifier, a timestamp and a target. Additional covariate columns are optional. Without the series and time columns the rows cannot be assembled into time series.

Question 13Architect low-code AI solutions

A team's training data already lives in BigQuery and the analysts are fluent in SQL but not Python, and they want rapid iteration. Which approach fits best?

  1. Export the data to Cloud Storage and train AutoML tabular
  2. Move the data into Dialogflow CX and train an agent
  3. Build a custom TensorFlow model with a data pipeline
  4. Use BigQuery ML so the team keeps working in SQL Correct answer.

Explanation

When the data is already in BigQuery and the team works in SQL, BigQuery ML lets them train, evaluate and predict with SQL statements and no data movement or custom code. AutoML or custom training would add export steps and a new toolchain for the same goal. Dialogflow CX is unrelated to tabular prediction.

Question 14Architect low-code AI solutions

What is the role of Google Cloud Pipeline Components in a low-code Vertex AI workflow?

  1. A self-managed Kubeflow install on GKE
  2. Prebuilt Google Cloud Pipeline Components for each step Correct answer.
  3. A library that only supports custom containers
  4. A BigQuery feature for scheduling queries

Explanation

Google Cloud Pipeline Components are prebuilt pipeline steps, such as dataset creation, AutoML training, model evaluation and endpoint deployment, that you chain together in a Vertex AI pipeline with minimal wrapper code. They remove the need to write each step as a custom container. They do not replace the pipeline runner or the scheduling service.

Question 15Architect low-code AI solutions

A generative model keeps answering with invented policy details. Which change most directly improves factual accuracy against your own documentation?

  1. Fine-tune the model without any examples
  2. Increase the maximum output tokens
  3. Raise the temperature so answers vary more
  4. Ground responses in a Vertex AI Search data store Correct answer.

Explanation

Grounding retrieves relevant passages from an indexed corpus and supplies them to the model, so answers are anchored in your documentation and can cite sources. Sampling parameters and output length change style or verbosity but not the factual basis. Fine-tuning without examples provides no new knowledge at all.

Question 16Architect low-code AI solutions

What distinguishes few-shot prompting from zero-shot prompting?

  1. The request is split across several API calls
  2. The temperature is fixed at zero
  3. A few worked input-output examples are placed in the prompt Correct answer.
  4. The model is asked to explain its reasoning step by step

Explanation

Few-shot prompting supplies a small number of example input-output pairs inside the prompt so the model can infer the desired format and behavior, whereas zero-shot prompting gives only instructions. Neither definition depends on temperature, call splitting or step-by-step reasoning instructions.

Question 17Architect low-code AI solutions

You want to summarize text stored in a BigQuery table using a foundation model hosted on Vertex AI, without moving the data. What must you set up?

  1. Call ML.FORECAST on the text column
  2. Train a KMEANS model over the text column
  3. Create a remote model and call ML.GENERATE_TEXT Correct answer.
  4. Call ML.PREDICT on a locally trained boosted tree

Explanation

BigQuery ML connects to Vertex AI through a Cloud resource connection, and a remote model registered over that connection lets you call ML.GENERATE_TEXT on table columns so inference happens on the hosted model while the data stays in BigQuery. Boosted trees, forecasting and clustering are unrelated model families that cannot call a generative endpoint this way.

Question 18Architect low-code AI solutions

Which evaluation output is produced for a Vertex AI AutoML image classification model?

  1. A silhouette score and a Davies-Bouldin index
  2. RMSE and MAE computed on held-out rows
  3. A confusion matrix at a threshold plus precision and recall Correct answer.
  4. R-squared and adjusted R-squared

Explanation

Classification models are evaluated with a confusion matrix at a selectable confidence threshold and derived metrics such as precision, recall and average precision per label. Error-based regression metrics and clustering indices belong to other model families. Reviewing per-label metrics is how you detect classes the model handles poorly.

Question 19Architect low-code AI solutions

A team keeps retraining models and needs to track several versions, compare their metrics and control which version is deployed. Which Vertex AI component is intended for this?

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

Explanation

Model Registry holds model resources with multiple versions, stores per-version metrics and artifacts, and lets you deploy or roll back a specific version to an endpoint. Feature Store serves feature values, ML Metadata records lineage, and Pipelines orchestrates workflow steps.

