- How the Official Exam Guide Is Organized
- Domain Weights at a Glance
- Domain 1: Architecting Low-Code AI Solutions
- Domain 2: Collaborating Within and Across Teams
- Domain 3: Scaling Prototypes into ML Models
- Domain 4: Serving and Scaling Models
- Domain 5: Automating and Orchestrating ML Pipelines
- Domain 6: Monitoring AI Solutions
- What the Questions Look Like
- Sequencing Your Preparation Around the Weights
- Logistics That Touch the Domains
- Frequently Asked Questions
- The Professional Machine Learning Engineer exam guide lists six sections, weighted approximately 13/16/21/20/18/13%.
- Scaling prototypes into ML models (~21%) and Serving and scaling models (~20%) together carry roughly two-fifths of the exam.
- The June 1, 2026 guide uses Gemini Enterprise Agent Platform terminology and covers Model Armor and LLM-as-a-judge.
- Scoring is a single pass/fail result across the whole exam, so a weak domain can be offset by strength elsewhere.
How the Official Exam Guide Is Organized
Google Cloud publishes a five-page Professional Machine Learning Engineer exam guide, and the version current as of June 1, 2026 is the definitive statement of what can be tested. It divides the exam into six sections. Those six headings describe what is examined; they are not a training curriculum and do not tell you in what order to learn things. That distinction matters, because many candidates read the list top to bottom and assume it mirrors a course syllabus.
The 2026 guide is also noticeably different from older editions. Earlier versions opened with a section on framing ML problems; that heading is not part of the current structure, and you should not build a plan around it. The current guide uses Gemini Enterprise Agent Platform terminology and adds updated generative-AI material on security and evaluation, including Model Armor and LLM-as-a-judge. If you are working from a study resource written a year or two ago, check its vocabulary against the current guide before trusting its domain map.
This article walks through each of the six sections, explains what candidates should be able to do in each, and shows how the weights should shape your preparation. For a broader plan that includes timelines and resources, see the PMLE Study Guide 2026: How to Pass on Your First Attempt. If you are new to the credential itself, What Is PMLE Certification? covers the basics.
Domain Weights at a Glance
Google publishes approximate weights for each section. As published, they are approximately 13%, 16%, 21%, 20%, 18% and 13%, which sums to 101% by simple arithmetic. That is a rounding artifact of approximate figures, and it is best to preserve the numbers as published rather than normalize them. It is equally unwise to convert them into an exact count of questions per domain: the exam draws 50-60 questions, the split between scored and unscored items is not disclosed, and the weights describe emphasis, not a fixed allocation.
| Domain | Section name | Approximate weight |
|---|---|---|
| 1 | Architecting low-code AI solutions | ~13% |
| 2 | Collaborating within and across teams to manage data and models | ~16% |
| 3 | Scaling prototypes into ML models | ~21% |
| 4 | Serving and scaling models | ~20% |
| 5 | Automating and orchestrating ML pipelines | ~18% |
| 6 | Monitoring AI solutions | ~13% |
Domain 1: Architecting Low-Code AI Solutions (~13%)
This is the lightest section by weight, but it is the one most likely to catch experienced practitioners off guard. Many working ML engineers write custom training code every day and have little reason to touch managed, low-code tooling. The exam, however, expects you to recognize when a managed approach is the better engineering decision.
What candidates must understand
The core skill is selection: given a business requirement, a data situation and a team profile, choose the right level of abstraction.
- When a prebuilt or pretrained capability is sufficient versus when custom training is justified
- When a managed AutoML-style workflow fits a team without deep ML expertise
- How to assemble solutions from managed services, including foundation models and generative capabilities, without writing a full training stack
- How data location, format and volume constrain which low-code path is viable
Questions in this area tend to be scenario-driven. You are given a team, a dataset and a constraint such as time-to-market, limited labeled data or a requirement for explainability, and asked which approach best fits. The trap is over-engineering: a custom model is rarely the best answer when the scenario describes a small team that needs a working result quickly.
Because this domain is smaller, it is a sensible place to bank points efficiently. A few focused sessions on the managed-service landscape usually pay off more per hour than the same time spent on a domain you already know well.
Domain 2: Collaborating Within and Across Teams to Manage Data and Models (~16%)
The title of this section signals something that surprises candidates who expect a purely technical exam: a meaningful share of the content concerns how data and models are shared, governed and handed off between roles. In real organizations, data engineers, ML engineers, analysts and application developers all touch the same assets, and the exam tests whether you can set that up sensibly.
Topics to master
Think about the lifecycle of data and model artifacts from ingestion through reuse.
