Professional ML Engineer (PMLE) Study Guide: Google Cloud

GCP Professional Machine Learning Engineer (PMLE) Study Guide 2026 — hero

The Google Cloud Professional Machine Learning Engineer (PMLE) exam is 2 hours, with 50-60 multiple choice and multiple select questions, and costs USD 200 (plus tax). Google does not publish a passing score. The certification is valid for 2 years.

Practice this exam in timed mode with the StudyKits Professional Machine Learning Engineer app. The Google Cloud certification hub lists every Google Cloud exam guide and practice app.

The current exam guide is dated June 1, 2026 and has six sections, led by scaling prototypes into ML models (~21%) and serving and scaling models (~20%). Google recommends 3+ years of industry experience, including 1+ year on Google Cloud. Facts are from the Google pages under Sources, checked October 15, 2026.

This guide walks through each section, the ML basics the exam assumes, and an 8-week study plan.

What is the PMLE exam format?

The Professional Machine Learning Engineer exam has 50-60 questions and you get 2 hours. Google does not publish an exact passing score. The exam costs USD 200 (plus tax) and is offered in English and Japanese.

Exam guide update (October 2026): Google’s PMLE exam guide is dated “as of June 1, 2026”. It renames the sections (for example “Architecting low-code AI solutions” and “Monitoring AI solutions”) and reflects the transition from Vertex AI to Gemini Enterprise Agent Platform naming. The weights below follow that guide; service names in this post still use the Vertex AI names you will see in older material.

Google recommends 3+ years of industry experience including 1+ years designing and managing ML solutions on Google Cloud. This is not a beginner certification. You need solid understanding of ML fundamentals (supervised/unsupervised learning, model evaluation, feature engineering) in addition to Google Cloud platform knowledge.

If you have already passed the Professional Cloud Architect or Professional Data Engineer exams, you have a strong foundation to build on.

What do the six PMLE sections cover?

Domain 1: Architecting Low-Code AI Solutions (~13%)

This domain focuses on using Google Cloud’s pre-built and low-code ML services.

AutoML on Vertex AI:

AutoML for tabular data — classification and regression without writing code
AutoML for image classification, object detection, and image segmentation
AutoML for text classification, entity extraction, and sentiment analysis
AutoML for video classification and object tracking
Understanding when AutoML is the right choice vs custom training
Data requirements: minimum dataset sizes, data quality considerations
Model evaluation metrics in AutoML: precision, recall, F1, AUC-ROC, confusion matrix

Pre-built APIs:

Vision API — image labeling, OCR, face detection, product search
Natural Language API — entity recognition, sentiment analysis, syntax analysis, content classification
Speech-to-Text and Text-to-Speech APIs
Translation API — neural machine translation
Video Intelligence API — shot detection, label detection, explicit content detection
Document AI — document parsing, form extraction
When to use pre-built APIs vs AutoML vs custom models

BigQuery ML:

Training models with SQL: linear regression, logistic regression, k-means, time series (ARIMA_PLUS), matrix factorization, deep neural networks, XGBoost
Importing TensorFlow models into BigQuery ML
Exporting BigQuery ML models to Vertex AI for serving
When BigQuery ML is the right choice: data already in BigQuery, SQL-skilled team, rapid prototyping

Domain 2: Collaborating within and across Teams (~16%)

This domain tests how ML engineers work with other teams and manage the ML lifecycle.

Vertex AI Workbench — managed Jupyter notebooks for ML development. Know the difference between managed notebooks and user-managed notebooks.
Experiment tracking — using Vertex AI Experiments to track metrics, parameters, and artifacts across training runs
Model Registry — versioning models, managing model metadata, comparing model versions
Feature Store — centralizing feature definitions, serving features for online and offline use, avoiding training-serving skew
Data labeling — Vertex AI Data Labeling service, active learning for efficient labeling
ML workflow management — roles and responsibilities of data engineers, data scientists, ML engineers, and MLOps engineers

Domain 3: Scaling Prototypes into ML Models (~21%)

This domain covers taking a proof-of-concept model and preparing it for production.

