Domain 1: Architecting Low-Code AI Solutions
- In BigQuery ML, ML.ROC_CURVE returns recall, false positive rate, and the corresponding threshold across a full sweep of thresholds, unlike ML.CONFUSION_MATRIX which reports counts at one fixed threshold.
- Choose the BigQuery ML model type by task: LOGISTIC_REG for binary or multiclass labels (for example churn), LINEAR_REG for continuous targets, KMEANS for unsupervised clustering with no label, ARIMA_PLUS for single time-series forecasting, and DNN_CLASSIFIER for deep non-linear patterns over large labeled data.
- ML.EVALUATE reports quality metrics for the model type, including mean absolute error, mean squared error, and r2_score for a linear regression model.
- BigQuery ML preprocessing functions include ML.STANDARD_SCALER to standardize numeric columns, ML.BUCKETIZE for fixed ranges, and ML.QUANTILE_BUCKETIZE for quantile-based buckets.
- Model Garden provides a catalog of open-source and Google foundation models that can be browsed and deployed to a Vertex AI endpoint with minimal, guided setup.
- Supervised fine-tuning adapts a foundation model's behavior using a curated dataset of input and desired-output pairs.
- Pretrained APIs cover common tasks with no training: Translation API auto-detects source language within the same request, content classification assigns text to predefined topic categories, and entity sentiment analysis attaches a sentiment score to each entity mentioned.
- Video Intelligence features are distinct: shot change detection marks camera-shot boundary timestamps, label detection tags objects and concepts, object tracking follows objects across frames, and text detection reads on-screen writing such as signs and captions.
- Speech-to-Text streaming recognition processes audio continuously for low-latency live captioning, and specialized models such as medical_conversation are tuned for clinical spoken vocabulary.
- AutoML forecasting datasets require a designated time column and a time-series identifier column alongside feature columns, and AutoML Edge image models export to TensorFlow Lite, Core ML, and TensorFlow.js.
- A Document AI processor trained on a document type such as receipts returns structured labeled fields instead of plain OCR text, and batch requests process many stored files as parallel jobs for large backlogs.
- Retail API managed recommendations serve personalized product suggestions without the team building or training a model.
Domain 2: Collaborating to Manage Data and Models
- In Vertex Explainable AI, sampled Shapley works for any model whether differentiable or not, integrated gradients suits differentiable models, and pixel-level integrated gradients is preferred over XRAI for low-contrast images such as X-rays where region segmentation performs poorly.
- A negative feature attribution value means that feature pulled the prediction opposite to the observed outcome, not that the value is missing or an error.
- Feature Store batch ingestion imports a full table or partition of precomputed feature values in bulk on a schedule, matching a nightly ETL job rather than a continuous stream.
- A feature view links a BigQuery table or view source to the online store and re-syncs on a configured recurring schedule; an entity is the object (such as a customer) whose feature values are tracked and served.
- Registered feature groups, cross-project IAM roles, and a synced feature view together let other teams safely discover and consume shared features.
- Dataset versions freeze an immutable snapshot of a non-tabular managed dataset's items and annotations so training pins to it for reproducibility; only text, image, and video (non-tabular) datasets support versions.
- CMEK and IAM are independent controls: CMEK sets which encryption key protects stored data while IAM governs who can access or manage the resource, and both remain necessary.
- CMEK requires the key ring to be regionally aligned with the resources, and the relevant service agent must hold encrypt and decrypt permission on the key.
- VPC Service Controls dry-run mode logs what a perimeter would block without enforcing it, letting teams observe impact before turning on enforcement.
- Model Registry offers integrated model card generation that populates structured documentation from a model's own metadata.
- Vertex ML Metadata records lineage: events link executions to the artifacts they consume or produce, and a context groups related artifacts and executions (such as one experiment) into a queryable set.
- Fairness terms are distinct: label bias is when ground-truth labels reflect systematic human inconsistency, and a proxy feature correlates strongly enough with a removed sensitive attribute to reintroduce its influence.
- Vertex AI TensorBoard is a regional resource used to visualize training curves and compare hyperparameter-tuning trials, so it must be created in a supported region.
- Parquet's columnar layout lets a reader fetch only needed columns to minimize bytes scanned, and APPROX_COUNT_DISTINCT gives a fast approximate cardinality estimate for very large tables.
Domain 3: Scaling Prototypes into ML Models
- scikit-learn's Pipeline API is a lightweight choice for quickly prototyping preprocessing plus a simple classifier on a modest dataset before moving to heavier frameworks.
- Fault-tolerant Spot VM training combines regular checkpointing, catching the preemption notice to save state before shutdown, and configuring automatic resumption from the latest checkpoint.
- Distributed TensorFlow strategies: MirroredStrategy is synchronous all-reduce on one machine, MultiWorkerMirroredStrategy is the same across machines, and ParameterServerStrategy needs dedicated chief, worker, and parameter-server pools.
