← All Google Cloud AI guidesAll sections
PMLE exam overview, renames and strategy
Start here. What the Google Cloud Professional Machine Learning Engineer exam covers (June 2026 guide), where study time pays off, the 2026 product renames, how to read Google's scenario questions, and a "which service?" chooser across the whole exam.
1. The exam at a glance
The Professional Machine Learning Engineer (PMLE) exam tests whether you can choose and operate the right approach and Google Cloud service at each stage of an ML system, for classic predictive ML and for generative AI (gen AI). It's a judgement exam: the guide says coding isn't assessed directly, but you should be able to read Python and SQL snippets.
| Item | Detail |
|---|---|
| Format | Multiple choice and multiple select, about 50–60 questions in 2 hours, at a test centre or online-proctored |
| Result | Pass or fail; no published passing score |
| Recommended experience | 3+ years in industry, including 1+ year designing and managing solutions on Google Cloud |
| Validity | 2 years |
Logistics (price, languages, question count) change; confirm on the official certification page. Sections and weights below are from the exam guide dated June 1, 2026.
| Exam section | Weight | Guides |
|---|---|---|
| 1. Architecting low-code AI solutions | ~13% | BigQuery ML, AutoML and AI APIs, Gen AI stack |
| 2. Collaborating within and across teams to manage data and models | ~16% | Data, Feature Store, notebooks, experiments, Fundamentals 1, Fundamentals 2, LLM evaluation |
| 3. Scaling prototypes into ML models | ~21% | Training, tuning and hardware, Fundamentals 2, Gen AI stack |
| 4. Serving and scaling models | ~20% | Serving and scaling |
| 5. Automating and orchestrating ML pipelines | ~18% | Pipelines and MLOps |
| 6. Monitoring AI solutions | ~13% | Security, responsible AI and monitoring, RAG failure modes, LLM evaluation |
2. Where study time pays off
Most candidates who already know Google Cloud lose points in two places, so most of your study time should go there:
| Area | Why it matters | Guides |
|---|---|---|
| ML fundamentals: feature engineering, leakage, class imbalance, bias-variance, regularization, evaluation metrics (PR AUC vs ROC AUC, thresholds, calibration) | Many scenario questions hinge on a fundamentals judgement wrapped in a Google Cloud story: "the model is 99% accurate but…", "training loss falls while validation loss rises…". The service is often obvious; the ML reasoning isn't. | Fundamentals 1, Fundamentals 2 |
| The gen AI stack: Gemini and Model Garden on Agent Platform, prompting vs grounding vs tuning, RAG and how it fails, LLM and agent evaluation, gen AI safety | The 2026 guide adds gen AI to nearly every section: model choice, tuning, cost and latency, evaluation with LLM-as-a-judge, Model Armor, monitoring gen AI. Product names also changed, which trips people who studied older material. | Gen AI stack, RAG failure modes, LLM evaluation |
| MLOps on Google Cloud: pipelines, serving, monitoring | Heavily weighted, but very pattern-based: learn the managed option and its gotchas | Pipelines, Serving, Monitoring |
Suggested study path
- This page, then the two fundamentals guides. Do every calculator until the numbers feel obvious.
- The gen AI trio: stack, RAG, evaluation.
- Section guides in exam order: low-code, data, training, serving, pipelines, monitoring.
- Hands-on labs in a sandbox project (each guide that has one ends its theory with it).
- Quizzes again, a week later. Re-read the "Common gotchas" list of every guide the night before.
3. The 2026 renames
In April 2026 Google folded Vertex AI into Gemini Enterprise Agent Platform ("Agent Platform") and renamed Cloud Composer to Managed Service for Apache Airflow. The exam uses the new names; most courses, blogs and code still use the old ones. APIs stayed backward compatible: the Python package is still google-cloud-aiplatform, the API host is still aiplatform.googleapis.com, and the CLI is still gcloud ai and gcloud composer.