Question 20Architect low-code AI solutions

You must extract vendor name, invoice date and line-item totals from scanned invoices with as little custom code as possible. Which Document AI choice fits best?

  1. A Dialogflow CX agent with invoice intents
  2. An AutoML text classification model trained on invoices
  3. An Invoice Parser processor that returns invoice fields Correct answer.
  4. A generic OCR processor that returns raw text only

Explanation

The Invoice Parser is a pretrained specialized processor that returns typed fields such as supplier, invoice date and line items directly, so no training or post-processing code is needed. Raw OCR gives untyped text, classification only assigns a label, and a conversational agent does not parse documents.

Question 21Architect low-code AI solutions

You store embeddings for a large product catalog and need semantic nearest-neighbor lookups with low latency at serving time. Which service is designed for this?

  1. Vector Search, for low-latency nearest-neighbor queries Correct answer.
  2. BigQuery ML with a KMEANS model over the embeddings
  3. Vertex AI Search, because it indexes documents automatically
  4. Dialogflow CX with a knowledge base of the catalog

Explanation

Vector Search is built to store and serve embeddings and answer approximate nearest-neighbor queries at low latency, which is what an embedding-based semantic lookup requires. Vertex AI Search is an end-to-end document search product, KMEANS clusters points rather than answering per-query neighbors, and Dialogflow CX handles conversations.

Question 22Architect low-code AI solutions

How must training data be labeled to train a Vertex AI AutoML video classification model?

  1. Assign one label to the whole video file
  2. Label time segments of each video with start and end times Correct answer.
  3. Provide one label per sampled frame automatically
  4. Provide only the audio transcript for each video

Explanation

AutoML video classification learns from labels attached to time segments, so a video can carry different labels in different intervals and unlabeled stretches are ignored. This segment-level labeling is what lets the model localize when a class appears rather than labeling the entire clip. Frame-level manual labeling is not the required input format.

Question 23Architect low-code AI solutions

How do you make a trained Vertex AI model available for online predictions?

  1. Export the model into a BigQuery dataset
  2. Deploy the model to a Vertex AI endpoint Correct answer.
  3. Publish the model to a Pub/Sub topic
  4. Upload the model to Artifact Registry

Explanation

Online prediction requires deploying the model to an endpoint, which provisions serving resources and returns a URL that clients call with prediction requests. Artifact Registry stores container images, BigQuery holds data, and Pub/Sub carries messages, so none of them serve model predictions.

Question 24Architect low-code AI solutions

You plan to tune a foundation model in Vertex AI with supervised tuning on your own labeled examples. What training input does the service expect?

  1. A TFRecord of images with bounding boxes
  2. A CSV of unlabeled product descriptions in BigQuery
  3. A JSONL file with one input/output example per line Correct answer.
  4. A single long text prompt stored in a table

Explanation

Supervised tuning consumes example pairs, formatted as JSONL where each line pairs an input with the expected output, stored in Cloud Storage. The model learns the mapping demonstrated by these examples. Unlabeled rows, image records and a lone prompt provide no target outputs to learn from.

Question 25Architect low-code AI solutions

You are training a custom translation model in Vertex AI AutoML Translation. What should the dataset contain?

  1. A TSV file of parallel source and target sentences Correct answer.
  2. A plain text file of unaligned paragraphs
  3. A CSV of target-language sentences only
  4. A JSONL of images paired with captions

Explanation

Translation training is supervised and needs aligned sentence pairs, supplied as a TSV where each row holds the source sentence and its target-language translation. Only aligned pairs let the model learn the mapping between languages. Single-language text, images or unaligned paragraphs provide no correspondence to learn from.

Question 26Architect low-code AI solutions

You are importing a CSV file into Vertex AI to train an AutoML tabular model. How do you tell Vertex AI which column the model should learn to predict?

  1. Place the column last in the CSV file so it is read as the label.
  2. Add a second CSV that lists the label value for each row.
  3. Select the column as the target column in the dataset schema after the data is imported. Correct answer.
  4. Name the column "target" so Vertex AI detects it automatically during import.

Explanation

Vertex AI tabular datasets carry a schema, and the column to predict is designated as the target in that schema after import. Vertex AI then uses the target to choose the applicable training objectives and to build evaluation metrics. Column order and column names have no special meaning.

