GCP Professional ML Engineer practice questions and exam guide

A team uses BigQuery ML CREATE MODEL but wants to call PaLM 2 for text summarization. Correct construct?

1. ML.GENERATE_TEXT over a remote model wrapping the LLM
2. CREATE MODEL with model_type='LINEAR_REG' over the text
3. A Cloud Function that calls the LLM API row by row
4. Not possible; BigQuery ML cannot reach an LLM endpoint

Answer: A — ML.GENERATE_TEXT over a remote model wrapping the LLM

A) Correct — a remote model registers the hosted LLM endpoint inside BigQuery, and ML.GENERATE_TEXT then invokes it from SQL over a table. B) A numeric regression type cannot summarize text. C) Workable but leaves SQL entirely and loses set-based processing. D) Remote models exist precisely for this.

A SQL-only analyst wants k-means clustering over a customer table. Best approach?

1. BigQuery ML CREATE MODEL with model_type='KMEANS'
2. Custom training running scikit-learn KMeans code
3. The AutoML image classification training objective
4. A Dataproc cluster running Spark MLlib KMeans

Answer: A — BigQuery ML CREATE MODEL with model_type='KMEANS'

A) Correct — K-means is a native BigQuery ML model type, trained and scored with SQL over the customer table. B) Requires Python the analyst does not write. C) Wrong modality entirely. D) A cluster is far more machinery than a single SQL statement needs.

BigQuery ML is BEST suited for:

1. Computer vision over labeled image datasets at scale
2. Modeling data already resident in BigQuery using SQL
3. Reinforcement learning with an environment feedback loop
4. Pre-training large language models from scratch on text

Answer: B — Modeling data already resident in BigQuery using SQL

A) Image work belongs in AutoML or custom training, not SQL. B) Correct — BigQuery ML runs regression, classification, clustering, time series and boosted trees through CREATE MODEL with no data movement. C) Not part of the BigQuery ML model catalog. D) Remote models can call an LLM, but BigQuery ML does not pre-train one.

An engineer is choosing a foundation model for a task that summarizes hour-long meeting recordings into action items. The MOST appropriate selection is:

1. Imagen, because it handles multimodal input and can render summary graphics
2. A long-context multimodal Gemini model selected from Model Garden today
3. Veo, because it is the model family designed for long-form video and audio
4. A speech-to-text call followed by a small encoder-only classification model

Answer: B — A long-context multimodal Gemini model selected from Model Garden today

B) Correct — the task needs long-context understanding of audio plus reasoning to produce action items, which is what a long-context multimodal Gemini model provides in one step. A) Imagen generates images; it is not a summarization model. C) Veo is a video generation family, not an analysis or summarization one — a common mix-up. D) Transcription plus an encoder classifier can label utterances but cannot synthesize free-form action items.

Vertex AI Vector Search is appropriate for:

1. Replacing relational databases for OLTP work
2. Approximate nearest-neighbor search over vectors
3. Computing gradients during distributed training
4. Labeling images and video for supervised datasets

Answer: B — Approximate nearest-neighbor search over vectors

A) It stores vectors, not transactional rows. B) Correct — the managed vector index serves low-latency ANN lookups at scale for retrieval, recommendation and similarity. C) Gradient computation happens in the training job. D) Labeling is a separate data service.

Which change to a prompt is LEAST likely to improve output consistency for a structured extraction task?

1. Raising the sampling temperature so the model explores more phrasings
2. Stating the required output schema explicitly and demanding it be followed
3. Supplying two or three worked examples of input and desired output
4. Moving the task instruction ahead of the long document to be processed

Answer: A — Raising the sampling temperature so the model explores more phrasings

A) Correct — higher temperature increases variability, which is the opposite of what a structured extraction task needs; near-deterministic decoding is the right setting. B) An explicit schema constrains the output shape and is standard practice for extraction. C) Few-shot examples are among the most reliable consistency levers. D) Instruction placement matters on long inputs, and leading with the task reliably improves adherence.

Cost-attribution across ML teams in one project. Best practice?

1. Read the single project bill and split it evenly across the teams
2. Label jobs, endpoints, and datasets; aggregate the billing export
3. Have each team submit manual invoices for the resources it used
4. Disable shared resources so each team must create its own project

Answer: B — Label jobs, endpoints, and datasets; aggregate the billing export

A) Opaque — an even split hides real usage. B) Correct — label-based chargeback over the billing export is the standard attribution method. C) Manual and error-prone. D) Not viable operationally.

