CertQuests
Why the PMLE exam was updated for 2026
The original Professional Machine Learning Engineer exam was designed when BigQuery ML, AI Platform Training, and Kubeflow Pipelines were the standard toolset. Google Cloud has since consolidated nearly all ML workloads onto Vertex AI, which now unifies training, serving, feature engineering, pipeline orchestration, and model monitoring under a single managed platform. Candidates who prepared against 2022–2023 study material would arrive at the exam knowing AI Platform Training but lacking the Vertex AI Pipelines, Vertex AI Feature Store, and Vertex AI Model Monitoring knowledge that modern production deployments require.
The second driver is generative AI. Vertex AI Studio and Model Garden entered general availability on Google Cloud in 2023 and have since become core infrastructure for enterprise AI deployments. The updated PMLE exam reflects this by adding content on prompt engineering, fine-tuning foundation models, evaluating LLM outputs with BLEU/ROUGE/human preference metrics, and integrating the Gemini API into ML workflows. These topics were entirely absent from previous exam versions.
The third driver is responsible AI maturity. Google Cloud has significantly expanded its responsible AI tooling — Explainable AI, Model Cards, the What-If Tool, and new fairness evaluation integrations in Vertex AI Experiments. The updated exam now tests candidates on applying these tools in real deployment contexts: detecting model drift that disproportionately affects demographic subgroups, documenting model limitations through Model Cards, and setting up monitoring pipelines that surface fairness regressions before they reach end users.
Exam at a glance
Professional Machine Learning Engineer
| Exam code | Professional Machine Learning Engineer |
| Questions | 60 multiple choice / multiple select |
| Duration | 2 hours |
| Pass score | Not published (performance-based bands) |
| Cost | $200 USD |
| Validity | 2 years |
| Prerequisite | None (3+ years ML experience recommended) |
| Delivery | Remote proctored or in-person testing centre |
| Languages | English, Japanese |
The six exam domains
The PMLE exam guide organises content into six weighted domains. Google publishes the weighting in approximate percentage bands rather than exact figures, and the distribution shifts slightly between exam versions. The 2026 weighting places the most emphasis on MLOps and model serving — a clear signal that production reliability, not training accuracy, is the primary lens through which ML engineers are now evaluated.
Domain 1 — Architecting Low-Code ML Solutions (~12%)
This domain tests whether candidates can select the right Google Cloud managed service for a given ML problem without writing training code. Key topics: BigQuery ML for in-database model training (linear regression, k-means, ARIMA+, XGBoost, TensorFlow), AutoML on Vertex AI for structured data, image, text, and video classification tasks, and Vertex AI Vision for pre-built vision APIs versus custom training decisions. Candidates must be able to identify when a managed solution is cost-effective and appropriate versus when a custom training pipeline is required — the exam regularly tests this decision boundary with constraint scenarios (latency requirements, dataset size, labelling budget, retraining frequency).
Domain 2 — Collaborating to Manage Data and Models (~16%)
This domain covers the organisational and tooling layer around ML work: Vertex AI Feature Store for sharing and reusing features across teams, Vertex AI Experiments for tracking hyperparameter search and run comparisons, Vertex AI Model Registry for version control and lineage, and the Data Catalog for dataset discovery and governance. The 2026 update added content on Vertex AI Model Cards and documentation best practices for cross-team model handoff. Candidates must understand RLS (row-level security) in BigQuery feature pipelines, how to use Vertex AI Experiments to reproduce a prior training run from metadata alone, and how to design a feature store schema that supports both real-time serving and batch training without duplication.
Domain 3 — Scaling Prototypes into ML Models (~18%)
This is the core model training domain: distributed training strategies (data parallelism, model parallelism, parameter server vs. all-reduce), hardware selection (TPU v4/v5 vs. GPU A100/H100 for different model sizes and batch patterns), Vertex AI Training custom containers, hyperparameter tuning with Vizier, and handling imbalanced datasets at production scale. The 2026 update significantly expanded generative AI fine-tuning content — supervised fine-tuning (SFT) and parameter-efficient fine-tuning (PEFT) techniques including LoRA and prompt tuning on PaLM 2 and Gemini models via Vertex AI. Candidates must understand when full fine-tuning is warranted versus PEFT, how to manage the training dataset format requirements for Gemini SFT jobs, and how to evaluate fine-tuned model quality using automated metrics before deploying to Model Registry.
Domain 4 — Serving and Scaling Models (~20%)
This domain tests production serving infrastructure: Vertex AI Endpoints (dedicated vs. shared), online prediction latency optimisation, batch prediction jobs for high-throughput offline inference, model compression techniques (quantisation, pruning, distillation) for latency-constrained deployments, and Vertex AI Prediction with custom serving containers. The 2026 update added Vertex AI Model Garden and the Gemini API as serving primitives — candidates must understand when to call the Gemini API directly versus deploying a fine-tuned model to a Vertex AI Endpoint, how to manage quotas and rate limiting for high-traffic generative AI applications, and how grounding with Vertex AI Search reduces hallucination in production RAG pipelines. Multi-model serving configurations and A/B traffic splitting via Vertex AI Endpoints are heavily tested.
