AI Wisdom

Google Cloud professional level

Google Cloud Professional Machine Learning Engineer

Code
PMLE
Length
120 minutes
Price
$200 plus tax
Valid for
2 years
Study time, our estimate
80 to 120 hours

A broad exam that covers classic ML and gen AI side by side, from BigQuery ML to Model Armor. It rewards knowing which Google product fits a constraint more than deep math, and the guide itself says it does not directly assess coding.

Take it if

You run ML on Google Cloud. It suits engineers who deploy and operate models.

Skip it if

You only work with foundation model APIs. Much of this is classic ML operations.

Before you start

Check yourself against this list first. If something is missing, close that gap before you pay for an exam.

  • Three years in industryGoogle recommends this, with one year on Google Cloud.
  • Python and SQLEnough to read code snippets. Coding itself is not directly assessed.
  • ML basicsKnow metrics, model types and overfitting.

Helpful first: GAIL

All exam details
Format
50 to 60 multiple-choice questions
Length
120 minutes
Price
$200 plus tax
Valid for
2 years
Languages
English, Japanese
Experience
3+ years in industry, 1+ year on Google Cloud
Guide date
June 1, 2026
If you fail
Professional: wait 14, 60, then 365 days

What the exam covers

  • Architecting low-code AI solutions~13%
  • Collaborating within and across teams to manage data and models~16%
  • Scaling prototypes into ML models~21%
  • Serving and scaling models~20%
  • Automating and orchestrating ML pipelines~18%
  • Monitoring AI solutions~13%

What to focus on

  • Picking the product: BigQuery ML, AutoML, custom training, or a Model Garden model.
  • Hardware: CPU, GPU and TPU, plus data and model parallelism for distributed training.
  • Serving: batch or online, Cloud Run or GKE, canary and A/B rollouts.
  • Pipelines: Agent Platform Pipelines, Airflow, Ray, and keeping preprocessing the same in training and serving.
  • Monitoring: training-serving skew, data drift, concept drift, and feature attribution drift.
  • Gen AI: LLM-as-a-judge evaluation, tuning Gemini, and Model Armor.
Everything on one mapEvery topic and key idea in this certification.
mindmap
  root((PMLE))
    Low-code and collaboration
      BigQuery ML
      AutoML
      Feature Store
    Scaling prototypes
      Hardware
      Parallelism
      Tuning
    Serving
      Batch or online
      Rollout
      Targets
    Pipelines
      Orchestration
      Consistency
      CI/CD/CT
    Monitoring and risk
      Model Monitoring
      Model Armor
      LLM-as-a-judge
Which drift is it?Monitoring questions turn on these distinctions.
flowchart TD
  A{"What changed?"} -->|"Serving differs from training"| B["Training-serving skew"]
  A -->|"Input distribution"| C["Data drift"]
  A -->|"Meaning of inputs for the outcome"| D["Concept drift"]
  A -->|"Which features matter"| E["Feature attribution drift"]

Lesson 1 of 5, ~13% + ~16% of the exam

Low-code and collaboration

Nearly 30% is picking the simplest tool that works.

BigQuery ML
Train with SQL where the data already lives.
AutoML
Managed training on Agent Platform.
Feature Store
Define features once and reuse them.
How much do you need to build?Choose the simplest tool that meets the need.
flowchart TD
  A{"What fits the task?"} -->|"A prebuilt API works"| B["Vision or Document AI API"]
  A -->|"Data is in BigQuery"| C["BigQuery ML"]
  A -->|"Tabular, little code"| D["AutoML"]
  A -->|"Full control"| E["Custom training"]

Common trap

Writing custom training when BigQuery ML meets the need.

Try it yourself

Train the same model in BigQuery ML and AutoML.

Official docs: ML in BigQueryFeature Store

Lesson 2 of 5, ~21% of the exam

Scaling prototypes

The largest domain.

