Google Cloud Professional Machine Learning Engineer

About the Professional Machine Learning Engineer exam

Google Cloud Professional Machine Learning Engineer is for people who take models from a notebook to production on Google Cloud and keep them healthy there. It covers classic ML (choosing a model type, training, tuning, serving, monitoring for drift) and, increasingly, generative AI: picking a foundation model, tuning Gemini, evaluating outputs and guarding against malicious prompts. Google recommends three or more years of industry experience, including at least a year designing and managing solutions on Google Cloud. There is no prerequisite.

The exam does not test coding directly, but some questions include Python or SQL snippets you need to read. Most questions are scenarios: a team needs predictions with low latency, a model does well offline and badly in production, or a pipeline has to retrain when data changes. You pick the approach that solves it with the least custom work. Wrong options often build by hand something a managed service already does.

The current exam guide, dated June 1, 2026, uses the name Gemini Enterprise Agent Platform for the ML platform previously called Vertex AI, so expect that name, and "Agent Platform", on the exam.

Exam format

• Questions: 50 to 60, multiple choice and multiple select.
• Time: 2 hours.
• Passing score: Google reports pass or fail and does not publish a cut score.
• Cost: 200 USD plus tax, online with a remote proctor or at a Pearson VUE test center. English and Japanese.

What the exam covers

The official exam guide has six sections:

• Architecting low-code AI solutions (~13%). BigQuery ML and AutoML for classification, regression and forecasting; Document AI, Vision and Translation APIs; choosing and tuning models from Model Garden.
• Collaborating to manage data and models (~16%). Preprocessing with BigQuery, Dataflow or Spark, Feature Store, handling PII, notebooks in Workbench and Colab Enterprise, and tracking experiments and lineage.
• Scaling prototypes into ML models (~21%). Choosing model types and products, custom training, hyperparameter tuning, fine-tuning foundation models, and picking CPUs, GPUs or TPUs with data or model parallelism.
• Serving and scaling models (~20%). Batch versus online inference, prebuilt and custom containers, Model Registry, canary and A/B rollouts, private endpoints, and scaling the serving backend.
• Automating and orchestrating ML pipelines (~18%). Agent Platform Pipelines and Kubeflow, keeping preprocessing identical between training and serving, retraining policies, and CI/CD/CT with Cloud Build.
• Monitoring AI solutions (~13%). Model Armor and safety filters, responsible AI and bias, explainability, and Model Monitoring for training-serving skew, data drift and concept drift.

How the practice questions here are organised

There are two 10-question practice tests, and each covers all six sections in roughly the exam's proportions. Test 1 starts with the fundamentals: low-code image classification, Feature Store online serving, target leakage, moving a notebook to a training job, Vizier tuning, batch versus online inference, traffic splitting, pipeline orchestration, skew detection and explanation methods. Test 2 goes further: BigQuery ML forecasting, de-identifying PII before training, experiment tracking, accelerator choice, Reduction Server, custom serving containers, private endpoints, retraining policy, CI/CD for pipeline components and screening prompts with Model Armor. Every question explains the right answer and each wrong one, with a documentation link.

A study plan that works

• Take Test 1 cold, read every explanation, then take Test 2 a day or two later.
• Learn the managed-first ladder: a pre-trained API, then BigQuery ML or AutoML, then a tuned foundation model, then custom training. Questions often ask for the lowest rung that meets the requirement.
• Be clear on the monitoring vocabulary: training-serving skew, data drift, concept drift and feature attribution drift each have a different cause and fix.
• In a test project, train a BigQuery ML model on a public dataset, then deploy a small model to an endpoint and split traffic between two versions.
• Read the current exam guide closely for product names, since several changed in 2026.

These questions are original, written from the public Google Cloud documentation and the official exam guide. They are not real exam questions, and this site is not affiliated with Google.