- Vertex AI automatically injects the TF_CONFIG environment variable into each worker so distributed TensorFlow strategies coordinate without manual cluster configuration.
- As worker count grows, communication overhead can outweigh added compute, causing diminishing or negative returns, so scaling must be validated rather than assumed linear.
- Accelerator selection: NVIDIA H100 gives the highest throughput and interconnect for the largest pretraining jobs, and A100 offers NVLink high-bandwidth GPU-to-GPU communication with greater memory for demanding distributed training.
- Migrating a PyTorch pipeline with custom CUDA kernels to TPUs requires reworking those kernels for a TPU-friendly path, so weigh that engineering cost against expected savings.
- In a tf.data pipeline, caching after a deterministic transform stores its output so later epochs reuse it, which is only safe when the step produces the same output every time.
- Data-format tradeoffs: columnar Parquet lets a reader fetch only requested columns and tensorflow-io's IODataset reads it directly, JPEG bytes save space at a decode cost, and CSV parses text field-by-field with no built-in compression.
- Training code reports each trial's metric to the hyperparameter-tuning service using the cloudml-hypertune library's reporting call.
- In hyperparameter tuning, higher parallelism finishes faster but gives Bayesian optimization less prior feedback per batch, often needing more total trials; a categorical parameter represents an unordered set of named choices such as optimizers.
- Pre-built Vertex AI training containers still require the code packaged as a Python source distribution plus the module to execute; the job YAML defines worker pools and machine types while args are passed through to the script for argparse.
- Requesting a quota increase for the specific region and accelerator type ahead of time avoids delays when the job is ready to run.
Domain 4: Serving and Scaling Models
- Shadow testing runs a candidate model on mirrored production traffic while hiding its predictions from users, ideal for sensitive cases like fraud scoring where an untested model must not reach a customer.
- Standard Vertex AI online prediction does not scale below its configured minimum replica count, so a minReplicaCount of 1 or higher keeps replicas running and billing even when idle.
- Setting minReplicaCount to zero lets an endpoint scale fully down to save cost, but the next request after idle triggers a cold start while a replica provisions and loads the model.
- Batch prediction fits latency-tolerant workloads that avoid idle infrastructure cost, reads inputs from Cloud Storage or BigQuery only, and accepts accelerator_type and accelerator_count directly in the job request.
- Vertex AI batch prediction jobs record errors for individual malformed instances and continue rather than failing the whole job.
- Vertex AI injects AIP_STORAGE_URI so a custom serving container knows the Cloud Storage location of the model artifacts to load, and rawPredict passes the request body directly without the standard instances/parameters wrapper.
- Model optimization techniques: knowledge distillation trains a smaller student to mimic a teacher's output distribution, magnitude-based pruning removes the least important weights, and fine-tuning after pruning recovers accuracy that one-shot pruning would lose.
- A prediction service SLO should specify a target latency percentile and a target success or error rate over a time window, with availability captured as the ratio of successful responses to total requests.
- Deploying endpoints within European regions and routing European customer traffic only to them satisfies a data-residency constraint.
- A load balancer with health checks can detect an unhealthy regional endpoint and route new requests to a healthy region instead.
- Placing a caching layer keyed on a hash of the input in front of an endpoint lets identical repeated requests be served from cache instead of recomputed.
- An endpoint cannot be deleted while models are still deployed to it, so every deployed model must be undeployed first, and functionally unrelated models that must not share scaling or routing belong on separate endpoints.
- Requesting a regional quota increase in advance is the direct fix when online prediction requests are throttled by a regional quota.
Domain 5: Automating and Orchestrating ML Pipelines
- Vertex AI Pipelines components do not retry automatically; a retry policy with a retry count and backoff must be explicitly configured on a task.
- Pipeline caching is evaluated per component from its own inputs and definition, so changing one component's inputs re-executes only it and its downstream steps while unchanged upstream components reuse cached outputs.
- Point-in-time correctness matters: joining labels to present-day feature values instead of their historical state leaks future information and inflates offline metrics unrealistically.
- A Kubeflow exit handler wraps the pipeline's tasks so a designated component always runs once the pipeline finishes, whether it succeeded or failed, ideal for a single notification or cleanup step.
- Component authoring: a typing.NamedTuple return type exposes several named outputs, base_image selects the starting image, packages_to_install adds pip dependencies at runtime, and dsl.importer registers an existing external file as a typed artifact.
- Artifacts are written under pipeline_root nested by run identifier and then by producing task name, and publishing versioned components to a shared Artifact Registry repository lets other teams reuse them.
- TFX Transform produces a transform graph of the preprocessing logic, both pre-transform and post-transform statistics artifacts, and the transformed examples, letting engineers compare feature statistics before and after.
- Event-driven orchestration: an Eventarc trigger on a Pub/Sub message can start a Cloud Workflows execution immediately, avoiding scheduled polling delay, and Workflows has built-in connectors for common Google Cloud APIs.