Question 27Architect low-code AI solutions

Training jobs and online prediction must use identical feature values, and prediction needs low-latency reads of those values. Which Vertex AI capability is designed for this?

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

Explanation

Feature Store holds feature values for both offline training and online serving, so the same definitions are reused and low-latency lookups are available at prediction time. This avoids training/serving skew created by reimplementing feature logic separately.

Question 28Architect low-code AI solutions

An AutoML model must be retrained every Monday with no one starting the job manually, and the team wants to stay inside Vertex AI. What should you configure?

  1. Export the model each week and retrain it in a local notebook.
  2. Set the training job's node-hour budget so it resets weekly.
  3. Configure model monitoring to retrain the model when it detects drift.
  4. Create a Vertex AI pipeline schedule that runs on a recurring basis. Correct answer.

Explanation

Vertex AI Pipelines supports schedules that launch pipeline runs at a recurring cadence, which is the managed way to automate retraining. The schedule holds the pipeline specification plus parameters, so results stay reproducible across runs.

Question 29Architect low-code AI solutions

A team wants AutoML training to stop after a fixed amount of compute so costs stay predictable. Which setting controls this?

  1. The machine type of the endpoint that serves the model.
  2. The batch size used by the batch prediction job.
  3. The number of rows held out in the dataset's test split.
  4. The training budget in node hours set when the job is started. Correct answer.

Explanation

AutoML training jobs expose a training budget expressed in node hours, and Vertex AI stops training when that budget is consumed. The budget therefore bounds both cost and training time, at the risk of a less converged model.

Question 30Architect low-code AI solutions

A Gemini-based assistant must answer "where is my order?" using an internal orders API without inventing tracking data. Which approach should you use?

  1. Register the orders API as a function (tool) that the model can call. Correct answer.
  2. Raise the temperature so the model produces more varied answers.
  3. Fine-tune the model on a one-time export of the current orders table.
  4. Add the full orders table to every prompt as context.

Explanation

Function calling lets the model emit a structured request for a declared tool; your code calls the orders API and returns the real result for the model to phrase. The answer is then grounded in live data rather than in the model's prior.

Question 31Architect low-code AI solutions

A retailer must score 20 million stored customer records each night and write the scores to BigQuery. Nobody needs interactive responses. What is the most appropriate approach?

  1. Export the model to a notebook and score records in small chunks.
  2. Use model monitoring to compute the scores on a schedule.
  3. Deploy the model to an endpoint and send 20 million online prediction requests.
  4. Run a batch prediction job with BigQuery as both input and output. Correct answer.

Explanation

Batch prediction is built for large offline scoring jobs and can read from and write to BigQuery directly, with no endpoint to keep warm. It is cheaper and simpler than online serving when latency does not matter.

Question 32Architect low-code AI solutions

After deploying an AutoML tabular model, you want an alert when incoming feature values drift away from the training distribution. What should you configure?

  1. Vertex AI Model Monitoring with the training data as the baseline. Correct answer.
  2. Model Registry version aliases on the deployed model.
  3. Vertex AI Pipelines with caching turned on.
  4. The AutoML training budget in node hours.

Explanation

Model Monitoring compares prediction traffic against a baseline, typically the training dataset, and raises alerts through Cloud Monitoring when drift or skew exceeds thresholds. This gives early warning that the deployed model no longer matches production inputs.

Question 33Architect low-code AI solutions

A Vertex AI pipeline runs nightly, but on most nights only the final step's input changes while an expensive training step reruns anyway. What is the most direct fix?

  1. Change the schedule so the pipeline runs less frequently.
  2. Enable task caching so unchanged tasks reuse earlier artifacts. Correct answer.
  3. Reduce the training budget to a single node hour.
  4. Increase the number of worker replicas so retraining completes faster.

Explanation

Vertex AI Pipelines can cache task executions and reuse the previously produced artifacts when the component code and inputs are unchanged. Because the training step's inputs did not change, caching skips it while the changed downstream step still runs.

Question 34Architect low-code AI solutions

A regulated team needs, for every online prediction from an AutoML tabular model, a ranked contribution showing which input features pushed the score up or down. What should you enable?

  1. An increased training budget so every feature is logged.
  2. Model monitoring with skew and drift detection on the endpoint.
  3. Feature attributions (explanations) enabled on the model's endpoint. Correct answer.
  4. Batch prediction writing results to a BigQuery output table.

