Google Cloud Professional Machine Learning Engineer Practice Exam

What the Google Cloud Professional Machine Learning Engineer exam covers

• Architecting Low-Code AI Solutions104 questions
• Collaborating to Manage Data and Models122 questions
• Scaling Prototypes into ML Models153 questions
• Serving and Scaling Models158 questions
• Automating and Orchestrating ML Pipelines174 questions
• Monitoring AI Solutions105 questions

Free Google Cloud Professional Machine Learning Engineer practice test questions

A sample of 10 questions with answers and explanations. Sign up free to practice all 816.

1. Question 1Architecting Low-Code AI Solutions

   A team wants to see how recall and false positive rate trade off across a whole range of classification thresholds, rather than at just one fixed threshold, to help pick the best operating point. Which BigQuery ML function is built for this?

   • AML.EVALUATE, which reports one row of aggregate quality metrics for the whole model
   • BML.ROC_CURVE, which reports recall and false positive rate across many thresholdsCorrect
   • CML.CONFUSION_MATRIX, which reports prediction counts at a single fixed threshold
   • DML.GLOBAL_EXPLAIN, which reports global feature importance across the entire model
   ✓ Correct answer: B

   ML.ROC_CURVE returns recall, false positive rate, and the corresponding threshold across many threshold values, letting a team see the full tradeoff curve before picking an operating point. ML.CONFUSION_MATRIX only reports counts at one fixed threshold rather than a full sweep, ML.EVALUATE returns a single aggregate row rather than a per threshold breakdown, and ML.GLOBAL_EXPLAIN reports feature importance rather than threshold tradeoffs.

   Why the other options are wrong

   • AML.EVALUATE reports one row of aggregate quality metrics for the whole model, not values across a range of thresholds.
   • CML.CONFUSION_MATRIX reports counts at a single fixed threshold, not a full sweep across many threshold values.
   • DML.GLOBAL_EXPLAIN reports overall feature importance, it says nothing about the recall and false positive rate tradeoff.
2. Question 2Architecting Low-Code AI SolutionsSelect all that apply

   A pricing team at Talworth Goods trains an AutoML Tabular regression model to predict product demand and reviews the evaluation metrics. Which two metrics would appear in this regression evaluation? Choose two.

   • AMean absolute error measuring average prediction deviationCorrect
   • BRoot mean squared error weighting larger errors more heavilyCorrect
   • CA confusion matrix comparing predicted classes against actual outcomes
   • DArea under the precision recall curve for classification
   • ELog loss measuring the accuracy of predicted probabilities
   ✓ Correct answer: A, B

   Both metrics quantify how far predicted demand values fall from actual values, with root mean squared error weighting larger errors more heavily. Neither metric depends on class labels, which distinguishes them from classification specific outputs. Confusion matrices, precision recall curves, and log loss are all classification concepts and would not appear in a regression evaluation.

   Why the other options are wrong

   • CA confusion matrix compares predicted and actual classes, which does not apply to numeric regression output.
   • DPrecision recall curves are built from classification confidence scores, not continuous regression predictions.
   • ELog loss measures classification probability accuracy and has no equivalent role in regression evaluation.
3. Question 3Collaborating to Manage Data and Models

   A financial institution wants to prevent an insider with valid IAM permissions from copying training data out of Gemini Enterprise Agent Platform and BigQuery into a personal Cloud Storage project outside the organization's environment. Which control should they add on top of IAM?

   • AA Cloud KMS customer-managed key applied to the datasets
   • BAdditional predefined IAM roles restricting dataset viewer access
   • CA VPC Service Controls perimeter around the ML resourcesCorrect
   • DA DLP de-identification template applied before training
   ✓ Correct answer: C

   Because the risk here comes from a principal who already has valid IAM permissions, an additional IAM role would not close the gap, but a service perimeter adds a network-level boundary that blocks API calls moving data to services outside the perimeter regardless of IAM. A CMEK key protects data at rest and a DLP template transforms sensitive values, but neither restricts where the data can be copied to.

   Why the other options are wrong

   • AA CMEK key encrypts stored data with a customer-managed key but does not restrict where an authorized principal can copy that data.
   • BThe insider already has valid IAM permissions, so adding more IAM roles would not address a principal who can already access the data.
   • DA de-identification template transforms sensitive values within the data but does not prevent the transformed data from being copied elsewhere.
4. Question 4Scaling Prototypes into ML Models

   After tuning prefetch and num_parallel_calls, a team observes the accelerator utilization stays near 100 percent throughout training, and further parallelizing the input pipeline does not change step time. How should they classify this training job?

