Professional Machine Learning Engineer · Sample questions
10 real questions from the CertifyDeck Professional Machine Learning Engineer bank, spread across the domains of the official Google Cloud Certified Professional Machine Learning Engineer skills outline, with the correct answer and a full explanation of why the other options are wrong. They are original questions written from the public syllabus, not exam dumps, and they are the same style you will find in the practice tests and mock exams.
Correct answer: A. Call ML.TRANSFORM with the model and a query that selects the affected customers' raw input columns.
ML.TRANSFORM applies the transformations captured in a model's TRANSFORM clause, with the statistics calculated during training, and returns the preprocessed data, which is exactly what the model sees at serving. Copying the expressions into a new query would recompute analytic statistics over the selected rows instead of reusing the training ones. ML.EXPLAIN_PREDICT and ML.FEATURE_IMPORTANCE return attributions or importance scores, not the transformed feature values.
Correct answer: A. Layout Parser, to extract text, tables and lists and return context-aware chunks for retrieval.
Layout Parser extracts document elements such as text, tables and lists and creates context-aware chunks intended for information retrieval in generative AI applications, which is exactly the chunking problem described. Enterprise Document OCR returns text and layout but leaves the chunking to you, so fixed-size splitting would keep cutting tables. Form Parser targets key-value pairs and tables in structured forms, and Custom Splitter separates a file into documents rather than producing retrieval chunks.
Correct answer: C. Lengthen the schedule interval to match how slowly the data changes and keep Model Monitoring alerts for shifts.
Google's guidance is that retraining frequency depends on how often the data patterns change and how expensive retraining is, and that you should consider how much your data changes when choosing the cadence; monitoring alerts cover the case where the data does shift. Stopping retraining permanently gives up adapting to change. Training on one day of data or on bigger machines keeps the daily runs that add no value.
Correct answer: C. Create the run as a training pipeline with the SDK's CustomTrainingJob class and set model_serving_container_image_uri.
A training pipeline (CustomTrainingJob) runs the custom job and can import the exported artifacts into Agent Platform as a Model resource, using the serving container you specify. The documentation states that training code cannot export artifacts in a way that directly creates a Model resource, so a plain CustomJob does not register anything, and a hyperparameter tuning job uploads a model only when it is itself wrapped in a training pipeline.
Correct answer: B. Use the Agent Platform SDK for Python to create a CustomJob from the local script with the prebuilt scikit-learn container URI.
A single Python file run with the SDK as a CustomJob is packaged and installed into a prebuilt container, which the serverless training overview describes as suitable for prototyping and as the low-effort option. A custom container is rated high setup effort, GKE with Kubeflow Trainer adds cluster operations, and AutoML would discard the code the data scientist already wrote.
Correct answer: A. Call aiplatform.init with the experiment name, then aiplatform.autolog() before training.
Autologging in the SDK automatically logs parameters and metrics from training runs to Experiments and supports frameworks that include scikit-learn and XGBoost, creating experiment runs for you if you do not name them. Writing TensorBoard logs or custom ML Metadata records means more code for each run. A self-managed MLflow server adds infrastructure to run, while autologging already uses MLflow's autologging inside the managed service.
Correct answer: C. Undeploy the model from the endpoint, then delete the endpoint and keep the model in Model Registry.
For open models in Model Garden you are charged for the compute resources used to deploy the model to an endpoint, so you undeploy the model and then delete the endpoint; the endpoint can only be deleted after all models are undeployed. Deleting the endpoint first is not allowed while a model is deployed, and deleting the model from Model Registry loses the model the team wants to keep. The runtime template affects notebook runtimes, not the endpoint.
Correct answer: D. Create a batch inference job for the Model with the BigQuery table as the input source and a BigQuery table as the output location.
Batch inference is an asynchronous request (a BatchPredictionJob) sent directly to a Model that isn't deployed to an endpoint, meant for accumulated data when you don't need an immediate response, and it reads from and writes to BigQuery or Cloud Storage. Deploying to an endpoint associates compute resources with the model all day, which is what you want to avoid. A local endpoint is meant for development and testing, and a Cloud Run service called row by row adds packaging and orchestration work for a job batch inference already handles.
Correct answer: D. Save the booster with save_model as model.bst in Cloud Storage, and import it with the prebuilt XGBoost container for that version.
Prebuilt containers provide HTTP inference servers that serve with minimal configuration; for XGBoost the artifact must be model.bst saved with Booster.save_model (or model.joblib), and you create the Model with the prebuilt container matching the training version. A custom container works but means building and maintaining an image, which the stem rules out. The .mar archive is the format for the prebuilt PyTorch container, and the SavedModel format is for TensorFlow.
Correct answer: A. A computation-based metric such as ROUGE that compares each response with the lawyers' reference summary.
Computation-based metrics evaluate responses with deterministic algorithms such as ROUGE or BLEU and are meant for cases where ground truth is available, as the reference summaries are here. Rubric-based and pairwise metrics rely on an LLM judge, and the judge model can show randomness; multi-sampling improves consistency but does not make the score deterministic, so it fails the compliance requirement.
Answer each question before opening the explanation, and write down why you discarded the other options. On Professional Machine Learning Engineer, most mistakes come from choosing an option that is technically possible but does not meet every requirement in the scenario.
These 10 questions are a taste of the format, not a diagnostic. To know which domains you fail, practice by domain for free and take a timed mock exam that mirrors the official weights and duration.
Start for free, no sign-up or card. Every question includes an explanation like the ones above, and the mock exam reproduces the official Professional Machine Learning Engineer format.