| Old name | 2026 name |
|---|---|
| Vertex AI | Gemini Enterprise Agent Platform (Agent Platform) |
| Vertex AI Studio | Agent Studio |
| Vertex AI Model Garden | Agent Platform Model Garden |
| Vertex AI Workbench / Colab Enterprise | Agent Platform Workbench / Colab Enterprise |
| Vertex AI Training | Agent Platform Managed Training (and Serverless Training, Managed Training Clusters) |
| Vertex AI Prediction / Endpoints / Batch Prediction | Agent Platform Inference: Online Inference, Endpoints, Batch Inference |
| Vertex AI Pipelines, Experiments, Feature Store, Model Registry, ML Metadata, Model Monitoring | Same names, on Agent Platform |
| Vertex AI Search | Agent Search |
| Vertex AI Vector Search | Vector Search (Vector Search 2.0 is now Agent Retrieval) |
| Vertex AI Agent Engine | Agent Runtime (with Sessions, Memory Bank, Code Execution) |
| Gen AI evaluation service | Agent Platform Evals |
| Cloud Composer | Managed Service for Apache Airflow (Managed Airflow) |
| Cloud DLP | Sensitive Data Protection (earlier rename; the DLP API keeps its name) |
An answer that says "Vertex AI Pipelines" and one that says "Agent Platform Pipelines" describe the same product. Never eliminate an answer because it uses an old name.
4. How to read Google scenario questions
Most questions describe a company, a constraint and a goal, then offer four designs that would all work in some sense. The right one meets the binding constraint with the least custom work.
- Find the binding constraint. "Minimal code" or "analysts know SQL"; "lowest latency"; "lowest cost"; "data must not leave the VPC"; "must explain each prediction"; "small labelled dataset"; "fastest time to market"; "existing Spark/Kubeflow investment".
- Climb the managed ladder from the bottom and stop at the first rung that meets every requirement.
- Move code to the data. Data in BigQuery points to BigQuery ML or a BigQuery-sourced pipeline, not a CSV export.
- Prefer managed services (Pipelines over cron scripts, Model Monitoring over hand-built histogram jobs) unless the question rules them out.
- Check the ML reasoning even when the service is obvious: leakage, imbalance, the right metric, the right split.
- Eliminate the over-engineered and the under-powered. TPU custom training for 10,000 tabular rows is wrong; so is AutoML when the team needs a custom loss.
| Rung | Option | Use when |
|---|---|---|
| 1 | Prebuilt API (Vision, Document AI, Translation, Speech) | Generic task; standard classes are fine |
| 2 | BigQuery ML or AutoML | Your labels, tabular or image data, little or no ML code |
| 3 | Gemini or another Model Garden model: prompt, then ground, then tune | Language, documents, multimodal, generation |
| 4 | Managed custom training on Agent Platform | Your architecture, loss, framework or scale |
| 5 | Self-managed on GKE, Ray or Compute Engine | Existing platform, special scheduling or control the managed path can't give |
"API, SQL, Auto, Gemini, Custom, DIY." From least to most effort. The exam usually wants the lowest rung that satisfies every stated requirement.
"Technically correct but more work" is the most common wrong answer. When two answers both work, choose the more managed one that keeps data where it is and meets the constraint.
Multiple-select questions
- The stem says how many to pick ("Choose two"). Each chosen option must be independently true and relevant.
- Watch for one option that is true but doesn't address the stated problem.
5. "Which service?" chooser
Pick an item to see what it means, why, and the tempting wrong choice.
Pick one above.
6. Scenarios
Exam-style situations: how to diagnose each one, the likely cause and the fix.
7. Searchable reference
| Item | Meaning |
|---|
8. Flashcards (click to flip)
9. Common gotchas in one list
- Old and new names are the same product: Vertex AI = Agent Platform; Composer = Managed Airflow; DLP = Sensitive Data Protection.
- Lowest rung that meets every requirement; extra custom work is the usual wrong answer.
- Keep the data where it is: BigQuery data suggests BigQuery ML or BigQuery-sourced training.
- Check the fundamentals inside cloud stories: leakage, imbalance, metric choice, time-based splits.
- Fresh facts need grounding, not tuning.
- Managed beats DIY unless the stem names an existing platform or a requirement only DIY meets.
- Multiple select: every chosen option must be true and address the problem.
- Security words change the answer: "no public internet", "exfiltration", "PII", "CMEK".
10. Practice quiz (16 questions)
Score: 0 / 0