   • AThe job is compute bound by the model's own operationsCorrect
   • BThe job is limited by the shuffle buffer size
   • CThe job is input bound by Cloud Storage read throughput
   • DThe job is limited by insufficient sharding across files
   ✓ Correct answer: A

   If the data pipeline were the bottleneck, adding more parallel reads or preprocessing threads would typically reduce step time by feeding the accelerator faster. Since step time stays fixed despite these changes and the accelerator stays busy, the model's own forward and backward computation is what limits throughput.

   Why the other options are wrong

   • BShuffle buffer size affects randomization and memory use, not the accelerator utilization pattern described here.
   • CCloud Storage read throughput would matter if the accelerator were sitting idle waiting for data, which is not what is observed here.
   • DMore file sharding would help if input reads were the bottleneck, but here further input parallelism has no effect on step time.
5. Question 5Scaling Prototypes into ML Models

   A dataset was written to TFRecord shards sorted by class label. During training the engineer sets a shuffle buffer size of only 100 examples on a dataset containing millions of examples per class. What problem is most likely to occur?

   • AThe dataset will silently drop examples that fall outside the buffer window
   • BGPU utilization will drop because shuffling with a small buffer is CPU intensive
   • CBatches will mostly share one class since the buffer cannot mix distant labelsCorrect
   • DTraining will fail because tf.data enforces a minimum buffer equal to the batch size
   ✓ Correct answer: C

   Because the shuffle buffer only holds a small sliding window of the dataset, and the underlying files are sorted by label, most examples pulled into any given window still share the same class. This produces batches that are far less randomized than intended, which can bias gradient updates during training.

   Why the other options are wrong

   • Atf.data buffers examples in a sliding window rather than dropping data outside it, so this described failure mode does not occur.
   • BA small shuffle buffer actually reduces memory pressure and computation compared to a large one, so this is not the resulting problem.
   • DThere is no such enforced minimum relating shuffle buffer size to batch size in tf.data.
6. Question 6Serving and Scaling Models

   Pinecrest Analytics configured its endpoint's minimum replica count to zero to save cost overnight. What tradeoff should they expect when traffic resumes in the morning?

   • AThe endpoint will reject every request until it is manually redeployed by an operator
   • BThe model's prediction accuracy will start to degrade until it is retrained again
   • CThe first requests may see cold-start latency while a new replica spins upCorrect
   • DThe endpoint will permanently cap its maximum replica count at zero from then on
   ✓ Correct answer: C

   When idle, an endpoint with no minimum floor can scale down to zero running replicas to save cost. The tradeoff is that the first requests after that idle period have to wait for a new replica to provision and load the model before they can be served, adding latency compared to a warm replica.

   Why the other options are wrong

   • AA scale to zero endpoint still accepts requests, it just needs time to provision a replica rather than rejecting traffic outright.
   • BScaling behavior has no effect on the model's accuracy, only on latency while capacity ramps up.
   • DMinimum replica count does not affect the maximum, which is configured separately.
7. Question 7Serving and Scaling Models

   A lightweight model served through Cloud Run handles requests that each take only a few milliseconds of CPU time. The team notices each container instance is only handling one request at a time, leaving capacity unused. What should they adjust?

   • AIncrease the maximum number of Cloud Run instances allowed to scale out further
   • BIncrease the container concurrency settingCorrect
   • CIncrease the memory allocated to each Cloud Run instance
   • DIncrease the request timeout configured for the service
   ✓ Correct answer: B

   For a fast, lightweight model, one instance can safely handle many concurrent requests before it needs another instance to spin up. Concurrency is the setting that controls this directly, while raising the instance limit, memory, or timeout does not change how many requests one instance can serve at once.

