Google CloudUpdated 2026Last reviewed

Google Professional Machine Learning Engineer (PMLE) Practice Test

Build, Serve, and Monitor ML on Google Cloud.

500 scenario-based practice questions covering all 6 PMLE exam sections: low-code AI, data and model collaboration, scaling prototypes, serving, ML pipelines, and monitoring. Mapped to the official exam guide.

No credit card required · Upgrade to 500 questions for $11.99

500 Exam-Style Questions
7-Day Money-Back Guarantee

Is this for you?

You train models already but want practice deciding when BigQuery ML or Agent Platform AutoML is enough and when the task calls for custom training or a Model Garden model
You need to choose between tuning a Gemini model and using it as is, and want scenarios that weigh cost, latency and availability for Gemini-based applications
You are moving notebook prototypes from Agent Platform Workbench or Colab Enterprise into Agent Platform Pipelines and want practice keeping preprocessing consistent between training and serving
You run models in production and want to tell training-serving skew, data drift, concept drift and feature attribution drift apart, and pick the Model Monitoring setup for each

Not for you if:

You want to discuss generative AI strategy rather than build systems: the Generative AI Leader exam is aimed at business decisions, while PMLE expects you to train, deploy and monitor models
You want copies of real exam questions: these are original practice questions with explanations, not braindumps

The Google Professional Machine Learning Engineer (PMLE) is a professional-level Google Cloud certification. The exam has 50–60 questions, runs 120 minutes, costs $200, and requires a passing score of Not published by Google. CertSharp provides 500 exam-style PMLE practice questions — 30 free with no credit card — each with a detailed explanation.

Start 2–3 weeks before your exam for best results
Updated for 2026 exam blueprint
Detailed explanations for every question

Exam Overview

PMLE Exam Details

Know exactly what to expect on exam day

Questions

50–60

Duration

120 minutes

Passing Score

Not published by Google

Exam Cost

$200

Compare this fee with other certifications in our exam cost comparison, or check your readiness with the free PMLE readiness quiz.

Coverage

Exam Domains Covered

Every question maps to an official PMLE domain

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%)

From the exam guide

What the PMLE exam actually tests

The PMLE guide covers both conventional ML and generative AI built on foundation models, across the full path from low-code models to monitored production pipelines. It does not test coding directly, but questions may include Python or SQL snippets, and most ask you to pick the right product, hardware or rollout strategy for a stated constraint.

Picking the model and the product

Section 3.1 asks you to choose a model type such as ARIMA, a DNN or an LLM, and a product such as Agent Platform AutoML, BigQuery ML or Agent Platform Pipelines, given cost, complexity, latency, scalability and interpretability needs. Section 1.2 adds industry APIs like Document AI API, Vision API and Translate API as options when a prebuilt API fits.

Training hardware and distribution

The guide covers CPU, GPU and TPU options and distributed training with data and model parallelism. You also need to troubleshoot training failures, run hyperparameter tuning, and decide whether a foundation model from Model Garden should be fine-tuned at all.

Serving, scaling and rollout

Section 4 covers batch versus online inference on Agent Platform, Cloud Run and GKE, prebuilt versus custom containers for PyTorch and XGBoost, public versus private endpoints, and serving features from Feature Store. Model Registry versioning and A/B or canary rollouts are how you compare versions in production.

Securing and evaluating generative AI

Section 6 asks which tool stops data exfiltration, malicious prompting or sensitive data reaching an LLM: regex checks, safety filters or Model Armor. Section 2.3 adds evaluating genAI output with metrics and LLM-as-a-judge, and tracking lineage with Experiments and ML Metadata.

Why CertSharp

Built for Exam Day, Not Just Study Time

Study guides teach concepts. We prepare you for the actual PMLE exam.

1

Scenario questions cover all six exam-guide sections, from BigQuery ML and AutoML through Model Garden, Feature Store, pipelines, and Model Monitoring

2

Every question has a worked explanation that says why the correct option fits and why each distractor fails, so you learn the product tradeoffs the exam tests

3

Each question was checked against Google Cloud documentation, and many link to the official page for the service in question

4

Written for the current exam guide, including generative AI topics such as Model Garden, Gemini tuning, LLM evaluation, and Model Armor

Free Sample Questions

PMLE Practice Questions (Free Sample)

Real PMLE exam-style questions from our 500-question Google Professional Machine Learning Engineer bank, each with the correct answer and a detailed explanation.

Question 1 — PMLE

A team wants to release a new version of a recommendation model but is concerned about regressions. They want to expose it to a small share of live traffic and compare metrics before full rollout. What should they do?

