PMLE study guide: ML Engineer domains and a 4-week plan

The Google Cloud Professional Machine Learning Engineer (PMLE) exam tests whether you can build, deploy, automate and monitor both traditional and generative AI solutions on Google Cloud. It is a professional-level certification for practising ML engineers. You answer 50 to 60 multiple-choice and multiple-select questions in two hours, online or at a test centre.

PMLE at a glance

Details as of September 2026 — confirm on the official exam page before booking.

Who this exam is for

Google recommends three or more years of industry experience, including at least one year designing and managing solutions on Google Cloud. There is no formal prerequisite. The exam guide describes someone with strong programming skills who handles large datasets, writes reusable code, and is comfortable with model architecture, data and ML pipelines, MLOps and metrics interpretation, plus a working knowledge of prompt and context engineering.

That profile matters for preparation. The exam rarely asks what a service is. It asks which service or design fits a stated constraint on cost, latency, scale, team skills or governance.

What the exam covers

The current guide is dated 1 June 2026. Google's exam page notes that it reflects the move from Vertex AI to Gemini Enterprise Agent Platform, updates to the data and analytics stack, and a preference for Google Cloud native solutions. If your notes say "Vertex AI", map each feature to its current Agent Platform name as you study.

Architecting low-code AI solutions (~13%)

This domain covers building models with BigQuery ML or AutoML on Agent Platform (classification, regression, forecasting, clustering), feature engineering and prediction in BigQuery ML, and fine-tuning Gemini models through BigQuery. It also covers choosing a model from Model Garden, building on industry APIs such as Document AI, Vision and Translation, and optimising Gemini-based applications for cost, latency and availability. The key skill is recognising when a low-code path is enough and custom training is overkill.

Collaborating within and across teams to manage data and models (~16%)

Expect scenarios on exploring and preprocessing data, choosing between BigQuery SQL, Dataflow, Spark and in-memory Python based on scale, building features in Feature Store, and handling PII. It also covers notebook prototyping in Agent Platform Workbench and Colab Enterprise, with collaboration and security settings, plus experiment tracking, lineage and evaluation, including LLM-as-a-judge for gen AI outputs.

Scaling prototypes into ML models (~21%)

The largest domain. You choose a model type (for example ARIMA, a deep network or an LLM), a product (AutoML, BigQuery ML, pipelines) and a deployment strategy, with interpretability requirements in play. Training questions cover organising data in Cloud Storage and BigQuery, training through different SDKs and on GKE, troubleshooting failed runs, hyperparameter tuning, and when fine-tuning a foundation model is justified. Hardware questions ask you to pick CPU, GPU or TPU and to reason about data versus model parallelism.

Serving and scaling models (~20%)

This domain tests batch versus online inference across Agent Platform, Model Garden, Cloud Run and GKE; prebuilt versus custom containers for frameworks such as PyTorch and XGBoost; versioning in Model Registry; and rollout patterns such as A/B and canary. Scaling questions cover Feature Store for online serving, public versus private endpoints, hardware choice and scaling to throughput.

Automating and orchestrating ML pipelines (~18%)

You build end-to-end pipelines with Agent Platform Pipelines, Managed Service for Apache Airflow (formerly Cloud Composer) or Ray, validate data and models, and keep preprocessing identical between training and serving. Retraining questions ask you to set a sensible retraining policy and wire models into CI/CD/CT with Cloud Build.

Monitoring AI solutions (~13%)

The final domain covers AI security (data exfiltration, malicious prompting, sensitive data sent to LLMs) and the matching controls, from regex and safety filters to Model Armor. It also covers bias monitoring, explainability, and Model Monitoring for training-serving skew, data drift, concept drift and feature attribution drift, plus evaluating gen AI solutions in production.

A 4-week study plan

This plan assumes eight to ten hours a week and hands-on access to a Google Cloud project. Time follows the weights, with serving and scaling given the most room.

Week 1: Low-code and data (13% + 16%). Train one model in BigQuery ML and one with AutoML on a tabular dataset, and note what each asked you to decide. Then practise choosing a preprocessing tool by data size and team. Close with mixed papers on both domains.

Week 2: Scaling prototypes (21%). Focus on model and product choice under constraints, distributed training and hyperparameter tuning. Write out, for each accelerator type, the kind of workload it suits. Use topic papers on training failures, where scenario detail decides the answer.

Week 3: Serving and pipelines (20% + 18%). Deploy a model to an endpoint, then compare batch inference. Build a small pipeline and trace how the same transformation runs at training and serving time. Study rollout strategies and retraining triggers together, because the exam often combines them.

Week 4: Monitoring (13%) and rehearsal. Study the drift types until you can tell them apart from symptoms alone, then cover AI security controls. Sit a full-length mock in two hours, target weak topics, then sit a second. The [AI certification study plan template](/blog/ai-certification-study-plan-template) helps if you need to stretch this over more weeks.

Where candidates lose marks

• Picking the most powerful option rather than the sufficient one. When BigQuery ML or AutoML meets the requirement, a custom training pipeline is usually wrong because it adds cost and maintenance. The guide's preference for native, managed solutions points the same way.
• Mixing up drift types. Training-serving skew, data drift and concept drift have different causes and different fixes. Read the symptom in the stem before you look at the options.
• Missing the constraint word. "Lowest latency", "minimal operational overhead", "data must not leave the VPC" each eliminate options. Underline them mentally before answering.
• Multiple-select questions. When a question asks for two or three answers, choose exactly that many and test each one on its own against the scenario. An option that is true in general but does not meet the stated requirement is still wrong.
• Studying renamed products. Older material teaches Vertex AI, Cloud Composer and deprecated features such as AutoML for text and video. Check current documentation before memorising a capability.
• Treating gen AI as a separate topic. The current guide threads foundation models through every domain, from BigQuery fine-tuning to LLM-as-a-judge and Model Armor.

Practising for PMLE on CertifyAIx

CertifyAIx has 541 original PMLE practice questions, written against the June 2026 exam guide, covering every domain and pitched at professional-level difficulty. Each question explains every option. Our article on [how CertifyAIx practice questions are written](/blog/how-certifyaix-practice-questions-are-written) describes the method.

• 3 full-length mock exams, timed at the real exam's two hours
• 25 mixed domain papers and 29 topic papers of up to 25 questions each (60 minutes per 25)
• Multiple-select questions alongside single-answer ones

Start with the [PMLE practice questions](/certifications/gcp-professional-machine-learning-engineer).

Frequently asked questions

How hard is the PMLE exam?

It is a professional-level exam, and Google recommends three or more years of industry experience with at least one year on Google Cloud. The difficulty comes from scenario judgement across the whole ML lifecycle rather than from recall.

How long should I prepare for PMLE?

For an engineer who already trains and deploys models on Google Cloud, four to six weeks of focused study is a reasonable range. If your ML experience is on another cloud, allow longer to learn Google's managed services and their current names.

Are CertifyAIx questions taken from the real exam?

No. All CertifyAIx questions are original and written from Google's public exam guide. Memorising leaked questions breaks the candidate agreement and does not prepare you for scenario reasoning.

Is the PMLE exam still about Vertex AI?

The concepts are the same, but Google's exam page says the exam now reflects the transition from Vertex AI to Gemini Enterprise Agent Platform. Study with current product names.

Should I take GAIL before PMLE?

It is not required. The [Generative AI Leader study guide](/blog/gail-study-guide) covers a business-level exam; PMLE is an engineering exam, and most candidates go straight to it if they already work in ML.

CertifyAIx is independent and not affiliated with Google. Professional Machine Learning Engineer is a trademark of its owner. All CertifyAIx practice questions are original.