Google Cloud Professional Machine Learning Engineer

About the Google Cloud Professional Machine Learning Engineer exam

The exam at a glance

Questions

50 to 60 questions, multiple choice and multiple select

Time

120 minutes

Pass mark

Not published by Google; pass or fail is reported

Exam fee

US$200 plus tax where applicable

Valid for

2 years; renew by passing the exam again, from 60 days before expiry

Where you sit it

Pearson VUE test centre or online proctored (OnVUE)

Prerequisites

None required; Google recommends 3 or more years of industry experience, including 1 or more on Google Cloud

As published in the vendor’s exam guide at the time of writing. Fees vary by country and change; the vendor’s page is the source of truth.

What is on the exam, domain by domain

1. Architecting low-code AI solutions

   Building classification, regression, forecasting and clustering models in BigQuery ML or AutoML, feature engineering and predictions in BigQuery ML, fine-tuning Gemini from BigQuery, and building with Model Garden, Gemini, Veo, models as a service and the Document AI, Vision and Translation APIs. The guide still names Imagen, which Google has retired in favour of Gemini image models.

   13% of the paper

2. Collaborating within and across teams to manage data and models

   Exploring and preprocessing data with BigQuery, Dataflow, Spark or Python, Feature Store, handling PII, prototyping in Workbench and Colab Enterprise, and tracking experiments, evaluation and lineage.

   16% of the paper

3. Scaling prototypes into ML models

   The largest section. Choosing model type and product, organising training data, custom training and AutoML, troubleshooting training, hyperparameter tuning, fine-tuning foundation models, and choosing CPUs, GPUs or TPUs with data or model parallelism.

   21% of the paper

4. Serving and scaling models

   Batch and online inference on Agent Platform, Cloud Run and GKE, prebuilt and custom containers, Model Registry versioning, A/B and canary rollouts, public and private endpoints, and scaling the serving backend.

   20% of the paper

5. Automating and orchestrating ML pipelines

   Building pipelines with Agent Platform Pipelines, Managed Service for Apache Airflow (formerly Cloud Composer) or Ray, validating data and models, keeping preprocessing consistent between training and serving, retraining policies and CI/CD/CT with Cloud Build.

   17% of the paper

6. Monitoring AI solutions

   Protecting AI systems from data leaks and malicious prompts with safety filters and Model Armor, monitoring for bias, explainability, Model Monitoring, and spotting training-serving skew, data drift, concept drift and feature attribution drift.

   13% of the paper

How to prepare

1. 1.Read the June 2026 exam guide and study to its bullets. Learn both product names: the exam says Agent Platform, while most courses and documentation pages still say Vertex AI.
2. 2.Build one model end to end in BigQuery ML: CREATE MODEL, ML.EVALUATE and ML.PREDICT on a public dataset. Then train the same problem with AutoML and compare the effort.
3. 3.Run a small pipeline in Agent Platform Pipelines that trains, evaluates and registers a model, deploy it to an endpoint, then turn on Model Monitoring. The exam rewards knowing how the pieces connect.
4. 4.Learn the trade-offs that decide most questions: low-code against custom training, batch against online prediction, CPU against GPU against TPU, and the managed option with the least operational overhead.
5. 5.Take Google's free sample questions, then two timed practice papers above 80 percent before you book.

Sample questions

Three questions in the exam’s style, written by Square 1 against the published domains. They are ours, not the vendor's. The answer and the reasoning are under each one; the practice papers on this page draw from a bank written the same way, with an explanation on every question.

1. Question 1 · Architecting low-code AI solutions

   Your BigQuery ML boosted tree model scores loan applications in a nightly batch. The web team now needs a score within a second while an applicant fills in a form. You want to avoid building and maintaining a serving container. What should you do?

   1. ARegister the model in Model Registry on Agent Platform and deploy it to an endpoint
   2. BRun ML.PREDICT from the web application as a BigQuery query on every form submission
   3. CExport the model to Cloud Storage and serve it from a custom container on GKE
   4. DSchedule the ML.PREDICT query every minute and have the web app read the results table

   Answer: A. Registering a BigQuery ML model in Model Registry lets you deploy it to an Agent Platform endpoint for online inference without a serving container. A custom container on GKE meets the latency need but adds exactly the build and maintenance work the team wants to avoid.

2. Question 2 · Serving and scaling models

   You work for an energy retailer. A churn model is registered in Model Registry, and the marketing team needs scores for all 8 million customers every Monday morning, written to BigQuery. Today a script calls an always-on endpoint row by row, and the endpoint sits idle the rest of the week. You want to cut cost and operational overhead. What should you do?

   1. AKeep the endpoint but lower its minimum replica count to one node
   2. BMove the scoring script to a Compute Engine VM that runs weekly
   3. CRun a weekly Agent Platform batch inference job writing to BigQuery
   4. DSend the rows to the endpoint in larger request payloads

   Answer: C. A batch inference job reads the input, scores it and writes to BigQuery with no endpoint, and its nodes stop when the job ends. Lowering the minimum replica count still leaves a paid node running all week and keeps the row-by-row client code.

3. Question 3 · Monitoring AI solutions

   You work for a payments company. Model Monitoring shows no feature drift on your fraud model, but chargeback data that arrives weeks later shows the model is missing more fraud each month. Fraudsters have changed how they split transactions. What should you do?

   1. ALower the drift alert thresholds on all features to 0.1 so that smaller changes are flagged
   2. BEnable output inference drift monitoring to compare prediction distributions over time
   3. CSwitch skew detection to use last month's serving data as its baseline
   4. DJoin chargeback labels to logged predictions, track recall on a schedule and retrain on recent data

   Answer: D. Stable inputs with falling accuracy is concept drift: the relationship between inputs and labels has changed, and only evaluation against ground truth shows it. Tighter drift thresholds still only measure input distributions, which have not changed.

Questions people ask before they book

Does the exam still cover Vertex AI?

Yes, under its new name. Google renamed Vertex AI to Gemini Enterprise Agent Platform in 2026, and the exam guide dated 1 June 2026 uses the new names, such as Agent Platform Pipelines and Model Registry. The services work the same way, so Vertex AI study material is still useful if you map the names.

How much does it cost and how long is it valid?

US$200 plus tax where applicable. The certification is valid for two years, and you renew it by passing the exam again. Renewal opens 60 days before it expires, and Google gives a 50 percent renewal discount code when you first certify.

Do I need to code for the exam?

Not directly. Google says the exam does not test coding, but some questions include short Python or SQL snippets, so you should be able to read them.

Do I need another Google Cloud certification first?

No, there are no prerequisites. Google recommends three or more years of industry experience, including a year or more on Google Cloud. The Associate Cloud Engineer helps if Google Cloud itself is new to you.

How long should I study?

Eight to twelve weeks part-time for most people who already train and deploy models, and longer if Google Cloud or MLOps is new to you.