  • A

    Overwrite the current model artifact in Cloud Storage so that the next endpoint restart automatically serves the new version to users

  • B

    Run batch inference with the new model on last month's data and promote it directly if the average score looks higher than before

  • C

    Create a second endpoint and ask client teams to hard-code a random number generator that picks the endpoint for each call made

  • D

    Deploy the new model to the existing Agent Platform endpoint with a small traffic split, then shift traffic as metrics hold

Correct answer: D — Explanation

An endpoint can host multiple deployed models with a configurable traffic split, enabling canary rollout and quick rollback. Overwriting the artifact removes the rollback path and tests nothing on live traffic. Offline batch scoring on history misses production behavior, and client-side routing logic is brittle and duplicates platform capability.

  1. Question 2Automating and orchestrating ML pipelinesEasy

    A team trains models in notebooks, hands the artifacts to an engineer who deploys them manually a few times a year, and has no pipeline or monitoring. At which MLOps maturity level is the team?

    • A

      Level 1, because they have a model in production, which is the defining feature of ML pipeline automation

    • B

      Level 2, because they release a few times a year, which counts as a continuous delivery cadence for the system

    • C

      Level 0, a manual process with notebook-driven training, infrequent releases, and little or no monitoring

    • D

      Level 3, because the engineer performs deployment steps carefully and the process is documented in team notes

    Show answer & explanation

    Correct answer: C — Level 0, a manual process with notebook-driven training, infrequent releases, and little or no monitoring

    Level 0 is characterized by manual, notebook-based work, a disconnect between data scientists and engineers, rare releases, and no active performance monitoring. Having a deployed model does not imply automation, a rare cadence is not CI/CD, and the widely used maturity model has no level 3.

  2. Question 3Scaling prototypes into ML modelsEasy

    A team trains a TensorFlow image classifier on millions of small JPEG files in Cloud Storage and the GPUs sit idle waiting on I/O. Which data format change most directly improves input throughput?

    • A

      Pack the images and labels into sharded TFRecord files and read them with tf.data in parallel

    • B

      Convert the labels to a CSV file and keep reading each JPEG individually from the bucket path

    • C

      Copy the JPEG files to the boot disk of the training VM at the start of each epoch

    • D

      Store the images as rows in a BigQuery table and query them one row at a time per step

    Show answer & explanation

    Correct answer: A — Pack the images and labels into sharded TFRecord files and read them with tf.data in parallel

    TFRecord shards give large sequential reads that tf.data can parallelize and prefetch, which removes per-object request overhead. A CSV of labels does not change the many-small-object access pattern. Recopying files each epoch adds startup cost, and row-at-a-time BigQuery queries are far too slow for image bytes.

  3. Question 4Collaborating within and across teams to manage data and modelsEasy

    A data science team's Agent Platform Workbench instances sit idle overnight and over weekends, driving up cost. What is the simplest managed way to reduce spend?

    • A

      Ask each scientist to remember to stop their own instance by posting a reminder in the team chat each evening

    • B

      Delete each instance at the end of the day and recreate it from scratch the next morning before work starts

    • C

      Enable idle shutdown so the instances stop automatically after a configured period of inactivity

    • D

      Move the notebooks to a larger machine type so jobs finish sooner and the instances sit idle less often

    Show answer & explanation

    Correct answer: C — Enable idle shutdown so the instances stop automatically after a configured period of inactivity

    Idle shutdown stops an instance after a period of inactivity, which removes compute charges without relying on people. Reminders are unreliable, daily deletion risks losing work and adds setup time, and a larger machine raises the hourly cost.

  4. Question 5Serving and scaling modelsEasy

    A hospital requires that online prediction traffic from its VPC to Agent Platform use internal IP addresses and not traverse the public internet. Which option fits best?

    • A

      A public Agent Platform endpoint combined with a firewall rule that blocks outbound traffic on port 80 from the client subnet

    • B

      A Cloud Storage bucket with uniform bucket-level access that receives the prediction requests as uploaded objects

    • C

      A Private Service Connect endpoint in the hospital VPC that forwards to a private Agent Platform endpoint

    • D

      A Pub/Sub topic with message encryption enabled that the clients publish to instead of calling the endpoint

    Show answer & explanation

    Correct answer: C — A Private Service Connect endpoint in the hospital VPC that forwards to a private Agent Platform endpoint

    Private Service Connect exposes the service through an internal IP in the consumer VPC, so traffic stays on Google's network. A public endpoint is reachable through the internet regardless of a port 80 rule. Buckets and topics are not prediction endpoints.