- Test types map to stages: data validation checks incoming data against a known-good schema and baseline, an integration test runs the full pipeline in staging on representative data, and a load test checks concurrency under a latency ceiling.
- CI/CD maturity progresses from automated retraining with manual pipeline updates, up to the most mature stage that automatically builds, tests, deploys, and validates pipeline changes and can trigger retraining without manual steps.
- Retraining triggers differ: drift-driven retraining starts automatically on a monitoring-detected distribution shift, and automated rollback reverts to the previous model version when an error-rate spike is detected after deployment.
- When the input schema or architecture has changed significantly, training from scratch is preferred over warm starting because prior weights no longer align, and distribution skew is when the serving dataset's overall distribution differs from the training set's.
Domain 6: Monitoring AI Solutions
- Vertex AI Model Monitoring is built for structured tabular features, depends on request-response logging to capture serving data, and adds storage and compute costs that scale with usage.
- Model Monitoring compares numerical feature distributions using Jensen-Shannon divergence and categorical feature distributions using L-infinity distance.
- Batch prediction monitoring is tied to the specific job's input data at execution time rather than running as a continuous background schedule like online-endpoint monitoring.
- Audit log types differ: Admin Activity logs (always on, cannot be disabled) show who performed a deployment, while Data Access logs, once enabled, show who has been invoking predictions.
- Focus monitoring on the subset of business-critical features rather than every column at maximum sensitivity, which keeps signal meaningful and controls cost and alert noise.
- Under severe class imbalance, the precision-recall curve is far more sensitive to minority-class performance than a ROC curve.
- Fast proxy labels, such as early delinquency signals or expedited manual sampling, provide a usable feedback loop to track live model quality well before slow ground truth completes.
- Output guardrails enforce and filter unsafe responses at serving time while safety-attribute monitoring tracks trends and volumes of unsafe content over time; the two are complementary, not interchangeable.
- When a skew or accuracy drop coincides exactly with a schema change, fix the upstream ingestion or schema issue first, since retraining on the corrupted data would only teach the model on flawed inputs.
- Confirmed genuine drift plus ample fresh labeled data are the conditions that make retraining the right response, whereas a corrupted pipeline is not.
- Operational metrics matter: error_count records failed online prediction requests as the basis for a failure rate, and replica count relative to maximum plus per-replica accelerator utilization together reveal scaling headroom.
- Cost governance uses Cloud Billing budgets with threshold rules that notify at a spend percentage, plus the detailed usage cost export to BigQuery for SKU- and label-level cost queries.
- Model Monitoring alerts route through Cloud Monitoring notification channels, which include integrations such as PagerDuty, and each distributed-training log entry carries worker-pool and task-index labels for per-machine filtering.
Google Cloud Professional Machine Learning Engineer exam tips
- Learn the BigQuery ML model-type-to-task mapping cold (LOGISTIC_REG, LINEAR_REG, KMEANS, ARIMA_PLUS, DNN_CLASSIFIER, BOOSTED_TREE), since many low-code questions hinge on picking the one model that fits the stated label and data shape.
- For any low-code scenario, default to the highest level tool that satisfies the requirement: pretrained API first, then AutoML or BigQuery ML, then custom training, and reach for custom code only when nothing simpler fits.
- Memorize the online-versus-batch prediction tradeoffs, including that standard online endpoints never scale below their minimum replica count while batch jobs run to completion and avoid idle cost.
- Distinguish the monitoring distance metrics (Jensen-Shannon for numerical, L-infinity for categorical) and the drift-versus-skew-versus-corrupted-pipeline decision, since the correct fix differs for each.
- When two options seem valid, favor the managed, event-driven, or reproducible choice (Eventarc over polling, dataset versions for reproducibility, dedicated service accounts over broad credentials), which matches Google's recommended patterns.
Study guide FAQ
How long is the exam and what score do I need to pass?
The exam runs 120 minutes with 767 questions in this bank's sampling, and Google does not publish a fixed passing score (results are pass/fail). It is scenario-heavy, so expect multi-sentence situational questions rather than simple recall.
How much Vertex AI knowledge does this exam assume?
A great deal. Vertex AI is the backbone across training, Feature Store, Model Registry, endpoints, batch prediction, Pipelines, and Model Monitoring, so you should be comfortable with each service's role and configuration options.
Do I need to write or debug real ML code for the exam?
You will not write code, but you must recognize correct usage of tools like BigQuery ML SQL functions, tf.data pipelines, distribution strategies, cloudml-hypertune, and Kubeflow component definitions well enough to choose the right one.
How much of the exam is low-code versus custom training?
Both are heavily tested. You need to know when a pretrained API, AutoML, BigQuery ML, or Model Garden solves a problem without custom code, and separately how to scale a custom prototype with GPUs, TPUs, distributed strategies, and hyperparameter tuning.