Google Cloud Professional Machine Learning Engineer
225 practice questions
Last reviewed: April 2026
Personal notes and resource links for your study journey
Filter by Certification
The Google Cloud Professional Machine Learning Engineer (PMLE) validates the ability to design, build, and productionize ML models on Google Cloud — covering Vertex AI end-to-end, AutoML, custom training, model deployment, MLOps pipelines, and the operational realities of serving ML at scale. The exam emphasizes Vertex AI Pipelines (Kubeflow), Vertex AI Model Registry, Feature Store, Endpoints (online and batch), TensorFlow Extended (TFX), explainability with Vertex Explainable AI, monitoring for drift and skew, and integration with BigQuery ML and Generative AI offerings (Gemini family, Model Garden). Question style is scenario-heavy and rewards candidates who think about production ML lifecycle (CI/CD/CT), not just modeling.
AutoML for tabular / vision / language, BigQuery ML, pre-trained APIs (Vision, Speech, Translation, Document AI), and choosing between low-code and custom paths. Smallest domain at 12% but high-density.
Vertex AI Workbench, Feature Store (online and offline), data labeling and annotation, model versioning and metadata, experiment tracking with Vertex AI Experiments. 16%.
Custom training (single-node, distributed, GPU/TPU), hyperparameter tuning with Vizier, container-based training, Vertex AI Tuning, dealing with dataset bias. 18%.
Vertex AI Endpoints (online with autoscaling, traffic split), batch prediction, TensorFlow Serving, latency / throughput / cost tradeoffs, edge deployment. 19%.
Largest domain at 21%. Vertex AI Pipelines (Kubeflow Pipelines SDK and TFX), CI/CD/CT, retraining triggers, Cloud Build integration. Heavy on lifecycle automation.
Vertex AI Model Monitoring (training-serving skew, drift, attribution drift), Vertex Explainable AI, performance and cost monitoring with Cloud Operations. 14%.
Services you'll encounter on the exam and why each one matters.
Unified ML platform covering training, tuning, prediction, pipelines, model registry, feature store, and monitoring under a single API surface.
Why it's on the exam: Vertex AI is the umbrella across every PMLE domain — expect questions on choosing between AutoML, custom training, and pre-built containers for a given workflow.
Managed Jupyter-based development environment with built-in BigQuery, Dataproc, and Cloud Storage integrations for prototyping models.
Why it's on the exam: Domain 3 (Scaling Prototypes) tests Workbench as the canonical notebook surface for moving from experiment to production-grade training.
Managed custom and pre-built container training jobs on CPU/GPU/TPU with distributed training, hyperparameter tuning, and reduction-server support.
Why it's on the exam: Scaling training across accelerators and choosing managed-vs-custom containers is a recurring Domain 3 scenario.
Managed online and batch prediction with autoscaling, traffic-splitting between model versions, and private endpoint support over PSC.
Why it's on the exam: Domain 4 (Serving and Scaling Models) tests endpoint sizing, autoscaling thresholds, and canary rollouts between model versions.
Serverless orchestration of Kubeflow Pipelines and TFX DAGs with artifact lineage, caching, and Vertex ML Metadata integration.
Why it's on the exam: Domain 5 (Automating and Orchestrating ML Pipelines) names Pipelines as the AWS-native MLOps orchestrator versus generic Workflows or Composer.
Central catalog for trained model versions with deployment tracking, lineage to training jobs, and approval workflows for production rollout.
Why it's on the exam: Domain 2 (Collaborating to Manage Data and Models) tests how teams version, approve, and govern model artifacts across environments.
Managed online (low-latency) and offline feature repository with point-in-time correctness and BigQuery-backed offline storage.
Why it's on the exam: Feature Store is the canonical Domain 2 answer for preventing training/serving skew and sharing features across teams.
Drift and skew detection on deployed endpoints with feature-attribution monitoring, alerting via Cloud Monitoring, and BigQuery-backed analysis.
Why it's on the exam: Domain 6 (Monitoring and Optimizing) tests how to detect training/serving skew and prediction drift on live endpoints.
