Google Cloud Professional Machine Learning Engineer (PMLE) tests whether you can build, deploy and run ML and generative AI solutions on Google Cloud. It covers BigQuery ML, AutoML, Vertex AI training and serving, pipelines, Feature Store, monitoring and responsible AI. The current version (exam guide 3.1) adds generative AI: Model Garden, Vertex AI Agent Builder for RAG, fine-tuning foundation models and evaluating gen AI solutions.
This guide follows Google's official exam guide: all six sections with weights, the Vertex AI services to know, an 8-week plan and 15 original scenario questions.
Quick facts (Google Cloud, November 2026)
| Item | Detail |
|---|---|
| Length | 2 hours |
| Questions | 50–60, multiple choice and multiple select |
| Fee | USD 200 (plus tax where applicable) |
| Languages | English, Japanese |
| Delivery | Online-proctored or at a test centre |
| Prerequisites | None; 3+ years of industry experience including 1+ year designing and managing Google Cloud solutions recommended |
| Coding | Not directly assessed; basic Python and SQL help you read code snippets |
Read the official Professional ML Engineer page and exam guide before booking.
Who should take PMLE?
- ML engineers and data scientists deploying models on Google Cloud.
- MLOps engineers building training and serving pipelines.
- GenAI engineers who need the full ML lifecycle, not just prompting.
Strengthen Python and ML basics first with Python for AI engineers. For the business view, see Generative AI Leader.
Sections and weights
| Section | Weight |
|---|---|
| 1. Architecting low-code AI solutions | ~13% |
| 2. Collaborating within and across teams to manage data and models | ~14% |
| 3. Scaling prototypes into ML models | ~18% |
| 4. Serving and scaling models | ~20% |
| 5. Automating and orchestrating ML pipelines | ~22% |
| 6. Monitoring AI solutions | ~13% |
Section 1: Low-code AI (~13%)
- BigQuery ML: choosing the model type (linear and logistic regression, time series, matrix factorisation, boosted trees, autoencoders), feature engineering and predictions.
- ML APIs and foundation models: APIs from Model Garden; industry APIs such as Document AI; RAG apps with Vertex AI Agent Builder.
- AutoML: preparing data, tabular and forecasting models, training on text, images and video, and debugging models.
-- BigQuery ML: train and use a churn classifier with plain SQL
CREATE OR REPLACE MODEL analytics.churn_model
OPTIONS (model_type = 'BOOSTED_TREE_CLASSIFIER', input_label_cols = ['churned']) AS
SELECT tenure_months, monthly_spend, support_tickets, plan_type, churned
FROM analytics.customers_training;
SELECT customer_id, predicted_churned_probs
FROM ML.PREDICT(MODEL analytics.churn_model, TABLE analytics.customers_current);Section 2: Data and models across teams (~14%)
- Exploring and preprocessing data in Cloud Storage, BigQuery, Spanner, Cloud SQL, Spark and Hadoop; Vertex AI datasets; Dataflow, TFX and BigQuery preprocessing; Vertex AI Feature Store; privacy (PII and PHI).
- Prototyping in notebooks: Vertex AI Workbench, Colab Enterprise or Dataproc notebooks; security; Spark kernels; Git; TensorFlow, PyTorch, scikit-learn, Spark and JAX; Model Garden models.
- Experiments: Vertex AI Experiments, Kubeflow Pipelines and Vertex AI TensorBoard; evaluating gen AI solutions.
Section 3: From prototype to model (~18%)
- Choosing frameworks and architectures, including when interpretability matters.
- Training: organising data, file types, Vertex AI custom training, Kubeflow on GKE, AutoML, distributed training, hyperparameter tuning, troubleshooting, and fine-tuning foundation models.
- Hardware: CPU vs GPU vs TPU vs edge; distributed training with Reduction Server.
Section 4: Serving and scaling (~20%)
- Batch and online inference (Vertex AI, Dataflow, BigQuery ML, Dataproc); serving PyTorch and XGBoost; Model Registry; A/B testing.
- Scaling online serving: Feature Store, public and private endpoints, hardware choice, autoscaling and tuning for latency, memory and throughput.
| Need | Usual choice |
|---|---|
| Score 50 million rows overnight | Batch prediction (Vertex AI batch or BigQuery ML) |
| Real-time fraud score under 100 ms | Online prediction on a Vertex AI endpoint with low-latency features |
| Traffic must stay inside the VPC | Private endpoint |
| Compare a new model on 10% of traffic | Traffic split between models on one endpoint |
| Same features for training and serving | Vertex AI Feature Store |
Section 5: Pipelines and MLOps (~22%)
- End-to-end pipelines: data and model validation; consistent preprocessing for training and serving; hosting third-party tools such as MLflow; components, parameters and triggers; Vertex AI Pipelines, Kubeflow Pipelines and Cloud Composer; hybrid and multicloud; TFX.