   Why the other options are wrong

   • ARaising the instance limit allows more instances to start, but each one would still handle only one request at a time.
   • CMore memory does not change how many concurrent requests a single instance is configured to accept.
   • DRequest timeout controls how long a request may run, not how many requests one instance handles simultaneously.
8. Question 8Automating and Orchestrating ML PipelinesSelect all that apply

   Which two practices make an automated champion and challenger pipeline more reliable? (Choose two)

   • ASkipping the comparison window entirely for challengers trained using warm starting
   • BComparing performance using a statistically meaningful sample of live trafficCorrect
   • CRetiring the champion's monitoring pipeline once a challenger is deployed
   • DRolling back automatically if the challenger's key metrics regressCorrect
   • EPromoting any challenger the moment its first hour of traffic looks better
   ✓ Correct answer: B, D

   Waiting for a meaningful sample of live traffic before drawing conclusions avoids promoting on noise, and an automatic rollback protects production if the challenger's metrics regress during the comparison window. Promoting after only an hour of traffic, disabling the champion's monitoring, or skipping the comparison window for warm started challengers all remove safeguards rather than adding reliability.

   Why the other options are wrong

   • AHow a challenger was trained does not change the need to validate it against live traffic before promotion.
   • CRetiring the champion's monitoring removes visibility into its behavior exactly when a comparison against it is underway.
   • EPromoting based on just the first hour of traffic does not give a statistically meaningful sample and risks acting on noise.
9. Question 9Automating and Orchestrating ML Pipelines

   An engineer at Harborview Logistics has written several component functions and now needs to define the function that stitches them together into an ordered workflow with a name that will appear in the Agent Platform Pipelines console. Which decorator should wrap this orchestrating function?

   • Adsl.importer helper, registering an external artifact
   • B@pipeline decorator, defining the orchestrating workflow functionCorrect
   • C@container_component decorator, wrapping a single containerized step
   • D@component decorator, marking an individual step function
   ✓ Correct answer: B

   Wrapping the orchestrating function with the pipeline decorator lets the SDK recognize it as the top level workflow, including its name and description shown in the console, distinct from the individual component functions it calls. Component and container_component decorators mark individual steps rather than the overall workflow, and importer is used inside a pipeline to register external data, not to define the workflow itself.

   Why the other options are wrong

   • Adsl.importer brings an artifact into a pipeline, it is not used to define the workflow function.
   • CThe container_component decorator wraps a single containerized step, it does not define the overall workflow.
   • DThe component decorator marks an individual step function, not the orchestrating workflow.
10. Question 10Monitoring AI Solutions

    A platform team supports several product groups that all submit Gemini Enterprise Agent Platform training jobs and deploy endpoints in a shared project. Finance wants to attribute cost to each product group individually. What should the platform team apply to each job and endpoint to make this possible?

    • AA unique network tag per product group
    • BConsistent labels identifying the owning product groupCorrect
    • CA dedicated service account per product group
    • DA separate Cloud Monitoring workspace per product group
    ✓ Correct answer: B

    Labels flow through into billing export data, so finance can query or filter cost by the product group label without needing separate infrastructure per team. Service accounts, monitoring workspaces, and network tags may exist for other reasons but are not the mechanism billing reports use to break down cost by owner.

    Why the other options are wrong

    • ANetwork tags are used for firewall rule targeting and have no bearing on how cost is attributed in billing reports.
    • CService accounts control identity and permissions for a workload, but billing reports do not group cost by service account by default.
    • DMonitoring workspaces organize metrics and dashboards, not billing cost data.

Who this Google Cloud Professional Machine Learning Engineer practice exam is for

This practice set is for anyone preparing for the Google Cloud Professional Machine Learning Engineer exam at the advanced level - from first-time candidates building a foundation to experienced Google practitioners doing a final review before test day. If you learn best by working through realistic questions and reading why each answer is right or wrong, it is built for you.

How to use this Google Cloud Professional Machine Learning Engineer practice exam

1. Start with the free sample questions above to gauge your current baseline.
2. Read the full explanation on every question, including why each wrong option is wrong.
3. Track your weak domains and focus your study where you are losing the most marks.
4. Once you are scoring consistently well, take a timed, full-length mock exam.
5. Use your readiness score to decide when you are ready to book the real Google Cloud Professional Machine Learning Engineer exam.

Related Google resources

• Google Cloud Professional Machine Learning Engineer study guideKey concepts
• Google practice examsAll Google
• Certification pathWhere this fits
• Certification exam guides & tipsBlog
• Plans & pricingFree & paid
• How these questions are written and reviewedMethodology
• Report a problem with a questionCorrections
• Google Cloud Professional Security Operations Engineer practice examRelated
• Associate Google Workspace Administrator practice examRelated
• GCP Associate Cloud Engineer practice examRelated

Google Cloud Professional Machine Learning Engineer practice exam FAQ