Google Cloud Professional Machine Learning Engineer (PMLE)

Learning plan

About 100 hours of focused study, split into 6 phases. Pick your weekly pace and the timeline adjusts.

Estimated finish: 10 weeks

What the exam covers

Source: official exam guide. Weights can change, so check before you start.

Before you start

• A Google Cloud project with billing enabled. New accounts get $300 in free credits, and the BigQuery sandbox needs no card
• Read-level Python and SQL. The exam doesn't test coding but shows snippets
• Know the rename: the June 1, 2026 guide uses Gemini Enterprise Agent Platform (and "Agent Platform") for what older material calls Vertex AI

1. Weeks 1–2 · 16 h

   Low-code ML and Google's foundation models

   • Build classification, regression, ARIMA_PLUS forecasting and k-means models in BigQuery ML, and use ML.PREDICT and TRANSFORM for feature engineering
   • Know when Agent Platform AutoML beats BigQuery ML, and when a pre-trained API (Document AI, Vision, Translation) beats both
   • Select models from Model Garden (Gemini, Imagen, Veo, open models), tune Gemini from BigQuery, and optimise Gemini apps for cost, latency and availability

   Hands-on: In the BigQuery sandbox, train a logistic-regression model on a public dataset (for example bigquery-public-data.ml_datasets.census_adult_income), inspect ML.EVALUATE and ML.FEATURE_IMPORTANCE, then train an ARIMA_PLUS forecast on a public time series and call ML.FORECAST.

   Use: Google Cloud Skills Boost – Machine Learning Engineer path · Introduction to BigQuery ML · BigQuery sandbox (free)

2. Weeks 2–4 · 16 h

   Data, notebooks and experiment tracking

   • Pick the preprocessing tool by scale (BigQuery SQL, Dataflow, Dataproc/Spark, in-memory pandas) and handle PII with Sensitive Data Protection
   • Create and serve features from Agent Platform Feature Store and explain how it prevents training-serving skew
   • Prototype in Agent Platform Workbench or Colab Enterprise with secure, shared settings (service accounts, VPC Service Controls, CMEK)
   • Track runs, metrics and lineage with Agent Platform Experiments and ML Metadata, and evaluate GenAI output with LLM-as-a-judge

   Use: Gemini Enterprise Agent Platform docs (formerly Vertex AI) · Best practices for ML on Google Cloud

3. Weeks 4–6 · 22 h

   Training at scale

   • Choose model type (ARIMA, DNN, tree ensemble, LLM) and product given cost, latency, interpretability and team skills
   • Run custom training jobs with prebuilt and custom containers, plus hyperparameter tuning jobs, and debug failed jobs from Cloud Logging
   • Choose CPU vs GPU vs TPU, and data vs model parallelism for distributed training; know Reduction Server and multi-worker strategies
   • Decide when to fine-tune a foundation model in Model Garden (supervised tuning, LoRA) versus prompting or grounding

   Hands-on: Package a scikit-learn or XGBoost training script, run it as an Agent Platform custom training job with a 5-trial hyperparameter tuning job, and log params and metrics to Experiments. Compare the best trial to your BigQuery ML baseline.

   Use: Google Cloud generative-ai samples (GitHub) · Vertex AI / Agent Platform samples (GitHub)

4. Weeks 6–8 · 26 h

   Serving and pipelines

   • Deploy to online endpoints (public and private) and batch prediction; serve from Cloud Run or GKE when you need custom runtimes
   • Version models in Model Registry and roll out with traffic splitting for A/B and canary tests
   • Build an Agent Platform Pipelines (Kubeflow Pipelines SDK) workflow with data validation, conditional deployment and caching; know when Managed Airflow or Ray fits instead
   • Set a retraining policy (scheduled vs drift-triggered) and automate CI/CD/CT with Cloud Build

   Hands-on: Turn your training job into a KFP pipeline (preprocess, train, evaluate, conditional upload to Model Registry, deploy) and run it on Agent Platform Pipelines. Send 10% of endpoint traffic to a second model version, then undeploy to stop billing.

   Use: MLOps: continuous delivery and automation pipelines in ML

5. Week 9 · 10 h

   Monitoring, security and responsible AI

   • Configure Model Monitoring for training-serving skew, data drift, concept drift and feature attribution drift, with alert thresholds
   • Use feature-based explanations (Sampled Shapley, Integrated Gradients) for explainability and bias checks
   • Protect LLM apps from prompt injection and data leakage with Model Armor, safety filters and regex screening

   Use: Model Armor · Gen AI evaluation service overview

6. Week 10 · 10 h

   Practice exams and gap review

   • Work through the official sample questions on the certification page, then a full 50–60 question timed practice exam in 2 hours
   • For every miss, write down which product the right answer used and why the runner-up was wrong (most questions are "choose the best service")
   • Translate any Vertex AI names in older practice material to the new Agent Platform names before reviewing

   Use: PMLE certification page (exam guide and sample questions) · Whizlabs PMLE practice tests

You're ready to book when…

• Given a scenario, you can choose between BigQuery ML, AutoML, custom training and a Model Garden FM in under a minute and defend the choice
• You can explain how Feature Store and a shared TRANSFORM step prevent training-serving skew
• You can name the monitoring signal for skew, data drift, concept drift and attribution drift and what triggers retraining
• You score 80% or higher on the official sample questions and 75% or higher on a full practice exam

Exam tips

• The June 2026 guide renamed Vertex AI products to Gemini Enterprise Agent Platform names, so expect the new names on the exam even though many courses still say Vertex AI.
• Google prefers its own managed, native services. When two answers work, pick the more managed Google Cloud option.
• Retakes: wait 14 days after a first fail, 60 days after a second and 365 days after a third, with at most four attempts in two years.