Custom training on Vertex AI:

Training with pre-built containers (TensorFlow, PyTorch, XGBoost, scikit-learn)
Custom containers for specialized frameworks
Distributed training — data parallelism and model parallelism
Hyperparameter tuning with Vertex AI Vizier
Training on GPUs and TPUs — when to use each, cost considerations

Model architecture considerations:

Transfer learning with pre-trained models
Model size and latency trade-offs
Quantization and pruning for smaller, faster models
Handling class imbalance: SMOTE, class weights, oversampling, undersampling

Data processing at scale:

Feature engineering with Dataflow
Data validation with TensorFlow Data Validation (TFDV)
Data transformation with TensorFlow Transform (TFT)
Feature selection techniques and dimensionality reduction

TPU (Tensor Processing Unit) knowledge:

TPU types: v2, v3, v4, v5e
When TPUs outperform GPUs (large batch sizes, transformer models)
TPU training strategies and limitations
TPU pricing and cost optimization

Domain 4: Serving and Scaling Models (~20%)

This domain tests deployment and serving strategies.

Vertex AI Prediction:

Online prediction (real-time) vs batch prediction
Endpoint management: deploying models to endpoints, traffic splitting
Autoscaling: minimum/maximum replicas, scaling based on CPU, GPU, or request count
Model monitoring: detecting data drift, concept drift, and prediction drift
A/B testing with traffic split percentages

Serving infrastructure:

GPU vs CPU for serving — latency and cost trade-offs
Custom prediction containers for pre/post-processing
Model optimization for serving: TensorRT, ONNX, TensorFlow Lite
Edge deployment with Vertex AI for edge devices

Model versioning and rollback:

Deploying new model versions with traffic split
Canary deployments: routing a small percentage of traffic to the new model
Rollback strategies when model performance degrades
Model warm-up for reducing cold start latency

Domain 5: Automating and Orchestrating ML Pipelines (~18%)

MLOps is a major focus of the PMLE exam.

Vertex AI Pipelines:

Building ML pipelines with Kubeflow Pipelines SDK or TFX
Pipeline components: data ingestion, preprocessing, training, evaluation, deployment
Conditional logic and branching in pipelines
Pipeline scheduling and triggering
Artifact tracking and lineage

CI/CD for ML:

Continuous training: retraining models automatically when new data arrives or performance degrades
Continuous deployment: automatically deploying models that pass evaluation criteria
Cloud Build integration for ML pipeline automation
Infrastructure as code for ML resources with Terraform

Monitoring and maintenance:

Vertex AI Model Monitoring for data drift and prediction drift
Setting up alerts when model performance drops below thresholds
Retraining strategies: scheduled retraining vs triggered retraining
A/B testing and champion/challenger patterns

Domain 6: Monitoring AI Solutions (~13%)

Model performance monitoring — tracking accuracy, latency, and throughput in production
Data quality monitoring — detecting schema changes, missing values, distribution shifts
Bias and fairness — Vertex AI Explainable AI for model interpretability, What-If Tool for exploring model behavior
Feature attribution — understanding which features drive predictions
Logging and debugging — Cloud Logging for prediction logs, Cloud Monitoring for infrastructure metrics
Cost monitoring — tracking training costs, serving costs, and optimizing resource usage

How does PMLE compare with AWS ML certifications?

If you are familiar with the AWS AI Practitioner or ML certifications, here is how the GCP PMLE maps:

GCP Service	AWS Equivalent	Key Difference
Vertex AI	SageMaker	Both are unified ML platforms, Vertex AI emphasizes AutoML more
AutoML	SageMaker Autopilot	Similar capabilities, different interfaces
BigQuery ML	SageMaker Canvas / Athena ML	BigQuery ML uses SQL, tighter data warehouse integration
Vertex AI Pipelines	SageMaker Pipelines	GCP uses Kubeflow/TFX, AWS uses proprietary SDK
Feature Store	SageMaker Feature Store	Similar concepts, different implementations
TPUs	Inf2/Trn1 (Inferentia/Trainium)	TPUs are Google-designed, broader ML support
Vertex AI Prediction	SageMaker Endpoints	Similar serving capabilities
Cloud Vision API	Rekognition	Pre-built vision APIs
Cloud Natural Language	Comprehend	Pre-built NLP APIs

AWS retired its Machine Learning - Specialty exam on March 31, 2026. The closest AWS exam now is Machine Learning Engineer - Associate (MLA-C01, moving to MLA-C02; AWS’s September 2026 announcement gives January 14, 2027 for general availability, and the exam page lists TBD). PMLE is a professional exam and goes deeper into ML theory, model optimization and MLOps. The AWS associate exam leans more toward service selection.