Features used in both training and online serving must be:

1. Computed by separate training and serving code paths
2. Stored in the managed feature store for both paths
3. Recomputed from raw source on every prediction request
4. Hardcoded as constants inside the serving application

Answer: B — Stored in the managed feature store for both paths

A) Two code paths drift apart and produce training-serving skew. B) Correct — the managed feature store serves one definition to both offline training reads and online lookups, with point-in-time correctness. C) Recomputation adds serving latency and still risks a different formula. D) Constants cannot track changing entity state.

Which artifact relationship does ML Metadata lineage MOST directly let an auditor answer?

1. Which dataset version and training run produced the model now in production
2. Whether the production model's predictions have drifted in the last 30 days
3. How much each input feature contributed to a specific prediction
4. Which IAM principals invoked the prediction endpoint last quarter

Answer: A — Which dataset version and training run produced the model now in production

A) Correct — lineage links artifacts and executions, so tracing a deployed model back to its dataset and run is its core function. B) Drift is answered by model monitoring, which watches serving distributions rather than provenance. C) Per-prediction attribution comes from feature attribution and explainability tooling. D) Caller identity lives in audit logs, an entirely different system.

A team wants to train a custom TensorFlow model with no infrastructure management. Best option?

1. Self-managed Compute Engine VMs with hand-installed CUDA
2. Managed custom training in a pre-built TF container image
3. Cloud Run containers autoscaled for stateless HTTP services
4. App Engine standard runtime for hosting web applications

Answer: B — Managed custom training in a pre-built TF container image

A) Puts VM provisioning, driver installs and patching on the team. B) Correct — the managed custom training service runs a pre-built TensorFlow container on compute it provisions and tears down for you. C) Serverless request serving, not a training service. D) A web application platform with no training primitives.

A team's tuning budget allows only 15 trials for a model with 9 hyperparameters. The MOST practical approach is to:

1. Fix the least influential parameters and tune the top three
2. Tune all nine parameters simultaneously using random search
3. Tune each of the nine parameters in isolation, three trials each
4. Skip tuning and use the framework defaults for all parameters

Answer: A — Fix the least influential parameters and tune the top three

A) Correct — with a tiny budget, reducing the search dimension to the parameters with the largest effect is the standard practical move. B) Fifteen draws over nine dimensions samples the space far too thinly to be informative. C) One-at-a-time tuning ignores interactions and spends the budget on a method known to be weak. D) Defaults are a reasonable baseline but forgo the budget entirely rather than using it well.

Reduction Server best applies to:

1. Single-machine training, where it removes the input bottleneck
2. Inference latency, by batching requests before the forward pass
3. Multi-worker GPU training, by accelerating the gradient all-reduce
4. Storage cost, by compressing checkpoints written to Cloud Storage

Answer: C — Multi-worker GPU training, by accelerating the gradient all-reduce

A) No benefit — there is nothing to reduce across workers. C) Correct — it speeds up gradient aggregation across many GPU workers. B) Unrelated to serving. D) Unrelated to storage.

A model registry holds versions 1 through 6. Version 6 is deployed to 100% of traffic and version 5 remains deployed at 0%. What is the MAIN benefit of leaving version 5 deployed?

1. It keeps the registry lineage record complete for auditors
2. It allows rollback by changing traffic weights, with no redeploy wait
3. It reduces the endpoint's overall cost because traffic is concentrated
4. It ensures monitoring baselines are computed from both versions

Answer: B — It allows rollback by changing traffic weights, with no redeploy wait

B) Correct — a deployed-but-unweighted version is warm capacity for instant rollback, which is the operational reason teams accept its cost. A) Lineage lives in the registry independently of whether a version is deployed. C) Keeping a second deployment running costs more, not less. D) Monitoring baselines derive from configured reference data, not from idle deployments.

An endpoint autoscales on a CPU utilization target, but the model is accelerator-bound and CPU stays near 20% while responses queue. The BEST correction is:

1. Lower the CPU target to about 10% so scaling triggers earlier
2. Scale on accelerator duty cycle or request concurrency instead
3. Pin the replica count to the observed peak and disable autoscaling
4. Move preprocessing onto the accelerator so CPU utilization rises

Answer: B — Scale on accelerator duty cycle or request concurrency instead

B) Correct — an autoscaler must key on the resource that is actually saturated, so switching the signal to accelerator duty cycle or concurrency makes scaling track real pressure. A) Tuning a CPU target is treating a metric that is not the bottleneck; the correlation is coincidental and fragile. C) Fixed peak provisioning works but pays peak cost permanently and is a retreat rather than a fix. D) Deliberately inflating CPU to game the autoscaler is a hack that couples unrelated concerns.