Hardware
Choose among CPU, GPU and TPU.
Parallelism
Data parallelism copies the model. Model parallelism splits it across devices.
Tuning
Fine-tune for behavior, not for missing facts.
Data or model parallelismThe question is whether the model fits on one device.
flowchart LR
  subgraph D["Data parallelism"]
    D1["Full copy on GPU 1"]
    D2["Full copy on GPU 2"]
  end
  subgraph M["Model parallelism"]
    M1["Layers 1 to 6 on GPU 1"] --> M2["Layers 7 to 12 on GPU 2"]
  end

Common trap

Using data parallelism when the model does not fit on one device.

Try it yourself

Run one training job on CPU and on GPU and compare.

Official docs: Gemini Enterprise Agent Platform

Lesson 3 of 5, ~20% of the exam

Serving

A fifth of the exam.

Batch or online
Decide by how fast an answer is needed.
Rollout
A/B tests and canaries compare versions.
Targets
Agent Platform, Cloud Run and GKE.
Canary rolloutReal traffic, small risk.
sequenceDiagram
  participant U as Users
  participant E as Endpoint
  participant O as Current model
  participant N as New model
  U->>E: Requests
  E->>O: 90% of traffic
  E->>N: 10% of traffic
  N-->>E: Compare metrics
  E->>N: Increase the share if healthy

Common trap

Testing a new model only on held-out data.

Try it yourself

Deploy two versions and split the traffic.

Official docs: What is Cloud RunGemini Enterprise Agent Platform

Lesson 4 of 5, ~18% of the exam

Pipelines

Repeatable training is what separates a prototype from a system.

Orchestration
Agent Platform Pipelines, Airflow and Ray.
Consistency
Share preprocessing between training and serving.
CI/CD/CT
Automate retraining with a clear policy.
One preprocessing pathSharing the logic prevents skew.
flowchart LR
  A["Shared preprocessing"] --> B["Training pipeline"]
  A --> C["Serving path"]
  B --> D["Model"]
  D --> C
  C --> E["Predictions"]

Common trap

Coding preprocessing twice, which causes skew.

Try it yourself

Build a small pipeline with a validation step.

Official docs: Pipelines

Lesson 5 of 5, ~13% of the exam

Monitoring and risk

Covers drift and gen AI security together.

Model Monitoring
Sets continuous evaluation for production models.
Model Armor
Screens prompts and responses.
LLM-as-a-judge
Uses a model to evaluate gen AI output.
Screening a Gemini appModel Armor sits on both sides of the model.
sequenceDiagram
  participant U as User
  participant A as Model Armor
  participant G as Gemini
  U->>A: Prompt
  A->>A: Screen for malicious input and sensitive data
  A->>G: Clean prompt
  G-->>A: Response
  A-->>U: Screened response

Common trap

Calling every quality drop data drift.

Try it yourself

Add Model Armor to a Gemini app.

Official docs: Model MonitoringModel Armor

Study plan

  1. Weeks 1 and 2: low-code and data

    Train a model in BigQuery ML and in AutoML. Build features in the Feature Store and keep track of experiments.

  2. Weeks 3 and 4: training and serving

    Run custom training in a notebook, then on a pipeline. Deploy with a canary and compare the two versions.

  3. Week 5: monitor

    Set up Model Monitoring and work through each kind of drift. Add safety filters and Model Armor to a Gemini app.

  4. Exam week

    Work the official learning path's review, do the sample questions, then the mock below.

Exam day and after

  • Google renamed Vertex AI to Gemini Enterprise Agent Platform, and its exams were updated to match. Old notes use the former names, so check the exam guide for the current ones.
  • The certification pages do not state a pass mark for these exams.
  • Sample questions are free on each certification page, with no limit on attempts.
  • The exam favors Google Cloud native solutions, so choose the managed product unless the question rules it out.
Test yourself8 practice questions, scored by exam topic.
Start the mock exam

Keep going

Exam facts checked on October 31, 2026 against PMLE exam guide, June 1, 2026. Lessons are our own explanations, so use the official docs as the final word, and confirm price and dates before booking.