Google Cloud Machine Learning Engineer Professional

1. 1.A retail analytics team stores three years of daily sales history in BigQuery and wants to project the next 90 days of demand per store. They want to stay in SQL and avoid managing training infrastructure. Which BigQuery ML model type should they use?

   • AARIMA_PLUSCorrect answer
   • BLINEAR_REG
   • CKMEANS
   • DMATRIX_FACTORIZATION

   Why

   ARIMA_PLUS is the BigQuery ML time-series model built for univariate forecasting. It handles seasonality, holidays, and anomaly cleanup automatically, and a single CREATE MODEL statement can fit one series per store using the TIME_SERIES_ID_COL option. LINEAR_REG has no notion of temporal structure, KMEANS is unsupervised clustering, and MATRIX_FACTORIZATION is for recommendations.

2. 2.An analyst has trained a logistic regression model in BigQuery ML and now needs to score a table of 40 million new customer rows, writing the predicted class and probability back to BigQuery. Which approach requires the least additional infrastructure?

   • ARun a query that calls ML.PREDICT on the model and writes the result to a destination tableCorrect answer
   • BExport the model to Cloud Storage and deploy it to an Agent Platform online endpoint
   • CPackage the model in a container and serve it from Cloud Run behind a batch client
   • DCall ML.EVALUATE on the new table to produce the predictions

   Why

   ML.PREDICT runs inference directly inside BigQuery against the stored model, so the scoring job is just a SELECT that can be wrapped in CREATE TABLE AS. No endpoint, container, or pipeline is needed. Exporting the model to Agent Platform or Cloud Run would add serving infrastructure that batch scoring inside the warehouse does not require, and ML.EVALUATE measures quality on labeled data rather than producing predictions.

3. 3.A BigQuery ML practitioner bucketizes a numeric column and one-hot encodes a categorical column before training, then discovers that the same preprocessing has to be repeated by hand every time ML.PREDICT is called. What should they change?

   • ADeclare the feature engineering in the model's TRANSFORM clauseCorrect answer
   • BMaterialize the engineered features in a view and query the view
   • CMove the preprocessing into a scheduled query that runs before each prediction
   • DRewrite the preprocessing as a Dataflow job invoked by ML.PREDICT

   Why

   Preprocessing declared inside the model's TRANSFORM clause is stored with the model and replayed automatically at prediction time, which removes the manual step and eliminates training-serving skew. Creating a view or a scheduled query still leaves the caller responsible for applying the transformations, and moving the work to Dataflow adds a separate system that ML.PREDICT does not invoke.

4. 4.A team with no deep learning experience must build an image classifier for 12 product categories from 30,000 labeled photos. They need a deployable model quickly and cannot write custom training code. Which option best fits?

   • ATrain an image classification model with Agent Platform AutoMLCorrect answer
   • BWrite a custom PyTorch training job and submit it to Agent Platform custom training
   • CCreate a LOGISTIC_REG model in BigQuery ML over the raw image bytes
   • DFine-tune a ResNet on a self-managed GKE cluster with manual hyperparameter sweeps

   Why

   Agent Platform AutoML for image classification handles architecture search, augmentation, and hyperparameter selection from a labeled dataset, and produces a model that can be deployed to an endpoint or exported for edge use, all without training code. Writing a custom PyTorch training job or hand-tuning on GKE contradicts the no-code constraint, and BigQuery ML does not train image classifiers from raw photos.

5. 5.An insurance company receives scanned claim forms as PDFs and needs structured output containing named fields such as policy number, claim date, and total amount. Which Google Cloud API is the most direct fit?

   • ADocument AI APICorrect answer
   • BVision API text detection
   • CNatural Language API entity analysis
   • DTranslate API

   Why

   Document AI is purpose-built for parsing documents into structured entities, and its processors extract key-value pairs and named fields from forms and specialized document types. The Vision API returns raw OCR text and would leave field identification to the caller, the Natural Language API analyzes text semantics rather than document layout, and the Translate API only converts languages.

Practice all 65 questions

Frequently asked questions