50 Machine Learning Engineer Practice Questions: Question Bank 2025

A retailer wants to predict whether a customer will churn in the next 30 days. Only 2% of customers churn, and the business cares more about catching churners than avoiding false alarms. Which evaluation metric should the ML engineer prioritize?

You need a managed feature store to share and reuse features between training and online prediction, with consistent feature definitions and low-latency retrieval. Which Google Cloud product best fits this requirement?

A data engineering team stores raw clickstream logs in Cloud Storage as JSON lines. The ML team needs to create a clean, partitioned analytical dataset for training, and the transformation should be repeatable and scalable. What is the recommended approach?

Your tabular classification model shows excellent training performance but significantly worse validation performance. Which issue is the most likely cause?

A bank must train a fraud model on sensitive data. Training is performed on Vertex AI, but data must never traverse the public internet. The training pipeline reads from BigQuery and writes artifacts to Cloud Storage. What architecture best meets this requirement?

Your training data is time-ordered (e.g., daily sales). You need a validation strategy that best estimates future performance and avoids leaking future information into training. What should you do?

A team uses BigQuery for training data. They frequently see training/serving skew because some features are computed differently in batch training than in online predictions. What is the best practice to reduce this skew?

You are training a model on Vertex AI and want to track datasets, code versions, parameters, and resulting metrics across experiments so results are reproducible and comparable. Which approach is most appropriate?

A Vertex AI endpoint serves an online prediction model. After a new model version rollout, prediction latency increases and P95 exceeds the SLO, but CPU utilization on the deployed replicas is low. Which action is most likely to reduce latency?

You run a Kubeflow Pipelines (Vertex AI Pipelines) workflow nightly. Some pipeline steps are expensive but deterministic given the same inputs (e.g., feature generation from a fixed snapshot). You want to reduce cost and runtime without sacrificing reproducibility. What should you do?

You are building a Vertex AI tabular model for customer churn. Business stakeholders need to understand why individual predictions are made to satisfy internal audit requirements. What is the most appropriate approach on Google Cloud?

A team has images stored in Cloud Storage and wants a minimal-code pipeline to classify them, including managed training and hyperparameter tuning, without building custom training infrastructure. Which approach best fits?

Your organization must ensure that training and prediction traffic never traverses the public internet. Which configuration best satisfies this requirement for Vertex AI training jobs and online predictions?

You need to build a training dataset by joining clickstream events (high volume) with user profile data. The dataset must be reproducible for audits, and feature computation should be consistent between training and serving. What is the best approach?

A Vertex AI endpoint is showing increased latency and intermittent 5xx errors shortly after a new model version was deployed. You suspect the container is failing health checks due to slow initialization. What should you do first?

You are designing a system to retrain a model weekly when new labeled data arrives in BigQuery. The pipeline must be repeatable, parameterized by a training date, and track lineage between data, code, and model artifacts. What is the best solution on Google Cloud?

A team is training an NLP classifier on imbalanced data where the positive class is rare. The current model has high accuracy but misses most positives. Which metric and approach best address the problem?

Your training jobs frequently fail because a subset of TFRecord files in Cloud Storage are corrupted. You want the pipeline to automatically quarantine bad files, continue processing valid data, and produce an auditable report of the skipped inputs. What is the best design?

You serve a recommendation model on Vertex AI. Recent monitoring shows a statistically significant shift in the distribution of several categorical features (data drift), but model performance metrics are not immediately available because labels arrive weeks later. What should you do to reduce business risk while maintaining a reliable workflow?

A regulated company must ensure that only approved models can be deployed to production endpoints. They need a controlled promotion process with evidence of evaluation results, sign-off, and an immutable record of what was deployed. Which solution best meets these requirements?

You are framing an ML project to predict whether a customer will churn in the next 30 days. Business stakeholders want a single number that is easy to understand and robust to class imbalance (only 3% churn). Which primary evaluation metric should you recommend?

A data science team trains models in Vertex AI using a training script that reads data from Cloud Storage. They want to ensure experiments are reproducible and that the exact dataset snapshot used for each run can be retrieved later. What is the best approach?

You have a batch prediction job on Vertex AI that writes results to BigQuery. The business wants a daily dashboard showing key aggregates from the predictions. Which approach is the most appropriate and simplest?

A team needs to create training data by joining clickstream logs (in BigQuery) with user profiles (also in BigQuery) and generating session-level features daily. They want a managed service to define the pipeline with SQL and scale without managing clusters. What should they use?

You deployed a Vertex AI endpoint for online predictions. Latency is higher than expected, and Cloud Logging shows frequent cold-start-like spikes. The model is used continuously throughout the day, but traffic is bursty. What is the best mitigation?

A model trained on historical data performs well offline, but online performance degrades because the distribution of several input features changes weekly. You want an automated way to detect and alert on feature distribution shifts in production. What should you implement?