Push vs pull serving — direction reversal: which is appropriate for a streaming personalization use case where features change every second?

1. Pull the features from the feature store's offline store per request
2. Serve without features and rely on the model's popularity prior
3. Pull the features from a CSV file refreshed once each night
4. Push precomputed features to the online store for low-latency reads

Answer: D — Push precomputed features to the online store for low-latency reads

A) The offline store is built for training reads, not request-time latency. D) Correct — push fresh values into the online store and serve them at request time. C) Stale by a day. B) Discards the personalization signal.

A team must build the pipeline's container images automatically whenever component code changes, then push them to a registry for the pipeline to consume. The MOST appropriate managed service is:

1. The pipelines service itself, using a build component step
2. A managed build service triggered by the source repository
3. A workflow orchestrator scheduled to run a build script hourly
4. A notebook instance running the build manually on demand

Answer: B — A managed build service triggered by the source repository

B) Correct — repository-triggered managed builds are purpose-built for turning source changes into versioned images with auditable provenance. A) Building images inside the pipeline couples artifact creation to run time and rebuilds on every execution. C) Hourly scheduled builds are both late and wasteful compared with reacting to a commit. D) Notebook-driven builds are manual, unversioned, and unrepeatable.

Continuous Training (CT) vs Continuous Integration (CI) vs Continuous Delivery (CD) — which belongs to MLOps Level 2?

1. All three: CI builds and tests, CD deploys pipelines and models, CT retrains
2. Only CT, because Level 2 is defined solely by automated continuous retraining
3. Only CD, because Level 2 is defined by automated delivery of models to serving
4. None of them, since Level 2 describes manual experimentation with scripts

Answer: A — All three: CI builds and tests, CD deploys pipelines and models, CT retrains

A) The controlling idea is that Level 2 is full automation: the pipeline itself is built and tested by CI, delivered by CD, and executes continuous training on new data. B tempts because CT is the headline capability, but CT alone already arrives at Level 1. C tempts because delivery automation is visible, yet it omits both the build and the retraining halves. D tempts by describing Level 0 and contradicts the premise of the maturity ladder.

A pipeline must retrain only when data drift is detected. Best trigger?

1. A fixed daily schedule that retrains whether or not the data has shifted
2. No trigger at all, relying on the endpoint to adapt as traffic changes
3. A manual retraining request filed after an analyst reviews weekly reports
4. A monitoring alert to Pub/Sub, then Cloud Run, which submits a pipeline

Answer: D — A monitoring alert to Pub/Sub, then Cloud Run, which submits a pipeline

D) The controlling idea is event-driven retraining: the monitoring job detects drift, publishes an alert, and a lightweight service submits the training pipeline run. A tempts because schedules are automation, but a fixed cadence both misses drift between runs and wastes compute when nothing changed. C tempts because human review adds judgment, yet it is not a trigger and does not meet the requirement. B tempts only if one assumes models self-correct, which they do not.

Vertex AI Model Monitoring detects:

1. Network outages affecting the serving region network
2. Feature, prediction, and concept drift in serving
3. Application code bugs and unhandled exceptions
4. Cost overruns against the monthly billing budget

Answer: B — Feature, prediction, and concept drift in serving

A) Infrastructure availability is a Cloud Monitoring concern. B) Correct — the managed model monitoring service compares live input and output distributions against a training baseline and alerts when they diverge. C) Error Reporting surfaces exceptions. D) Budgets and alerts live in billing.

Half-right monitoring: team monitors only top features and ignores rare-but-critical features. Risk?

1. No real risk, because the top features dominate the model's predictions anyway
2. The monitoring pipeline runs faster, the intended tradeoff of narrow coverage
3. Rare high-impact features can shift undetected; prioritize by importance
4. Monitoring costs fall, and cost is the main consideration when scoping coverage

Answer: C — Rare high-impact features can shift undetected; prioritize by importance

C) The controlling idea is that frequency and importance are different axes: a feature that appears rarely can still drive the decisions that matter most, so coverage should follow attribution-based importance. A tempts because top features usually do dominate on average, but averages hide the critical tail. B tempts because narrower monitoring is genuinely faster, yet speed is not the risk being asked about. D tempts for the same reason, trading a real blind spot for a small saving.

