Machine Learning Engineer Practice Exam 2025: Latest Questions

You are building a binary classifier to detect fraudulent transactions. Only 0.3% of transactions are fraud. The business wants to minimize missed fraud while keeping false positives manageable for the review team. Which evaluation approach is most appropriate?

Your team wants a serverless way to run scheduled batch inference every night and write predictions to BigQuery. The inference code is packaged as a container. Which Google Cloud service is the best fit?

You have raw event logs in Cloud Storage and want to create a repeatable, scalable transformation into a curated analytics table in BigQuery with minimal operational overhead. The transformation includes windowed aggregations and joins. What should you use?

You need to expose an ML model for online predictions with autoscaling and minimal infrastructure management. The model is trained in Vertex AI and must serve low-latency predictions. What is the recommended approach?

A training dataset includes a feature 'days_since_last_purchase' computed using the entire dataset, including purchases occurring after the label timestamp. The model performs very well in offline evaluation but poorly in production. What is the most likely issue and best fix?

You need to train a large TensorFlow model on GPUs using Vertex AI custom training. The training code must be portable across environments and should not rely on VM-level configuration. What is the best practice?

You operate a Vertex AI endpoint for online predictions. Over time, prediction quality degrades due to changing user behavior. You want to detect distribution shift between training data and recent online prediction inputs, and trigger investigation. Which approach is most appropriate?

You want a reproducible, auditable ML pipeline that trains, evaluates, and conditionally deploys a model if it meets a quality threshold. The pipeline should be versioned and rerunnable, and each step should capture lineage of artifacts. What should you use?

You must provide near-real-time features for online predictions while also supporting offline training with identical feature definitions. Multiple teams will reuse features across models. You need point-in-time correctness for offline training and low-latency retrieval for online serving. Which design best meets these requirements on Google Cloud?

A regulated organization needs to deploy an LLM-based summarization service. Requirements: prevent data exfiltration to public endpoints, keep all traffic private, and restrict who can call the model. They plan to use managed models in Vertex AI. What is the best architecture?

Your team wants to quickly sanity-check whether a new model is learning anything useful before investing in feature engineering. You need a baseline approach that is fast, easy to implement, and provides a meaningful reference score for future iterations. What should you do?

You are building an NLP classifier on Vertex AI. Training data is in Cloud Storage and is updated daily. You want reproducible training runs and to ensure that each run uses an immutable snapshot of the dataset. What is the recommended approach?

A batch scoring job writes predictions to BigQuery each night. Analysts report intermittent duplicates and missing rows. The pipeline uses Dataflow to read from Pub/Sub and write to BigQuery. What change best improves correctness for exactly-once-like outcomes at the sink?

You need to orchestrate a repeatable end-to-end training workflow that includes data extraction, feature preprocessing, model training, evaluation, and conditional model registration only if metrics meet thresholds. Which Google Cloud approach best fits this requirement?

You are serving an online prediction model on Vertex AI. You must meet strict latency SLOs and want to reduce cold-start impact while allowing traffic to scale during peak hours. What should you configure?

Your binary classifier achieves high overall accuracy, but in production it misses a small number of high-impact positive cases. The business wants to minimize false negatives even if false positives increase. What is the most appropriate next step?

You train a model on historical data and observe strong offline performance. After deployment, performance drops significantly. Investigation shows that one key feature was computed using information that is only available after the prediction time (for example, a future aggregate). What is the most likely issue and best fix?

You want to improve model performance by adding a new feature derived from raw clickstream events. The feature must be identical for training and online serving, and it must be updated continuously as new events arrive. Which design best ensures training-serving consistency?

Your team has multiple models deployed. You need to detect data drift and performance regressions, investigate root causes, and set alerts when prediction distributions shift significantly. Which Vertex AI capability should you use?

You are using hyperparameter tuning for a custom training job. Some trials report poor metrics early and waste resources. You want to stop unpromising trials while still exploring the search space effectively. What should you do?

A retail company wants to train a model to predict whether a customer will return an item. Only 2% of orders are returned, and the business cares most about catching true returns while keeping the manual review workload manageable. Which evaluation approach is MOST appropriate?

You are designing a feature store strategy for multiple teams using Vertex AI. The same feature (e.g., "user_7d_purchase_count") must be computed once, reused across training and online inference, and remain point-in-time correct to prevent label leakage. What is the BEST approach?

A team trains tabular models on BigQuery data and wants a fully managed approach with minimal code that supports automatic feature preprocessing, hyperparameter tuning, and easy deployment to an endpoint. Which solution is MOST appropriate?

You operate a Vertex AI endpoint for real-time predictions. After a data pipeline change, prediction quality degrades gradually over a week, but latency and error rate remain normal. You want an automated way to detect this issue early and link it to input data shifts. What should you do?

A regulated healthcare company trains models in one Google Cloud project and serves predictions in a separate project owned by another team. The serving project must be able to invoke the model endpoint, but must NOT be able to list training datasets or access the training project's storage buckets. What is the BEST design?