THE DEV BENCH
🔵

Google Cloud Professional Machine Learning Engineer

One workspace for passing the Google Cloud PMLE — a professional-level exam about designing, building, productionizing, and monitoring ML on Vertex AI (now including generative AI and Gemini): framing & scaling prototypes, serving & scaling, pipeline automation, data & model management, low-code AI, and monitoring. Six section decks and an 88-question practice exam do the heavy lifting, authored against the official exam guide.

The honest read: this is a professional exam that tests judgment on operating ML on GCP, not the math of ML — it assumes real ML + Vertex AI experience and rewards picking the option that best fits the stated constraint (cost, latency, drift, governance), not just the technically correct one. Almost every question is “which Vertex AI / GCP service fits this scenario,” so the decks + exam here are the core study. Real hands-on Vertex AI needs a paid GCP project, so those labs point out honestly to the free Cloud Skills Boost labs + credits rather than pretending to host them.

Deep work · Google PMLEtotal
or log

The exam at a glance

verified July 2026
Cost$200 (+ tax)
Format50–60 questions · multiple choice / multiple select
Duration2 hours
Passing scoreNot published
Validity2 years
LevelProfessional

Google does not publish the PMLE passing score or a per-question scoring scheme; the 70% used in the practice exam here is a readiness benchmark, not the real cut. Certification validity is 2 years. Confirm current cost, format, section list, and policy on the official certification page before you book — the exam was refreshed for generative AI in 2026.

Certification roadmap

sequence
ACEAssociate Cloud EngineerOptional foundation

Associate · MCQ

Not required. Helps only if GCP itself (projects, IAM, networking) is unfamiliar. Skip if you already work in GCP day to day.

PMLEProfessional ML EngineerThe goal

Professional · 50–60 Q · 2 hours · 2-year validity · $200

The target of this workspace. The full ML lifecycle on Vertex AI — framing & scaling prototypes, serving & scaling, pipeline automation, managing data & models, low-code AI, and monitoring — plus generative AI. A professional-level exam: it assumes real ML + GCP experience and rewards judgment on trade-offs, not recall.

TrioTri-cloud MLOpsThe bigger story

AWS MLA · Azure AI-300 · Google PMLE

PMLE is the GCP leg of an 'MLOps across all three clouds' story alongside AWS MLA-C01 and Azure AI-300. Same concepts, three vendors' tooling — a strong portfolio signal. Both other paths have full workspaces here too.

Exam domains — weight your study by these

official weights
SectionWeightWhat it covers
Scaling prototypes into ML models~21%Framing the problem, choosing a build approach (custom training, AutoML, BigQuery ML, foundation models), the right framework & hardware (CPU/GPU/TPU), distributed & efficient training, hyperparameter tuning, and evaluating against the business goal.
Serving & scaling models~20%Online vs batch prediction, endpoints & deployed models, autoscaling and machine/accelerator sizing, traffic-split rollout & A/B, private endpoints (PSC/VPC-SC), and optimizing latency/throughput/cost — including serving generative-AI models.
Automating & orchestrating ML pipelines~18%Vertex AI Pipelines (KFP/TFX), components & artifacts, CI/CD with Cloud Build + Artifact Registry, triggers & scheduling, retraining loops, evaluation gates, and ML Metadata lineage.
Collaborating to manage data & models~15%Vertex AI Feature Store (no training/serving skew), managed Datasets, governance & lineage (Dataplex, BigQuery), IAM least-privilege, the Model Registry & metadata, PII/DLP, VPC-SC/CMEK, and reproducible team workflows.
Architecting low-code AI solutions~13%BigQuery ML, Vertex AI AutoML, pre-trained APIs (Vision, Natural Language, Speech, Translation, Document AI), Gemini & Model Garden, and Vertex AI Search / agent-building — knowing which low-code option fits which problem.
Monitoring AI solutions~13%Vertex AI Model Monitoring (drift & skew), endpoint observability (Cloud Monitoring/Logging), retraining triggers, Vertex Explainable AI, responsible-AI & safety for GenAI, and cost/performance monitoring.

Scaling prototypes (~21%) + serving/scaling (~20%) + pipelines (~18%) are roughly 60% of the exam. Weights are from the official exam guide (July 2026); “Collaborating to manage data & models” is the ~15% remainder the guide omits an explicit figure for — confirm in the current guide. The 99 deck cards are distributed to match these weights.

Flashcard decks — the recall layer

6 decks · 99 cards
Scaling Prototypes into ML Models (~21%)21 cards — the biggest section. Framing the problem, the four build approaches (BigQuery ML, AutoML, custom training, foundation models), CPU/GPU/TPU choice, data- vs model-parallel training, Vizier hyperparameter tuning, and the metrics/overfitting controls that tie a model to the business goal. Start here; it's over a fifth of the exam.