Which ML fundamentals does the exam assume?

The exam assumes strong ML fundamentals. Review these before diving into GCP-specific content:

Supervised learning: Linear regression, logistic regression, decision trees, random forests, gradient boosting (XGBoost), neural networks, SVMs. Know when to use each.

Unsupervised learning: K-means clustering, hierarchical clustering, PCA, autoencoders. Know use cases and limitations.

Deep learning: CNNs for image tasks, RNNs/LSTMs for sequence tasks, Transformers for NLP. Understand architecture choices.

Model evaluation: Precision, recall, F1 score, AUC-ROC, RMSE, MAE, R-squared. Know which metric to use for different problem types.

Feature engineering: Normalization, standardization, one-hot encoding, embedding layers, feature crosses, bucketization. Understand training-serving skew.

Regularization: L1 (Lasso), L2 (Ridge), dropout, early stopping, data augmentation.

Bias-variance trade-off: Overfitting vs underfitting, cross-validation, learning curves.

What does an 8-week PMLE study plan look like?

Weeks 1-2: ML Fundamentals and Vertex AI Overview

Review ML fundamentals: supervised/unsupervised learning, evaluation metrics
Explore Vertex AI platform: Workbench, datasets, training, models, endpoints
Study AutoML and pre-built APIs
BigQuery ML: train simple models with SQL
Hands-on: train an AutoML model on a tabular dataset
20 practice questions per day

Weeks 3-4: Custom Training and Feature Engineering

Custom training with TensorFlow and PyTorch on Vertex AI
Distributed training, GPUs, and TPUs
Hyperparameter tuning with Vizier
Feature engineering with Dataflow and TFT
Feature Store setup and usage
Hands-on: train a custom TensorFlow model with hyperparameter tuning
25 practice questions per day

Weeks 5-6: MLOps and Pipelines

Vertex AI Pipelines with Kubeflow Pipelines SDK
CI/CD for ML with Cloud Build
Model monitoring and drift detection
Continuous training and deployment patterns
Hands-on: build an end-to-end ML pipeline
30 practice questions per day

Weeks 7-8: Serving, Monitoring, and Practice Exams

Model serving: endpoints, autoscaling, traffic splitting
Explainable AI and fairness
Cost optimization for ML workloads
Take full-length practice exams
Review weak areas
40 practice questions per day
Schedule your exam

What exam tips help most?

The exam blends ML theory with Google Cloud implementation. Study both sides.
Vertex AI is the central platform. If a question asks about training, serving, or monitoring, the answer likely involves Vertex AI.
Understand the full ML lifecycle: data preparation, training, evaluation, deployment, monitoring, retraining. Many questions test your understanding of where things can go wrong in this lifecycle.
For “which approach should you use” questions, consider the team’s skills, timeline, and data size. AutoML for quick prototyping with limited ML expertise, custom training for complex problems with experienced teams.
Cost questions often hinge on choosing preemptible VMs for training, right-sizing GPU/TPU choices, and using batch prediction instead of online prediction when real-time is not required.

How can you practise for PMLE?

Use this guide as a map, check each topic against the June 2026 exam guide, and build hands-on experience on the platform.

The StudyKits PMLE app has original practice questions, but they follow the previous exam outline (Framing ML problems through Monitoring, optimizing, and maintaining ML solutions); update in progress. The PMLE app page lists how many questions the app has in each of those sections.

Frequently asked questions

Is the PMLE exam updated for 2026?

Yes. The exam guide is dated June 1, 2026. Google says the exam now reflects the move from Vertex AI to Gemini Enterprise Agent Platform and changes to its data and analytics stack.

What is the PMLE passing score?

Google does not publish one for any of its certification exams.

How long is the PMLE certification valid?

2 years. There is no shorter renewal exam for PMLE yet.

Do I need the Data Engineer certification first?

No. Google lists no prerequisite. A data engineering background helps with the pipeline and data questions.

Sources

Checked October 15, 2026:

GCP Professional Machine Learning Engineer (PMLE) Study Guide 2026