- Retraining: retraining policy; CI/CD with Cloud Build or Jenkins.
- Metadata: Vertex AI Experiments and ML Metadata; model and dataset versioning; lineage.
Section 6: Monitoring AI (~13%)
- Risks: protecting data and models from exploitation; Google's responsible AI practices; fairness and bias; explainability on Vertex AI.
- Monitoring: continuous evaluation with Vertex AI Model Monitoring and Explainable AI; training-serving skew; feature attribution drift; comparing against baselines and simpler models; common errors.
8-week study plan
| Week | Focus | Lab |
|---|---|---|
| 1 | BigQuery ML and AutoML | Churn model in BigQuery ML; AutoML tabular on the same data |
| 2 | Model Garden, ML APIs, Document AI, Agent Builder RAG | RAG search app over sample PDFs |
| 3 | Workbench, Colab Enterprise, Experiments, TensorBoard | Track three experiments and compare them |
| 4 | Custom training, hyperparameter tuning, GPUs and TPUs, distributed training | Custom PyTorch job with tuning |
| 5 | Fine-tuning and evaluating foundation models | Tune a model on a small dataset; evaluate against the base model |
| 6 | Model Registry, endpoints, batch vs online, traffic splitting, Feature Store | Deploy two versions with a 90/10 split |
| 7 | Vertex AI Pipelines, CI/CD, metadata and lineage | Pipeline: preprocess → train → evaluate → deploy only if better |
| 8 | Model Monitoring, skew and drift, explainability, fairness; sample questions | Enable monitoring; book the exam |
15 scenario questions with answers
These are original scenarios written for this guide, not exam questions.
1. Analysts who know only SQL want to forecast weekly sales from BigQuery data. What do you use?
A BigQuery ML time-series model (ARIMA_PLUS).
2. A team needs a good tabular model fast, with little ML expertise. What do you use?
AutoML for tabular data on Vertex AI.
3. Insurance claim forms must be read and key fields extracted. What service?
Document AI.
4. Employees need a chatbot grounded in internal policy documents. What is the quickest route?
A RAG app with Vertex AI Agent Builder over the documents.
5. A foundation model must learn your company's support-reply style. What do you do?
Fine-tune (supervised tuning) a foundation model on example replies, then evaluate it against the base model.
6. Training features are computed in Python but online features are computed differently in Java, and accuracy drops in production. What is this?
Training-serving skew. Use shared preprocessing and a Feature Store.
7. Training a large deep learning model takes too long on one GPU. What helps?
Distributed training across multiple GPUs or TPUs, for example with Reduction Server.
8. You must find the best learning rate and tree depth automatically. What do you use?
Vertex AI hyperparameter tuning.
9. Fifty million records need predictions every night. Online or batch?
Batch prediction.
10. A new model should get 10% of live traffic first. How?
Deploy both models to one Vertex AI endpoint with a 90/10 traffic split.
11. Model inputs in production drift away from training data. What detects this?
Vertex AI Model Monitoring for skew and drift.
12. A regulator asks why the model rejected an applicant. What helps?
Explainable AI feature attributions on Vertex AI.
13. The whole flow (preprocess, train, evaluate, deploy) must be repeatable and versioned. What do you build?
A Vertex AI Pipeline (Kubeflow or TFX components) with metadata tracking.
14. A new model should deploy only if it beats the current one on validation metrics. Where is that enforced?
In a pipeline evaluation step that conditionally deploys, triggered through CI/CD.
15. Data scientists need managed Jupyter notebooks with enterprise security. What do you provide?
Vertex AI Workbench or Colab Enterprise.
After PMLE
- Practise with GCP AI interview questions and Kubernetes for AI interview questions.
- Read GCP for AI engineers and see live AI and ML jobs.
Cloud Soft Solutions in Ameerpet, Hyderabad teaches machine learning, MLOps and generative AI on Google Cloud, AWS and Azure. See the APEX AI, ML, Cloud and Cyber Security program and all courses.
📞 Call +91 96660 19191 · WhatsApp · Flat No. 513, 5th Floor, Nilagiri Block, Aditya Enclave, Ameerpet, Hyderabad
Frequently Asked Questions
How long is the Professional Machine Learning Engineer exam?
2 hours, with 50 to 60 multiple-choice and multiple-select questions.
How much does the Google ML Engineer exam cost?
USD 200, plus tax where applicable.
Does the PMLE exam test coding?
Not directly. Google says basic Python and SQL proficiency is enough to read the code snippets in questions.
Which PMLE section has the most weight?
Automating and orchestrating ML pipelines, at about 22%.
Does PMLE cover generative AI?
Yes. The current version covers Model Garden, RAG with Vertex AI Agent Builder, fine-tuning foundation models and evaluating gen AI solutions.
What experience does Google recommend?
3+ years of industry experience, including 1+ year designing and managing solutions on Google Cloud.