No-code training of tabular, image, text, and video models with managed feature engineering and hyperparameter search.
Why it's on the exam: Domain 1 (Low-Code ML Solutions) names AutoML as the canonical choice when domain experts need a model without writing training code.
Track training runs, parameters, metrics, and artifact lineage; query Vertex ML Metadata for reproducibility and audit.
Why it's on the exam: Domain 2 tests experiment tracking and reproducibility — Experiments + Metadata is the AWS-native lineage store.
Black-box hyperparameter optimization service usable standalone or embedded in custom training jobs, with Bayesian and grid search strategies.
Why it's on the exam: Domain 3 questions on efficient hyperparameter tuning at scale name Vizier as the managed alternative to grid-search-on-Compute-Engine.
Approximate-nearest-neighbor service (formerly Matching Engine) for embeddings-based retrieval at sub-100ms scale.
Why it's on the exam: Recommendation and RAG-style retrieval scenarios under Domain 4 name Vector Search as the managed serving layer for embeddings.
Train and serve regression, classification, time-series, and embedding models with SQL directly on BigQuery tables — no data movement required.
Why it's on the exam: Domain 1 + Domain 3 cite BigQuery ML when the data already lives in BigQuery and an analyst needs models without an ML pipeline.
End-to-end TensorFlow MLOps framework: ExampleGen, Transform, Trainer, Evaluator, Pusher — runs natively on Vertex AI Pipelines.
Why it's on the exam: Domain 5 tests TFX as the open-source pipeline framework that compiles into Vertex AI Pipelines for portable MLOps.
Apache Beam-based service for batch and streaming inference, feature engineering at scale, and Vertex AI integration via RunInference transforms.
Why it's on the exam: Domain 4 (Serving) tests Dataflow for streaming inference and bulk preprocessing pipelines that feed Vertex AI training jobs.
Managed TensorBoard for visualizing training metrics, scalars, embeddings, and profiler traces with team-level sharing via IAM.
Why it's on the exam: Domain 3 + Domain 6 reference TensorBoard for debugging convergence issues and profiling GPU utilization during training.
Account-wide access control plus Workload Identity Federation for binding GKE/Vertex AI service accounts to short-lived credentials.
Why it's on the exam: Domain 2 + Domain 5 test least-privilege service accounts for training jobs, pipeline components, and cross-project model serving.
Managed cryptographic keys with CMEK support for Vertex AI training data, model artifacts, BigQuery datasets, and Cloud Storage buckets.
Why it's on the exam: CMEK on training corpora and model artifacts is the canonical Domain 2 answer for protecting model IP and compliance-bound data.
Unified logs, metrics, and alerts across Vertex AI training jobs, endpoint invocations, pipeline step durations, and custom model metrics.
Why it's on the exam: Domain 6 expects Cloud Monitoring for endpoint latency/error SLOs and Cloud Logging for training-job troubleshooting.
Unified data fabric for cataloging, classifying, and tracking lineage of BigQuery datasets, Cloud Storage objects, and ML feature artifacts.
Why it's on the exam: Domain 2 (Collaborating to Manage Data and Models) tests Dataplex as the GCP-native answer for ML data lineage and feature governance.
$145k–$210k–$320k USD annual
Range reflects US-based senior ML engineers where Vertex AI is the primary platform. FAANG L5 ML engineer TC clears $400k; staff and principal levels go higher. ML engineering is the highest-paying cloud-engineering specialty by base, and the GCP-specific candidate pool is small relative to AWS / multi-cloud, which helps PMLE holders at hiring time.
Source: levels.fyi 2025–2026 (Google L4–L6 ML engineers, FAANG and AI-startup senior ML), U.S. BLS OEWS May 2024 (15-2099 mathematical science occupations / data scientists, 15-1252 software developers). Figures are approximate; actual compensation depends on role, region, and experience.
PMLE demand surged through 2024–2026 as GenAI hiring pulled qualified ML engineers across the board. Heavy demand at Google Cloud partners with ML practices, AI-first startups building on Vertex AI, and Google itself for customer-engineering ML specialists. The cert is also valuable on multi-cloud ML platform teams. PMLE pairs naturally with Professional Data Engineer (PDE) for an end-to-end "data + ML" senior profile and with Generative AI Leader (GAIL) for a strategic-plus-technical pair. Holders consistently report strong recruiter response — ML-engineering candidate pools remain tight even as the GenAI hype peak normalizes.