Half-right monitoring: team monitors data drift but not model performance (accuracy, AUC, etc.). Risk?

1. Drift can be benign or severe; without ground truth you cannot tell either way
2. No real risk, since drift large enough to matter shows up in the input monitors
3. Training runs finish faster because fewer evaluation metrics are computed
4. Serving costs fall, since performance metrics are the costly part of monitoring

Answer: A — Drift can be benign or severe; without ground truth you cannot tell either way

A) The controlling idea is that input drift is a leading indicator, not a verdict: only labelled outcomes reveal whether accuracy or AUC actually degraded. B tempts because drift magnitude feels informative, but large shifts can be harmless and small ones can be fatal. C tempts because skipping metrics is faster, yet training duration is unrelated to production monitoring. D tempts by framing the omission as savings, but the saving is trivial next to the blind spot.

Vertex Explainable AI provides:

1. Automatic optimization of training source code
2. Per-feature attributions for individual predictions
3. Managing and rotating customer-managed encryption keys
4. Cost analysis and forecasting for training jobs

Answer: B — Per-feature attributions for individual predictions

A) No such code-rewriting service exists here. B) Correct — the explainability service returns per-feature attribution, using sampled Shapley or integrated gradients, for a specific prediction. C) That is Cloud KMS. D) That is billing tooling.

Which item is LEAST appropriate to include in a model card intended for downstream consumers?

1. The intended use cases and documented out-of-scope uses
2. Evaluation results broken down by relevant subgroups
3. The service account credentials used by the training pipeline
4. Known limitations and the data the model was trained on

Answer: C — The service account credentials used by the training pipeline

C) Correct as the misfit — credentials are secrets and must never appear in documentation intended for distribution. A) Stating intended and out-of-scope uses is a core purpose of a model card. B) Subgroup results let consumers judge where the model is reliable. D) Limitations and training data provenance are standard and expected content.

Right feature wrong stage: bias monitoring belongs where?

1. During training only, using fairness evaluations across slices of the training set
2. Only at deployment, as a one-off review gate before the model serves traffic
3. Both: slice-level fairness at training and slice-level monitoring after deploy
4. Nowhere, since fairness is a policy matter rather than an engineering concern

Answer: C — Both: slice-level fairness at training and slice-level monitoring after deploy

C) The controlling idea is that bias is not a one-time property: the training-time evaluation establishes a baseline, and post-deployment slice monitoring catches populations that shift after release. A tempts because training is where fairness is usually measured, but a fair model can become unfair as inputs change. B tempts because a release gate feels decisive, yet a single check expires immediately. D tempts by deferring responsibility, which leaves the failure undetected.

Distinguish supervised tuning, distillation, and RAG with grounding as ways to improve a foundation model's answers.

Supervised tuning adapts a model's behavior, style or task format using labeled examples • distillation transfers a larger teacher model's behavior into a smaller cheaper student to cut latency and cost • RAG Engine with grounding injects retrieved authoritative content at inference so answers cite current facts. Use RAG for knowledge freshness and factuality, tuning for behavior and format, distillation for cost and latency.

BigQuery ML CREATE MODEL syntax basics?

CREATE MODEL `project.dataset.model` OPTIONS (model_type='LOGISTIC_REG') AS SELECT … . Use ML.PREDICT() to predict.

When do you use batch prediction instead of an online endpoint?

When predictions are consumed asynchronously — scoring a full table nightly, backfilling scores, or feeding a downstream job — and no per-request latency requirement exists. Batch reads from Cloud Storage or BigQuery, writes results back in bulk, provisions compute only for the run, and costs far less than keeping an endpoint warm.

Compare Workbench instances and Colab Enterprise as managed notebook environments.

Workbench instances give a persistent managed VM you control, suited to long-running work, custom images and larger attached accelerators. Colab Enterprise gives a zero-setup collaborative notebook with governed runtime templates, suited to fast shared exploration. Both run inside the project's network and IAM boundary.

What is Model Registry's role in a deployment workflow?

It is the versioned inventory of trained models: each version carries its evaluation metrics, lineage and aliases, and deployment targets a registry version rather than a raw artifact path. This gives a single approved-model surface for promotion, rollback and governance across environments.

What must be configured to enable model monitoring on a deployed endpoint?

A baseline (the training dataset or a reference window of serving traffic), request-response logging on the endpoint, the features and objectives to monitor, an alert threshold per feature, and a monitoring sampling rate and interval plus alert destinations. Without a stored baseline there is nothing to compare live traffic against.