Domain 5 — Automating and Orchestrating ML Pipelines (~22%)
The highest-weighted domain in the 2026 exam. Vertex AI Pipelines (built on Kubeflow Pipelines v2 with the KFP SDK) is the central subject: component authoring, pipeline compilation, conditional and parallel execution, artifact lineage, and integration with Cloud Scheduler and Cloud Build for CI/CD-triggered retraining. This domain also covers Vertex AI Feature Store online serving latency and offline batch source configuration, event-driven pipeline triggers via Eventarc and Pub/Sub, and cross-project pipeline sharing. The 2026 update added content on integrating generative AI steps into ML pipelines — specifically, calling the Gemini API as a pipeline component for text augmentation, classification, or evaluation steps, and chaining these with traditional training components in a single Vertex AI Pipeline run.
Domain 6 — Monitoring, Optimising, and Maintaining ML Solutions (~12%)
Vertex AI Model Monitoring for feature skew and prediction drift detection, alerting thresholds and notification channels, data drift analysis using Jensen-Shannon divergence, and retraining trigger logic. Responsible AI evaluation tools: Explainable AI (feature attributions via SHAP/IG, example-based explanations), the What-If Tool for counterfactual analysis, and fairness metric configuration for protected groups. The 2026 update added content on evaluating generative AI outputs in monitoring pipelines — automated ROUGE/BLEU scoring for summarisation tasks, LLM-as-a-judge patterns for open-ended generation quality, and detecting output distribution shift in deployed generative AI applications. Candidates must understand how to set up a full monitoring loop from model deployment through alert to automated retraining pipeline trigger.
What changed from the 2023 exam version
Candidates who studied for the earlier PMLE exam versions will find roughly 55–65% of the core content transferable. The training fundamentals, BigQuery ML, and Kubeflow Pipelines knowledge carries over. The areas requiring fresh dedicated study are:
- Vertex AI as the unified platform — AI Platform Training, AI Platform Prediction, and the older Kubeflow on GKE deployment patterns are deprecated in the new exam. All training and serving content is now Vertex AI-native.
- Generative AI and foundation models — Vertex AI Studio, Model Garden, Gemini API, prompt engineering, SFT and PEFT fine-tuning, RAG with Vertex AI Search. Entirely absent from the 2023 exam.
- Vertex AI Feature Store v2 — the 2026 exam covers the redesigned Feature Store with online serving via Bigtable backends, not the original Featurestore API.
- Responsible AI depth — Model Cards, fairness subgroup monitoring, and LLM output evaluation are now explicit exam topics rather than background context.
- LLM evaluation in pipelines — automated scoring of generative outputs as a monitoring step is new and carries meaningful question weight in Domain 6.
If you have hands-on Vertex AI experience from 2024 or later, expect 3–4 weeks of focused study. If your GCP ML experience predates Vertex AI's current form, plan for 6–8 weeks — the platform architecture has shifted substantially enough that old hands-on familiarity with AI Platform Training can create false confidence in the exam.
The Google Cloud ML certification path
The PMLE sits at the top of the Google Cloud ML track and is the natural follow-on to the Professional Data Engineer credential for practitioners shifting toward production ML.
The PMLE has no formal prerequisite, but candidates without Professional Data Engineer knowledge frequently struggle with the BigQuery ML, Dataflow feature pipeline, and data governance sections of the exam. The recommended preparation path is ACE or PDE before attempting PMLE — not because the exam tests ACE/PDE content directly, but because the practical GCP literacy they build significantly reduces the cognitive load of learning Vertex AI-specific concepts on top of unfamiliar infrastructure.
Who should certify in 2026
The PMLE is aimed squarely at ML engineers and MLOps practitioners who work on production systems — not data scientists whose primary activity is exploratory analysis, and not software engineers who interact with ML outputs via APIs without operating the training pipelines. The ideal candidate trains and deploys models in Vertex AI, manages automated retraining pipelines with Vertex AI Pipelines, monitors model health in production, and is beginning to incorporate generative AI capabilities into existing ML workflows.
In the hiring market, the PMLE has strong recognition among Google Cloud customers and GCP-focused consulting firms. It carries less universal recognition than AWS MLS-C01 or the Microsoft AI Engineer Associate outside of GCP-heavy shops, but within organisations standardised on Google Cloud it is the expected credential for senior ML engineering roles. The 2026 update's emphasis on generative AI content has increased its relevance for teams building LLM-powered applications on Vertex AI, making it useful even for engineers who have not historically worked on traditional ML pipelines.
Google Cloud's official preparation path is the Preparing for the Google Cloud Professional Machine Learning Engineer Exam learning path on Google Cloud Skills Boost (formerly Qwiklabs), which includes labs covering Vertex AI Pipelines, AutoML, and Model Monitoring. The Machine Learning on Google Cloud specialisation on Coursera provides deeper conceptual grounding. For hands-on practice with generative AI content, the Vertex AI Studio quickstart labs and the Gemini API in Vertex AI codelabs cover the new exam topics directly. Set up a personal GCP project with free trial credits and build at least one end-to-end Vertex AI Pipeline that trains a custom model, registers it to Model Registry, deploys to an Endpoint, and triggers retraining on drift detection — this exercise covers all six exam domains in a single workflow. CertQuests has a dedicated GCP Professional ML Engineer practice pack covering all updated domains.
Practice GCP Professional ML Engineer questions on CertQuests — free, no sign-up required.
Start Practicing →