The other five sections

Serving & Scaling Models (~20%)19 cards — online vs batch prediction, endpoints & deployed models, the traffic-split for canary/A-B, autoscaling & machine/accelerator sizing, private endpoints (Private Service Connect + VPC-SC), latency/cost optimization (quantization, distillation), and serving Gemini/foundation models.Pipelines & Orchestration (~18%)18 cards — Vertex AI Pipelines (KFP vs TFX), components & artifacts, CI/CD with Cloud Build + Artifact Registry, schedules & event triggers (Pub/Sub, Eventarc), evaluation gates, retraining loops, ML Metadata lineage, step caching, and Composer vs Vertex Pipelines.Data & Model Management (~15%)15 cards — Vertex AI Feature Store (online vs offline, no skew), managed Datasets, governance & lineage (Dataplex), IAM least-privilege, the Model Registry & ML Metadata, PII/DLP, VPC-SC & CMEK, Experiments, and project-per-environment separation.Low-Code AI Solutions (~13%)13 cards — BigQuery ML (train with SQL, ARIMA+ forecasting, remote models), Vertex AI AutoML, the pre-trained APIs (Vision, Natural Language, Speech, Translation, Document AI), Gemini & Model Garden, Vertex AI Search for RAG grounding, and the low-code-vs-custom decision.Monitoring AI Solutions (~13%)13 cards — Vertex AI Model Monitoring (training/serving skew vs prediction drift), data vs concept drift, model-quality monitoring with labels, Cloud Monitoring/Logging observability, Vertex Explainable AI (attribution drift), automated retraining loops, and responsible-AI/safety for GenAI.

Practice exam

live · 88-question pool
Google PMLE — Professional ML Engineer Practice ExamA fresh 50-question draw each attempt from a section-weighted 88-questionpool, with a 120-minute soft timer. Scenario-style “which Vertex AI / GCP service” MCQs with worked explanations and a per-section weak-area breakdown. Google doesn't publish a pass mark, so the 70% here is a readiness benchmark. Progress saves in your browser — no login.

Use the per-section breakdown to find weak areas, then close them with the matching deck. These are original questions in the exam's scenario style — a strong readiness gauge, not a brain dump of real items.

Hands-on Vertex AI — the honest note

external · needs a GCP project

PMLE is multiple-choice, so you can pass on the decks + exam + reading. But building the judgment — actually running a train → pipeline → deploy → monitor loop on Vertex AI, and standing up a BigQuery ML model or a grounded Gemini app — needs a real GCP project, which we can't host for you. Rather than fake it, this path points you at the real thing:

Cost watch:deployed Vertex AI endpoints and Workbench notebook instances bill by the hour even when idle, and Gemini bills per token — the top surprise bills. Undeploy models, stop notebooks the moment you're done, and set a budget alert.

Skill map — where to practice each thing

the workspace
SkillPractice withStatus
Which Vertex AI / GCP service fits each job (all sections)The 6 PMLE decks (99 cards)ready
Exam-format scenario judgment & weak-area findingPMLE practice exam (draw 50, ~70% benchmark)ready
Hands-on Vertex AI (train → deploy → pipeline → monitor)Cloud Skills Boost labs + free GCP credits (external)external
The official section scope & question styleCloud Skills Boost ML Engineer path (external, free)external

A study routine

  1. 1.Learn Vertex AI as the SPINE — training, pipelines, endpoints, Feature Store, monitoring, Model Garden. Most scenarios resolve to 'which Vertex AI capability fits.' Do the Scaling Prototypes deck first (~21%).
  2. 2.Front-load the two biggest sections — scaling prototypes (~21%) and serving/scaling (~20%) — then pipelines (~18%). Together they're roughly 60% of the exam.
  3. 3.PMLE is JUDGMENT-heavy (professional level): practice picking the option that best fits the stated constraint — cost, latency, drift, governance, freshness — not just the technically-correct one. The scenario's single constraint usually decides.
  4. 4.Drill the recurring 'which one' decisions: online vs batch prediction, AutoML vs BQML vs custom, KFP vs TFX, RAG vs fine-tuning, skew vs drift, private vs public endpoint. These are the backbone of the question set.
  5. 5.Use the practice exam's per-section breakdown to find weak areas, then close them with the matching deck. Google doesn't publish a pass mark, so treat 70% here as a readiness benchmark — hit it three times before you book.

Curated resources — free-first

verified July 2026

Everything you need to LEARN the PMLE is free (the official exam guide + Cloud Skills Boost path + the Vertex AI and BigQuery ML docs). Only the exam voucher ($200) and real GCP usage cost money. Every link checked live (July 2026); prices/versions shift, so treat costs as “as seen” — and re-check the exam guide, since the exam was refreshed for GenAI in 2026.

Start the practice exam