You are training a deep learning model on Vertex AI and want to track hyperparameters, metrics, and artifacts consistently across runs, and later compare experiments. Which approach is recommended?

Your organization requires that all ML pipeline runs are fully auditable: you must capture lineage from raw data to features to model artifacts to deployment, including who approved production promotion. You are using Vertex AI Pipelines. Which design best meets this requirement?

A financial services company must train a model on sensitive data. Policies require that the training job cannot access the public internet, all traffic must stay within Google Cloud, and data exfiltration risk must be minimized. Which architecture is most appropriate on Vertex AI?

You are building a feature store-like system for online serving. Features must be computed from streaming events and be available for low-latency reads during prediction. You also need point-in-time correctness for offline training datasets. Which approach best satisfies both online and offline needs on Google Cloud?

You are scoping an ML project to predict customer churn. Business stakeholders insist on measuring success in a way that reflects the cost of incorrectly flagging a customer as “will churn” (unnecessary retention offer) versus missing an actual churner (lost revenue). What is the BEST evaluation approach?

You are designing a feature store strategy for multiple teams training models on the same user and product features. The teams want consistency between training and online serving and want to avoid duplicating feature engineering logic. Which Google Cloud approach BEST meets these goals?

A data engineering team needs to compute daily aggregates and windowed features from streaming click events. They want exactly-once processing semantics and the ability to backfill from historical data using the same code. Which solution is MOST appropriate?

You trained a classification model on Vertex AI and want to deploy it to an endpoint. Your requirement is to keep serving costs low for a model with predictable traffic and minimal burstiness. Which deployment option is the BEST fit?

A team is training models using features computed in BigQuery. During online serving, they recompute several of those features in a microservice, and model performance is significantly worse than offline evaluation. What is the MOST likely root cause?

You need to orchestrate an end-to-end ML workflow that includes data extraction, feature engineering, training, evaluation, and conditional model registration only when metrics meet a threshold. You want reusable components and lineage tracking. Which approach is MOST appropriate?

Your organization must ensure that only approved features are used for model training because of privacy policy (e.g., disallowing certain user attributes). You want an enforceable control that prevents accidental inclusion during dataset creation. What is the BEST solution on Google Cloud?

A model is deployed to a Vertex AI endpoint. After a new data source is added, prediction quality degrades over several weeks, but latency and error rates remain normal. You need to detect this issue early in the future. What should you implement?

You are training a deep learning model on large image data. Training is unstable and occasionally diverges. You suspect the input pipeline is producing nondeterministic preprocessing differences between runs, making debugging difficult. You want reproducible training runs end-to-end on Google Cloud. What is the BEST combination of actions?

A regulated enterprise needs to serve a model with strict data residency: inference requests and all transient logs must remain within a specific region. They also require CMEK for storage and want to minimize the risk of accidental cross-region access in the pipeline. Which architecture BEST satisfies these constraints?

You are scoping an ML solution to help a support team prioritize inbound tickets. The business wants a 0–100 priority score, but agents ultimately choose which tickets to handle next. Which problem framing is most appropriate to start with on Google Cloud?

A team wants to accelerate feature engineering for tabular ML models by reusing computed features across training and online serving. They need consistency between offline and online features and low-latency reads for serving. Which Google Cloud approach best meets this requirement?

You deployed a model to a Vertex AI endpoint and want to gradually shift traffic from the current model to a newly trained model while monitoring performance. What is the recommended way to do this?

A data pipeline ingests events into BigQuery. During training, you discover label leakage: some features include information only available after the outcome occurred. Which change best prevents leakage while keeping training data accurate?

You are training a custom model on Vertex AI using GPUs. The training job frequently fails with out-of-memory errors midway through an epoch. Which action is the most effective first step to stabilize training without changing the model architecture?

A team needs an automated pipeline that: (1) runs a daily Dataflow job, (2) trains a Vertex AI model, (3) evaluates against a holdout set, and (4) only deploys if quality gates pass. They want clear lineage and reproducible runs. Which approach best fits?

A product team wants near-real-time personalization. They need to compute user aggregates (e.g., last 10 minutes clicks) and serve predictions with low latency. The raw events arrive continuously. Which architecture is most appropriate on Google Cloud?

Your Vertex AI endpoint shows a gradual drop in model performance in production. Labels arrive with a 7-day delay. You need an approach to detect drift early and decide when to retrain. What should you do?

A regulated healthcare company must train models on PHI and ensure the training environment cannot exfiltrate data to the public internet. They also need to call Vertex AI managed training. Which configuration best meets this requirement?

You have a large, imbalanced dataset for fraud detection (0.2% positive). You need a model that supports calibrated probability outputs because downstream systems use thresholds that change by region and time. Which approach is most appropriate?