  5. Question 6Architecting low-code AI solutionsEasy

    A media team wants to produce short video clips from written storyboards using a managed Google model. Which option is correct?

    • A

      Imagen, which outputs a finished video file directly when the prompt includes a duration and a frame rate request

    • B

      Veo on Agent Platform, which generates video from text prompts and can be called from the Agent Platform APIs

    • C

      Speech-to-Text, which turns the storyboard narration into video frames while it transcribes the spoken description

    • D

      Video Intelligence API, which synthesizes new clips from a storyboard after label detection runs on the script text

    Show answer & explanation

    Correct answer: B — Veo on Agent Platform, which generates video from text prompts and can be called from the Agent Platform APIs

    Veo is the generative video model offered on Agent Platform. The Video Intelligence API annotates existing footage rather than creating it. Imagen is the image generation family. Speech-to-Text converts audio into text and does not produce visual content.

  6. Question 7Monitoring AI solutionsEasy

    A team enables feature attribution drift monitoring on an endpoint and the job configuration is rejected. What prerequisite is most likely missing?

    • A

      A Cloud Scheduler cron must exist in the project, because Model Monitoring jobs cannot run without one configured

    • B

      The model must be retrained in AutoML, because custom containers cannot produce feature attributions of any kind

    • C

      The endpoint must have request-response logging disabled, since attributions cannot be computed while payloads are saved

    • D

      The deployed model must have an explanation spec configured so that feature attributions can be computed for sampled requests

    Show answer & explanation

    Correct answer: D — The deployed model must have an explanation spec configured so that feature attributions can be computed for sampled requests

    Attribution drift monitoring needs Vertex Explainable AI to be configured on the model so attributions exist to compare. Request logging is independent of attributions. Custom models can have explanation specs when metadata is supplied. The monitoring service has its own schedule and does not require an external scheduler.

  7. Question 8Automating and orchestrating ML pipelinesMedium

    A data scientist ran the same pipeline with several hyperparameter and data settings and wants to compare the runs' parameters and metrics side by side in one place. Which service should the team use?

    • A

      Agent Platform Experiments, which tracks runs with their parameters and metrics for comparison

    • B

      Managed Service for Apache Airflow, which displays the task durations from each Airflow DAG run in a table

    • C

      Artifact Registry, which sorts container images by tag and shows their pushed dates

    • D

      Cloud Scheduler, which lists the last attempt time for each scheduled job it manages along with the target and cron setting

    Show answer & explanation

    Correct answer: A — Agent Platform Experiments, which tracks runs with their parameters and metrics for comparison

    Agent Platform Experiments (formerly Vertex AI Experiments) stores run parameters and metrics, including those from pipeline runs, and offers comparison views. Managed Service for Apache Airflow (formerly Cloud Composer) shows DAG task timing, not ML metrics for comparisons. Artifact Registry lists stored artifacts and has no metric tracking. Scheduler lists job attempts, not parameters or results.

  8. Question 9Scaling prototypes into ML modelsMedium

    A team wants a custom training script to save checkpoints and TensorBoard logs to locations Agent Platform manages for the job. How should the script find those locations?

    • A

      Hardcode a bucket path in the script and update it by hand before each job submission to the service

    • B

      Query the Agent Platform Model Registry for a checkpoint path at the start of the training job run

    • C

      Parse the container's stdout logs from a previous run to discover the output directory it used

    • D

      Read the AIP_CHECKPOINT_DIR and AIP_TENSORBOARD_LOG_DIR environment variables set for the job

    Show answer & explanation

    Correct answer: D — Read the AIP_CHECKPOINT_DIR and AIP_TENSORBOARD_LOG_DIR environment variables set for the job

    Agent Platform injects environment variables such as AIP_CHECKPOINT_DIR and AIP_TENSORBOARD_LOG_DIR, so code can stay portable across runs. Hardcoding is brittle and error-prone. The Model Registry stores model versions rather than live checkpoint locations, and scraping old logs is unreliable.

  9. Question 10Collaborating within and across teams to manage data and modelsMedium

    A model needs a GPU for a three-day training burst, but the Workbench instance is otherwise used for light exploration. What is the cost-conscious action?

