Why start with the exam guide
Before you write any BigQuery ML or deploy any endpoint, read the exam guide. Experienced engineers often fail this exam for a simple reason: they answer with what they'd build at work, not what the exam calls the best answer. This level gives you the map. It covers what each section tests, how much each is worth, which product names the guide uses now, and a routine for scenario questions.
Some ground rules. Google owns the Professional Machine Learning Engineer exam guide (as of June 1, 2026). This course doesn't reproduce it. You can read it free on Google's certification page, and you should keep it open alongside this course. This course is independent of Google and isn't endorsed by it. No pass is promised. At the time of writing, Google's exam page lists 2 hours, 50–60 multiple-choice and multiple-select questions, a $200 fee (plus tax), and English and Japanese. Check the page yourself before you book, because these details change.
One more point from the guide: the exam does not directly assess coding skills. You need basic Python and SQL so you can read the code snippets in some questions.
The six sections and their weights
The guide splits the role into six sections:
| # | Section | 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% |
Sections 3, 4 and 5 make up nearly 60% of the exam: training, hardware, serving and pipelines. Low-code AI and monitoring are the smallest sections, but at ~13% each they still cost you about 7 questions apiece if you skip them.
Worked example: budget your study time
Give each section study time in proportion to its weight. Here's the calculation as code:
# Approximate section weights from the exam guide (they sum to ~101% due to rounding)
weights = {
"1 Low-code AI": 13,
"2 Data & models collaboration": 16,
"3 Scaling prototypes": 21,
"4 Serving & scaling": 20,
"5 Pipelines": 18,
"6 Monitoring": 13,
}
study_hours = 60 # your total budget
questions = 55 # midpoint of the 50-60 range
total = sum(weights.values())
for section, w in weights.items():
share = w / total
print(f"{section:32} {share*study_hours:5.1f} h ~{share*questions:4.1f} questions")
Output:
1 Low-code AI 7.7 h ~ 7.1 questions
2 Data & models collaboration 9.5 h ~ 8.7 questions
3 Scaling prototypes 12.5 h ~11.4 questions
4 Serving & scaling 11.9 h ~10.9 questions
5 Pipelines 10.7 h ~ 9.8 questions
6 Monitoring 7.7 h ~ 7.1 questions
Practitioner tip: weight your weak spots more heavily too. If you serve models daily but have never set up drift monitoring, move hours from section 4 to section 6.
The names changed: a translation table
In April 2026 Google renamed and expanded Vertex AI as Gemini Enterprise Agent Platform. The exam guide uses the new names. Blog posts, Stack Overflow answers and older code still use the old ones. Learn both.
| Exam guide name | What you'll still see in docs, code, forums |
|---|---|
| Gemini Enterprise Agent Platform | Vertex AI |
| Agent Platform Workbench | Vertex AI Workbench |
| Agent Platform Pipelines | Vertex AI Pipelines |
| Agent Platform AutoML | Vertex AI AutoML |
| Agent Platform Model Registry | Vertex AI Model Registry |
| Agent Platform Feature Store | Vertex AI Feature Store |
| Managed Service for Apache Airflow | Cloud Composer |
The old name also survives in code. The Python SDK is still the google-cloud-aiplatform package:
# pip install google-cloud-aiplatform
from google.cloud import aiplatform # the 'aiplatform' name outlived the 'Vertex AI' brand
aiplatform.init(
project="my-ml-project", # your Google Cloud project ID
location="us-central1", # region where your models, endpoints and pipelines live
)
Gotcha: many documentation URLs still contain
/vertex-ai/even though the page titles say "Agent Platform". Also, Agent Registry (a catalogue of agents and MCP servers) is not the same thing as Agent Platform Model Registry (versioned ML models). They're easy to mix up under time pressure.
Scenario questions: finding the best answer
Most questions describe a company, a problem and some constraints, then give four answers that could all work. Your job is to pick the one that fits the constraints best. Use this routine:
- Read the last sentence first. It tells you what's being asked: train, serve, monitor or reduce cost.
- Find the binding constraint. Look for phrases like minimal operational overhead, lowest latency, team only knows SQL, cost-effective, data must not leave the region or PII.
- Rule out answers that work technically but break that constraint. For example, self-managing something on GKE breaks "minimal overhead".
- Prefer managed, Google Cloud-native services unless the scenario demands control. The exam was updated specifically to put Google Cloud-native solutions first.
- Choose the simplest answer that fully meets the requirements.
Worked example
A retail analytics team stores sales data in BigQuery. They know SQL but not Python. They need a demand forecast this quarter with minimal operational overhead. What should they do?
- A. Write a custom PyTorch model and train it on Agent Platform custom training. Eliminate: the team doesn't know Python.
- B. Export to Cloud Storage and run Kubeflow on GKE. Eliminate: too much infrastructure to manage.
- C. Build a forecasting model inside BigQuery with SQL. Keep: it's low-code, the data doesn't move, and the team already has the skills.
- D. Fine-tune an LLM from Model Garden. Eliminate: an LLM is the wrong kind of model for numeric forecasting.
All four could produce a forecast, but only C respects the constraints. You'll learn the BigQuery side of this in the next levels.
Trade-off to remember: in real projects you might choose the custom option for flexibility later. On the exam, the constraints given in the question decide the answer. Don't add requirements the question doesn't state.
What the exam asks about this
No question tests the weightings or the renames directly. But every question uses the new names, sometimes mixed with old ones in the answers, and almost every question is a scenario where you pick the best answer by its constraints. Being fluent in both sets of names and having an elimination habit will earn you points in all six sections.
Key takeaways
- The guide is Google's and free to read on its certification page. Check the exam page for current length, price and languages.
- The weights are ~13 / 16 / 21 / 20 / 18 / 13%. Sections 3–5 make up nearly 60%, so plan your study time to match.
- Gemini Enterprise Agent Platform = formerly Vertex AI. Workbench, Pipelines, AutoML and Model Registry are now prefixed "Agent Platform". The SDK is still
google-cloud-aiplatform. - Find the binding constraint, eliminate answers that work but don't fit, and prefer managed, native, simple solutions.
- You won't write code in the exam, but you'll need to read Python and SQL snippets.