Explanation

Vertex AI can attach explanations to a model's endpoint so that each online prediction response includes per-feature attributions, computed with a method such as Sampled Shapley. That satisfies per-decision transparency without custom scoring code.

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

A data science team must be able to submit Vertex AI training jobs and deploy models in a shared project, but it must not be able to change project IAM policies. Which predefined role should be granted?

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

Explanation

Predefined Vertex AI roles follow least privilege: roles/aiplatform.user covers submitting jobs, creating endpoints, and using Vertex AI resources without granting control over IAM policy bindings. Changing who can access a project is governed by resource-manager IAM roles, which are separate from the ML workload role.

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

Analysts in project B need to query a curated subset of tables that live in a dataset in project A, and they must not be able to read the underlying tables directly. What should the data owner in project A do?

  1. Grant the analysts BigQuery Data Owner on the source dataset in project A.
  2. Create a view in a dataset in project A, authorize it on the source dataset, and grant the analysts access to the view. Correct answer.
  3. Schedule a nightly job that copies the source tables into the analysts' project.
  4. Grant the analysts BigQuery Data Viewer on the source dataset in project A.

Explanation

An authorized view runs with the view owner's permissions against the source dataset, so consumers can query the view's output without receiving any grant on the base tables. Access is then granted on the dataset or view that holds the view definition.

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

Several teams build custom training containers and need one private, versioned location from which Vertex AI custom jobs can pull the images. Which service should host the images?

  1. Artifact Registry Correct answer.
  2. The Vertex AI Model Registry
  3. A Cloud Storage bucket with uniform bucket-level access
  4. A shared BigQuery dataset

Explanation

Artifact Registry stores container images and supports per-repository IAM, vulnerability scanning, and image digests, which makes it the standard source for custom containers used by Vertex AI training and prediction.

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

Which Vertex AI resource gives teams a central catalog where they can register model versions, attach aliases, and deploy a chosen version to an endpoint?

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

Explanation

The Vertex AI Model Registry is the central catalog for model versions; it stores versioned model resources with aliases and metadata and links them to endpoints and evaluations.

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

A Vertex AI Pipelines workflow must read and write Cloud Storage data without depending on any individual's personal Google identity. What should the team configure?

  1. The Compute Engine default service account with the Editor role
  2. A dedicated user-managed service account attached to the pipeline run Correct answer.
  3. A JSON key file for the project owner stored inside the pipeline definition
  4. The personal OAuth credentials of the pipeline's original author

Explanation

Attaching a dedicated user-managed service account to the pipeline run decouples the workflow from any person, keeps credentials short-lived, and produces auditable logs of which workload touched which data.

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

A shared BigQuery dataset must be encrypted with a customer-managed encryption key (CMEK). Which grant lets BigQuery encrypt and decrypt data with that Cloud KMS key?

  1. roles/cloudkms.viewer on the project, to every dataset reader
  2. roles/bigquery.dataOwner on the dataset, to the key ring
  3. roles/cloudkms.cryptoKeyEncrypterDecrypter on the key, to the BigQuery service agent Correct answer.
  4. roles/cloudkms.admin on the key ring, to the dataset owner

Explanation

CMEK requires the Google-managed BigQuery service agent to wrap and unwrap data keys, which needs the cryptoKeyEncrypterDecrypter role on the specific Cloud KMS key. The key must also reside in a location compatible with the dataset.

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

Which Vertex AI service records artifacts, executions, and their relationships so teams can trace how a particular model version was produced?

  1. Vertex ML Metadata Correct answer.
  2. Vertex AI Vizier
  3. Vertex AI Feature Store
  4. Vertex AI Model Monitoring

Explanation

Vertex ML Metadata keeps a lineage graph of artifacts such as datasets and models plus the executions that created them, which lets teams audit how a model version came to exist.

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

Multiple teams share one Google Cloud project and need Vertex AI costs attributed to each team in billing reports. What should the platform team do?

  1. Rely on the default project-level cost breakdown without further configuration.
  2. Create a separate project per team and reconcile the invoices by hand each month.
  3. Apply consistent resource labels to pipelines, endpoints, and training jobs. Correct answer.
  4. Record the team name in a free-text description on each job.

Explanation

Labels are propagated into billing exports and the Cloud Billing cost breakdown, so a consistent labelling convention lets finance and engineering attribute spend per team even inside a shared project.