- Exploring, preparing and organizing data so multiple teams can use it consistently
- Feature management and reuse across projects, and avoiding training/serving skew
- Tracking experiments, versioning models and keeping lineage so results are reproducible
- Access control and governance for datasets and models shared across teams
- Handling sensitive data responsibly, including appropriate privacy and compliance practices
Expect questions that read like workplace situations: two teams are producing inconsistent features, a model cannot be reproduced, or an analyst needs access to training data without exposure to sensitive fields. The right answer is typically the one that creates a repeatable, governed process instead of a one-off fix. Minimum Python and SQL proficiency helps here, because some questions include code snippets or queries you need to interpret, though the exam does not include a direct coding task.
Domain 3: Scaling Prototypes into ML Models (~21%)
This is the heaviest section at approximately 21%, and it is the heart of what an ML engineer does: taking something that works in a notebook and turning it into a robust, trainable, production-quality model. If you are allocating study hours by weight, this is where the largest share belongs.
What the section rewards
The emphasis is on training and development at realistic scale, not on toy examples.
- Choosing model architectures and approaches appropriate to the data type and the business objective
- Training at scale, including distributed training and the use of accelerators, and understanding when each is warranted
- Hyperparameter tuning and systematic experimentation
- Diagnosing overfitting, underfitting and data problems, and selecting suitable evaluation metrics
- Adapting and fine-tuning foundation models, and evaluating generative outputs, including LLM-as-a-judge approaches
The exam does not ask you to derive algorithms. It asks you to make sound engineering choices: which metric reflects the business cost of errors, how to handle class imbalance, how to reduce training time without sacrificing quality, or when fine-tuning is preferable to prompting. A candidate who understands the reasoning behind these decisions will do better than one who has memorized product names.
For a sense of how demanding this material is relative to the rest, How Hard Is the PMLE Exam? Complete Difficulty Guide 2026 discusses where candidates tend to struggle.
Domain 4: Serving and Scaling Models (~20%)
The second-largest section covers what happens after training: getting predictions to users reliably and economically. Together with Domain 3, it accounts for roughly two-fifths of the exam, so these two sections deserve the bulk of your attention.
Core decisions
Serving questions revolve around trade-offs between latency, throughput, cost and operational simplicity.
- Batch versus online prediction, and how the access pattern drives the choice
- Deploying models to managed endpoints, configuring autoscaling and choosing hardware
- Optimizing for latency and cost, including model optimization and right-sizing resources
- Safe rollout strategies such as staged traffic splitting and rollback
- Securing generative applications, including guarding against prompt injection and unsafe outputs with tools such as Model Armor
A typical question describes a workload, perhaps spiky traffic with strict response-time targets or a large nightly scoring job, and asks for the configuration that meets the requirement at the lowest reasonable cost. Read the constraints carefully. Words like "real time," "millions of records overnight," or "unpredictable bursts" almost always decide the answer.
The inclusion of Model Armor in the current guide reflects a real shift: serving a model is no longer only about performance but also about protecting the application around it. Know what category of risk each protective control addresses, and be able to place it in a deployment architecture.
Domain 5: Automating and Orchestrating ML Pipelines (~18%)
At approximately 18%, pipeline automation is the third-largest section. The central idea is that repeatable, versioned, automated workflows beat manual steps, and the exam tests whether you can design them.
What to be able to do
Move from ad hoc scripts to orchestrated, testable pipelines.
- Break an ML workflow into discrete, reusable components with defined inputs and outputs
- Orchestrate training, evaluation and deployment steps, including conditional logic such as deploying only if a metric threshold is met
- Apply CI/CD ideas to ML, covering code, data and model changes
- Trigger retraining on a schedule or in response to events and drift signals
- Maintain reproducibility through versioned artifacts, metadata and lineage
These questions often present a manual, error-prone process and ask how to productionize it. The best answers usually introduce automation, validation gates and artifact tracking rather than adding more people to review steps by hand. Because this section connects tightly to Domains 2, 4 and 6, studying it well also reinforces material elsewhere on the exam.
Domain 6: Monitoring AI Solutions (~13%)
The final section shares the lowest weight with Domain 1, but it is a common source of lost points because candidates treat monitoring as an afterthought. A model that is not watched will degrade silently, and the exam expects you to know how to catch that.
Areas of focus
Monitoring spans both classical ML signals and generative-AI behavior.
- Detecting data drift, concept drift and training/serving skew, and deciding when to retrain
- Tracking prediction quality, latency, errors and resource utilization in production
- Setting up alerting and logging so problems surface before users report them
- Monitoring generative applications for quality, safety and misuse over time
- Fairness and explainability considerations, and responsible AI practices in production
A common scenario: a deployed model's accuracy has quietly declined, and you must identify the likeliest cause and the right response. Distinguishing a shift in the input distribution from a change in the underlying relationship between inputs and labels is a recurring skill. Equally important is knowing what you can and cannot measure without ground-truth labels, since labels often arrive late in real systems.
