Certification guide
PMLEProfessional Machine Learning Engineer: the honest guide
Professional Machine Learning Engineer is an MLOps exam wearing a machine learning badge. Scaling prototypes, serving and scaling models, and automating pipelines are 59 percent of it between them, while building models with low-code tools is 13 percent.
That weighting is the most useful thing to know before preparing. Candidates who arrive as data scientists usually know the modelling and lose marks on the production half; candidates who arrive as engineers find the opposite.
Who ML Engineer is for
A good fit if
- You put models into production on Google Cloud and own what happens to them afterwards.
- You are a data scientist whose organization now expects you to ship, monitor and retrain rather than hand over notebooks.
- You are a platform or DevOps engineer supporting ML teams and need the vocabulary and the failure modes.
- Your organization needs a certified ML engineer for a Google Cloud partner specialization.
Probably not, if
- You want to learn machine learning. This exam assumes the theory and asks about operations, hardware choice and serving patterns.
- Your work stops at the notebook. More than half the curriculum is about what happens after a model is trained.
- You do not use Google Cloud. Vertex AI's shape is the subject, and the operational concepts transfer better than the specifics do.
Is ML Engineer worth it?
For an engineer responsible for models in production, this is one of the few certifications that tests the actual operational problem: retraining, skew, serving cost and pipeline automation rather than model architecture.
It is also the Google exam whose material dates fastest, which cuts both ways. The certification is valid two years and the platform will have moved inside that, so the value is highest when it reflects work you are doing now.
It is a poor fit for a research-oriented data scientist. Much of the curriculum is infrastructure, and a candidate who does not deploy will be studying somebody else's job.
What the exam actually asks you to do
Multiple choice and multiple select, in Google's own words. No labs, no console access, and nothing to configure during the exam.
Item formats
- Multiple choice
- Multiple response
Nothing here needs a lab. Reading carefully and eliminating options is the whole skill. Google Cloud, Professional Machine Learning Engineer exam details ↗
Domain breakdown and official weightings
From the official Google Cloud exam guides. Scaling prototypes into ML models is the heaviest domain at 21 percent, followed by Serving and scaling models at 20 percent.
- Architecting low-code AI solutions13%
- Collaborating within and across teams to manage data and models16%
- Scaling prototypes into ML models21%
- Serving and scaling models20%
- Automating and orchestrating ML pipelines18%
- Monitoring AI solutions13%
Where to focus: The operational half is where people underprepare. Serving, pipelines and monitoring together are 51 percent of the exam, against 34 percent for building and scaling models. Google's own section weights sum to 101 rather than 100, so treat every figure as approximate.
Study plans by experience level
ML engineer shipping on Vertex AI
6 to 8 weeksat 6 hours
- 1Week 1: read the exam guide against your own delivery process and mark the stages somebody else owns. Those are the gaps.
- 2Weeks 2 to 4: scaling prototypes and serving, 41 percent between them. Hardware selection, distribution strategy, online versus batch serving and the cost of each.
- 3Weeks 5 to 6: pipelines and automation at 18 percent, built end to end with a retraining trigger even if the model is trivial.
- 4Weeks 7 to 8: monitoring at 13 percent, including drift and skew detection, then a full pass through the guide.
Data scientist moving to production
10 to 12 weeksat 8 hours
- 1Weeks 1 to 2: the platform. Vertex AI's components, what each one replaces in a notebook workflow, and where cost accrues.
- 2Weeks 3 to 6: serving and scaling, which is the largest gap for this audience: endpoints, batch prediction, autoscaling and latency budgets.
- 3Weeks 7 to 9: pipelines and automated retraining, built once from nothing.
- 4Weeks 10 to 12: monitoring, low-code tools including BigQuery ML, then the exam guide worked line by line.
Platform engineer supporting ML teams
10 weeksat 8 hours
- 1Weeks 1 to 3: enough machine learning to reason about the operations. What training, evaluation and inference cost, and why a model degrades.
- 2Weeks 4 to 6: the pipeline and serving sections, which map closely onto delivery infrastructure you already understand.
- 3Weeks 7 to 8: experiment tracking and the collaboration section, where the exam expects team workflow rather than tooling trivia.
- 4Weeks 9 to 10: monitoring and risk, then the guide end to end.
Common mistakes
- Leaving an endpoint deployed
- A Vertex AI endpoint bills for its node whether or not it serves a prediction. Undeploying the model at the end of each session is the difference between a trivial lab bill and one that exceeds the $200 exam fee.
- Studying modelling instead of operations
- Low-code model building is 13 percent of the exam. Scaling, serving, pipelines and monitoring are 72 percent. A preparation plan weighted toward algorithms is weighted against the exam.
- Never building a retraining pipeline
- Automating pipelines and retraining is 18 percent, and it is the section that most reliably separates candidates who have shipped from those who have not. One pipeline that retrains on a schedule teaches the whole section, and the model itself can be trivial.
- Ignoring cost as an answer
- Many questions describe a constraint on latency, budget or hardware and expect the answer that respects it. Candidates who reach for the most capable accelerator by reflex miss questions that were about restraint.
What comes after passing
Two years of validity with a 60-day renewal window, which on this subject is genuinely short: the Vertex AI surface moves faster than any other area Google certifies.
The post-pass discount code applies to a renewal attempt.
Professional Data Engineer is the natural neighbour, and the two share the pipeline, monitoring and cost material.
The durable knowledge is the failure catalog: training-serving skew, drift, feedback loops and the way a model silently degrades. That survives every platform rename.
Costs across the full renewal cycle are on the ML Engineer cost page.
Frequently asked questions
Is this a machine learning exam or an MLOps exam?
Mostly MLOps. Scaling prototypes, serving and scaling models, and pipeline automation account for 59 percent of the exam, while building models with low-code tools is 13 percent. It assumes machine learning knowledge and tests what happens to a model after it exists.
How do I keep the lab cost down?
Undeploy endpoints at the end of every session and keep training runs small. Endpoint nodes bill continuously, which is the main way people overspend here, and the exam tests hardware and serving decisions rather than the size of anything you trained.
Do I need to be a strong programmer?
You need to read Python comfortably and reason about pipeline and serving code. The exam does not ask you to write production code under time pressure; it asks which approach fits a stated latency, cost or scale constraint.
How long is the certification valid?
Two years, with the renewal window opening 60 days before expiry. Given how quickly this platform area changes, treat renewal as real study rather than a formality.
Keep reading
- ML Engineer practice questions
Free sample questions with the full explanation on every answer.
- Free ML Engineer practice test
10 original questions, playable now. No account, no card.
- ML Engineer exam objectives
The full official blueprint, with practice pages on covered objectives.
- ML Engineer exam format
How many questions, how long, what the items look like, and how long it stays valid.
- ML Engineer passing score
The exact cut score, what kind of number it is, and the retake terms.
- How hard is ML Engineer?
An honest difficulty read from the format, the clock and the weights.
- What ML Engineer costs
The voucher price, the retake, and what renewal costs across the cycle.
Practice ML Engineer for free while you decide
Original questions written from the published objectives, with the concept, the reasoning, and a note on every wrong option. No account needed to start.
Start free ML Engineer questions