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

A Vertex AI custom training job runs in project B and must read a BigQuery table in project A. What is the minimum sharing configuration?

  1. Give the job's service account BigQuery Data Viewer on the project A dataset and BigQuery Job User in the project where the query runs. Correct answer.
  2. Give the job's service account BigQuery Data Owner on project A so it can manage the shared tables directly.
  3. Copy the required table into project B before each training run.
  4. Give the job's service account BigQuery Admin on both projects and rely on VPC Service Controls to limit the blast radius.

Explanation

Querying a table requires two distinct permissions: reading data in the dataset that owns the table, and creating a query job in the project that executes it. Granting Data Viewer on the source dataset plus Job User in the executing project satisfies both with least privilege.

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

A team retrains weekly, and a downstream team always deploys the current production model. How should the handoff be managed in the Model Registry?

  1. Have the downstream team hard-code the new model version ID each week.
  2. Register each week's model in a new project so resource names never collide.
  3. Email the resource name of the new model to the downstream team.
  4. Upload each approved version and move a shared alias such as 'prod' to it. Correct answer.

Explanation

Aliases are mutable pointers to a model version, so consumers reference a stable name while the owning team advances the pointer after approval. This decouples producer and consumer release cycles.

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

Two teams must query the same BigQuery table, and BigQuery itself must enforce that each team sees only the rows matching its own region. Which feature provides this?

  1. Authorized views in each team's own dataset
  2. Dataset-level IAM grants for each team
  3. Row-level access policies on the shared table Correct answer.
  4. Policy tags applied to the region column

Explanation

Row-level access policies attach a filter predicate to a table, and BigQuery applies it to every query regardless of how the SQL is written. Users with the bypass permission are the only ones who see unfiltered rows.

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

A pipeline component runs under a service account in project B and must write artifacts to a Cloud Storage bucket in project A. Which grant is required?

  1. roles/storage.admin on project B for the pipeline's service account
  2. roles/storage.objectAdmin on the bucket in project A for the pipeline's service account Correct answer.
  3. roles/aiplatform.user on project A for the pipeline's service account
  4. roles/editor on project A for all pipeline developers

Explanation

Cloud Storage permissions are checked on the bucket or object resource, so the writing identity needs an object-writing role granted on that specific bucket, even though the workload runs in another project.

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

A contractor needs access to a shared Vertex AI project for a 30-day engagement. What is the recommended way to grant it?

  1. Create a JSON key for the contractor's service account and delete it later.
  2. Grant the Editor role and rely on a calendar reminder to revoke it.
  3. Grant a predefined role with an IAM condition that sets an expiry timestamp. Correct answer.
  4. Add the contractor to the project's default Compute Engine service account.

Explanation

IAM conditions support time-based expressions, so the binding automatically stops granting access at the expiry time while preserving an audit trail of the grant. The role itself should still be the narrowest predefined role that fits the work.

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

A platform team manages access for 200 engineers who rotate between teams every quarter, and access reviews are slow because permissions are bound to individual users. Which change most improves manageability?

  1. Export IAM policies to BigQuery and audit them on a weekly schedule.
  2. Grant every engineer the Editor role so fewer bindings are needed in total.
  3. Consolidate all resources into a single project so fewer policies exist.
  4. Grant roles to Google Groups and manage group membership instead of individual bindings. Correct answer.

Explanation

Binding roles to groups makes the IAM policy small and stable while membership changes are handled by group owners. Access reviews then reduce to verifying group membership rather than editing resource policies.

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

Several engineers run training with different hyperparameters and want to compare runs side by side using tracked parameters and metrics. Which Vertex AI tool fits this requirement?

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

Explanation

Vertex AI Experiments records parameters, metrics, and artifacts for each run so runs can be compared and the best configuration selected before registering a model.

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

A governance group requires that every registered model document its intended use, limitations, and evaluation results in a standard location. What should the ML team use?

  1. A Vertex ML Metadata artifact created per training run
  2. A README inside the training container image
  3. Labels applied to the prediction endpoint
  4. A Model Card attached to the model version in the Model Registry Correct answer.

Explanation

Model Cards are the standard structured document for intended use, limitations, and evaluation results, and they are versioned alongside the model in the Model Registry so reviewers see the documentation that matches the deployed version.