    • A

      Stop the instance, change its machine type or accelerator for the burst, then scale back down after

    • B

      Run the light exploration work on the GPU instance as well so the accelerator is not left sitting idle

    • C

      Create a new project for each training burst and abandon it afterward to skip the cleanup work later

    • D

      Keep the GPU attached permanently so it is available whenever a teammate might need it for training

    Show answer & explanation

    Correct answer: A — Stop the instance, change its machine type or accelerator for the burst, then scale back down after

    Instances can be stopped and reconfigured, so accelerators are billed during the periods that need them. A permanent GPU or exploration on GPU hardware pays for unused capacity, and spawning projects adds governance overhead.

Want more PMLE practice questions?

Take the free PMLE practice test: 30 questions per session with instant explanations. Unlock all 500 questions plus timed exam mode for $11.99 one-time.

Start the Free 30-Question Practice Test

Pricing

Choose Your Plan

Start free, upgrade when you're ready

Free

Good for exploring the platform

$0
  • 30 practice questions
  • Practice mode only
  • Progress tracking
  • No exam simulation mode
Start Free
POPULAR

PMLE Access

Best for the final weeks before your PMLE exam, once you have used Agent Platform hands-on

$11.99one-time
  • 500 practice questions
  • Practice & Exam modes
  • Detailed explanations
  • Lifetime access
Get PMLE Access — $11.99

7-day money-back guarantee

BEST VALUE

Pro — All Certs

Planning multiple certifications

$9.99/month
  • ALL certifications included
  • Unlimited questions
  • New certs added free
  • Cancel anytime
View Pro Plans

Save 30% vs buying individually

Your PMLE Exam Won't Wait

Score 80% on two full CertSharp mock exams and you are calibrated to pass the live PMLE exam.

Don't risk $200 on the real exam without testing yourself first.

Start Free Practice Now

No credit card required · 30 free questions · Upgrade for $11.99 when ready

Official Certification Resources

Our practice questions align with official exam objectives. Study official materials first, then use CertSharp to test your readiness.

View Official PMLE Exam Details

FAQ

PMLE Frequently Asked Questions

Everything you need to know about Google Professional Machine Learning Engineer

What is the Google Professional Machine Learning Engineer certification?
The Professional Machine Learning Engineer (PMLE) validates that you can build, evaluate, productionize, and optimize AI solutions using Google Cloud, covering both conventional ML and generative AI. The exam fee is $200 (plus tax where applicable), with 50-60 multiple choice and multiple select questions over two hours.
How hard is the Google PMLE exam?
PMLE is a professional-level exam. It is scenario-driven and expects you to choose between services such as BigQuery ML, AutoML, custom training, and Model Garden based on cost, latency, and scale. Google recommends 3+ years of industry experience including 1 or more years designing and managing solutions on Google Cloud.
Is the Google PMLE certification worth it in 2026?
It is a good fit if your work involves training, deploying, or operating ML and generative AI systems on Google Cloud. It signals practical MLOps and platform knowledge rather than theory alone. Whether it moves your career depends on your role and employer, so check the job postings you are targeting.
What experience do I need for the PMLE exam?
Google lists no prerequisites. Its recommended experience is 3+ years of industry experience including 1 or more years designing and managing solutions using Google Cloud. The exam does not directly assess coding skill, but minimum proficiency in Python and SQL helps you interpret questions with code snippets.
What topics are on the Google PMLE exam?
Six sections: 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%), and monitoring AI solutions (~13%). Products include BigQuery ML, AutoML, Model Garden, Feature Store, Model Registry, Pipelines, and Model Monitoring.
How many questions are on the Google PMLE exam?
Google states 50-60 multiple choice and multiple select questions with a two-hour length. Google does not publish a passing score, so aim for solid, consistent results across all six sections rather than a target percentage.
How long should I prepare for the PMLE exam?
It depends on your background. Candidates who already run ML workloads on Google Cloud need less time than those new to the platform. A practical plan is to read the exam guide, practice hands-on in Google Cloud Skills Boost labs, and then work through scenario questions, spending extra time on your weakest sections.
How does PMLE compare with the Associate Cloud Engineer exam?
ACE is an associate-level exam on deploying and operating general Google Cloud infrastructure. PMLE is professional-level and specialized: it focuses on the ML lifecycle, including data preparation, training, serving, pipelines, and monitoring. You do not need ACE first, since Google lists no prerequisites for PMLE.
How much does the Google PMLE exam cost?
The registration fee is $200 USD plus tax where applicable. You can take the exam online-proctored from a remote location or onsite at a testing center. It is offered in English and Japanese.
Is there a free Google PMLE practice test?
Yes. CertSharp offers a free Google PMLE practice test: 30 free questions per session with detailed explanations, no credit card required. The full 500-question PMLE bank, including a timed exam mode of 60 questions, is $11.99 one-time with lifetime access.