What the Questions Look Like
The standard exam contains 50-60 multiple-choice and multiple-select questions to be completed in two hours. That range is not a fixed question count, and Google does not disclose how many items are scored versus unscored, so avoid treating any single number as a target. The format is scenario-heavy: you are rarely asked to recall a definition and far more often asked to pick the best option among several plausible ones.
- Multiple-select items require you to choose every correct answer, so partial knowledge is risky; read the instruction on how many to select.
- Code-snippet questions require reading Python or SQL, not writing it. There is no direct coding assessment.
- "Best" answers mean several options may work; the correct one fits all stated constraints such as cost, effort and managed-service preference.
Google provides an official sample question set, but sample performance does not predict your result and does not establish what topics will appear, so use it to get comfortable with phrasing rather than to gauge readiness. For realistic scenario practice, the PMLE practice tests are built around the six domains above.
Sequencing Your Preparation Around the Weights
You do not need a generic schedule; you need one that reflects how these domains depend on each other. The ordering below is one reasonable approach for someone with moderate Google Cloud familiarity, and you should stretch or compress it to fit your experience.
Domains 3 and 2
- Start with model development at scale (~21%) because everything else builds on it
- Pair it with data and model management (~16%), since features, lineage and experiments feed training
Domains 4 and 5
- Cover serving and scaling (~20%) while training concepts are fresh
- Move to pipeline orchestration (~18%), which stitches training and serving together
Domains 6 and 1
- Study monitoring (~13%) once you know what is being deployed and automated
- Finish with low-code architecture (~13%) as a decision-making lens over the managed services you have seen
Mixed practice
- Take timed, mixed-domain sets and review every miss against its domain
Key Takeaway
Spend your largest blocks of time on Domains 3 and 4, but do not neglect the lighter sections. Because scoring is pass/fail on the whole exam, strong performance in the heavy domains can cushion a weaker one, yet a domain you ignore entirely still costs you marks you could have earned cheaply. The PMLE Cheat Sheet 2026 is useful for a final skim of key facts.
Logistics That Touch the Domains
A few practical facts shape how you prepare. Google recommends three-plus years of industry experience including at least one year designing and managing Google Cloud solutions, but these are recommendations rather than admission requirements; there is no mandatory degree, credential or training hours. Candidates must be at least 18. See PMLE Requirements 2026: Eligibility, Prerequisites & How to Qualify for the full picture.
The exam fee is $200 USD plus applicable tax, with no member/non-member distinction; the PMLE Certification Cost 2026: Complete Pricing Breakdown covers retakes and renewal pricing. Testing is delivered by Pearson VUE, at a center or through OnVUE remote proctoring, accessed through CM Connect. Kryterion's last testing date was February 22, 2026, and Pearson testing began March 2, 2026. The exam is closed book, with no notes, phones or AI assistance, and remote candidates must complete an identity check and keep the camera on continuously with no ordinary breaks. Cancel or reschedule at least 24 hours before your appointment. For scheduling mechanics, see PMLE Exam Dates 2026: Testing Windows, Deadlines & Scheduling.
If you do not pass, the paid-retake policy allows a maximum of four attempts in two years, with waits of 14, 60 and 365 days after successive failures. Unsuccessful standard-exam candidates can obtain section feedback, which maps neatly back onto the six domains and can direct your next round of study. The credential is valid for two years and renews through the standard exam. For what the scoring rules do and do not reveal, read PMLE Passing Score 2026: Exactly What You Need to Pass.
Frequently Asked Questions
Scaling prototypes into ML models is the largest at approximately 21%, followed by Serving and scaling models at approximately 20%. Automating and orchestrating ML pipelines comes next at approximately 18%.
The published weights are approximate (13/16/21/20/18/13%) and total 101% by arithmetic. They are best read as indicators of emphasis, not exact question counts, and should not be normalized or converted into a fixed number of questions per domain.
No. The result is pass/fail across the whole exam, not separate domain passes. Candidates who do not pass the standard exam can obtain section feedback to see where they were weaker.
No. There is no direct coding assessment. The exam uses multiple-choice and multiple-select questions, though some include Python or SQL snippets you need to interpret, so basic proficiency in both is helpful.
Yes. The guide as of June 1, 2026 uses Gemini Enterprise Agent Platform terminology and covers updated generative-AI security and evaluation, including Model Armor and LLM-as-a-judge. Older headings such as Framing ML problems should not be assumed to apply.