There are no formal prerequisites. Google recommends three or more years of industry experience and one or more years architecting and operationalizing ML solutions on Google Cloud. In practice, PMLE is not a credible first GCP cert and is rarely a credible first ML cert — successful candidates have shipped at least one production ML model and have working knowledge of TensorFlow or PyTorch.
Strong Python fluency, working knowledge of scikit-learn / TensorFlow / Keras / PyTorch, and at least conceptual familiarity with Kubeflow or another ML pipeline framework are effectively required. Comfort with BigQuery SQL is helpful since BigQuery ML appears in many scenarios. The official ML Engineer Learning Path on Google Cloud Skills Boost (around 50–80 hours) is a good baseline; most successful candidates also build a non-trivial Vertex AI Pipelines project end-to-end.
PMLE is rated professional and is consistently hard for candidates without production ML experience. Plan on 100–150 hours of study over 10–14 weeks if PMLE is your first ML-engineering cert, or 50–80 hours over 5–8 weeks if you already hold an AWS or Azure ML certification and have shipped models on either platform. The exam is 50–60 multiple-choice / multiple-select questions in 120 minutes, delivered through Pearson VUE (Google migrated from Kryterion / Webassessor in early 2026).
The most common stumbling block is the MLOps lifecycle — when to retrain, how to detect drift vs. skew, how to wire Vertex AI Pipelines into Cloud Build for CI/CD/CT. The second stumbling block is choosing between AutoML, BigQuery ML, custom training on Vertex AI, and pre-trained APIs for a given scenario, where Google's "preferred" answer often hinges on team skill and time-to-value rather than pure technical fit. Google does not publish numeric scores — only pass/fail. The credential is valid for two years and recertification requires re-passing the current exam.
Current exam guide refreshed in late 2024 to add Generative AI integration scenarios (Gemini, Model Garden), expanded Vertex AI Agent Builder coverage, and updated Feature Store content.
Major refresh consolidating around Vertex AI as the unified ML platform, retiring the older AI Platform / AutoML Tables coverage.
Original general availability, replacing the earlier "Data Engineer with ML focus" path.
PMLE (Google Cloud Professional Machine Learning Engineer) is a a challenging, scenario-heavy exam that requires deep hands-on experience and the ability to make architectural trade-off decisions Professional-level exam. Most candidates need 150–300 hours of study spread over 3–6 months for professional and expert-level exams. These exams typically expect prior associate-level proficiency. Most candidates who score consistently above the passing threshold on practice exams pass on their first attempt.
Most candidates need 150–300 hours of study spread over 3–6 months for professional and expert-level exams. These exams typically expect prior associate-level proficiency. Time-to-pass varies widely by prior experience. Engineers with hands-on production experience in the underlying technology typically need less; candidates new to the platform should plan toward the upper end of that range.
PMLE is a recognized credential in the GCP ecosystem and signals validated knowledge to employers, recruiters, and clients. Whether it is worth the time and fee for you depends on your role and goals — it tends to pay off most for cloud engineers, architects, and consultants who work with GCP day-to-day or want to move into roles that do.
The passing score for PMLE is Not published. The exam contains 50 questions and lasts 2 hr.
The PMLE exam fee is $200 USD. Fees are set by GCP and may vary by region; always confirm the current price on the official GCP certification page before booking.
Google Cloud Professional certifications are valid for 2 years. Recertify by re-passing the current version of the exam.
Yes. You can take the exam online (proctored via the provider's secure browser, available 24/7 in most regions) or at an in-person Pearson VUE test center during business hours. Both formats use the same questions, time limit, and passing score.
CertLabPro provides 15 study modes across the practice question bank for PMLE. The exam-simulation mode mirrors the real exam: 50 questions in 2 hr, with the same passing threshold of Not published. Browse mode lets you read every Q&A statically.