Google Cloud Professional Machine Learning Engineer Practice Exam

What the Google Cloud Professional Machine Learning Engineer exam covers

• Architecting Low-Code AI Solutions99 questions
• Collaborating to Manage Data and Models107 questions
• Scaling Prototypes into ML Models139 questions
• Serving and Scaling Models153 questions
• Automating and Orchestrating ML Pipelines169 questions
• Monitoring AI Solutions100 questions

Free Google Cloud Professional Machine Learning Engineer sample questions

A sample of 10 questions with answers and explanations. Sign up free to practice all 767.

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 Solutions

   A company wants to add multilingual voice prompts to its phone IVR system, choosing from many natural sounding voices with minimal engineering effort. Which service is appropriate?

   • ATranslation API for prompt text
   • BText-to-Speech API for promptsCorrect
   • CNatural Language API for prompt text
   • DSpeech-to-Text API for prompt audio
   ✓ Correct answer: B

   The team simply sends text and a chosen voice and receives audio ready to play. Speech-to-Text instead converts audio to text, the Translation API converts text between languages but produces no audio, and the Natural Language API only analyzes text without producing speech.

   Why the other options are wrong

   • AThe Translation API converts text between languages but does not generate audio output.
   • CThe Natural Language API analyzes text, it does not produce speech.
   • DSpeech-to-Text converts audio to text, not the synthesis needed for voice prompts.
3. Question 3Collaborating to Manage Data and Models

   A team wants to delete an old, unused version of a model from Vertex AI Model Registry, but the delete operation fails with an error. What is the most likely cause?

   • AModel Registry versions cannot be deleted once they are created
   • BThe version is still deployed and needs undeploying before deletionCorrect
   • CThe version must first be given the default alias before it can be deleted
   • DOnly the organization administrator has permission to delete versions
   ✓ Correct answer: B

   Vertex AI requires that a version be undeployed from any endpoints serving it before the version itself can be removed from the registry. Once undeployed, the delete operation can proceed normally.

   Why the other options are wrong

   • AModel versions can be deleted; there is no restriction that makes them permanent once created.
   • CHolding the default alias is unrelated to deletion; in fact a version actively in use as an alias target is a separate concern from endpoint deployment.
   • DDeletion permission is governed by IAM roles such as aiplatform.user or aiplatform.admin, not restricted solely to an organization administrator.
4. Question 4Scaling Prototypes into ML Models

   A team trains an XGBoost model in a custom job and plans to later deploy it to a Vertex AI endpoint using the pre-built XGBoost prediction container without writing a custom prediction routine. How should they name and save the artifact in AIP_MODEL_DIR?

   • ASave the booster as a pickle file named artifact.pkl inside AIP_MODEL_DIR
   • BSave the booster with any filename, since the prediction container inspects file contents to identify the model
   • CSave the booster under a subfolder named after the experiment run inside AIP_MODEL_DIR
   • DSave the booster as model.bst directly inside the AIP_MODEL_DIR pathCorrect
   ✓ Correct answer: D

   Vertex AI's pre-built prediction containers look for specific, well-known filenames such as model.bst for XGBoost boosters placed directly in the model artifact directory. Nesting it in a subfolder or using an unexpected filename or format prevents the container from automatically loading it without a custom prediction routine.

   Why the other options are wrong

   • AThe pre-built container looks for the specific filename model.bst, an arbitrarily named pickle file will not be recognized automatically.
   • BThe pre-built container relies on the expected filename and location rather than inspecting arbitrary file contents to identify the model.
   • CPlacing the file in a subfolder means the container will not find it at the expected top-level path, requiring a custom prediction routine instead.
5. Question 5Scaling Prototypes into ML Models

   A team training on Spot VMs is deciding how often to checkpoint. Checkpointing too frequently adds I/O overhead that slows training, while checkpointing too infrequently risks losing significant progress on preemption. What should guide their choice of checkpoint interval?

   • ACheckpoint after every single training step no matter the overhead
   • BCheckpoint only a single time at the very end of training
   • CSkip checkpointing entirely and simply restart training from scratch after every single preemption event
   • DBalance the interval against preemption frequency and rework costCorrect
   ✓ Correct answer: D

   Because both extremes carry cost, teams should choose a checkpoint interval informed by how often preemption is likely to occur and how much lost progress is acceptable, striking a practical balance. Checkpointing after every step adds overhead that can meaningfully slow training. Skipping checkpoints entirely means a preemption wipes out all progress, and checkpointing only at the end provides no protection against interruption during the run.

   Why the other options are wrong

   • ACheckpointing after every single step adds overhead that can noticeably slow down overall training throughput.
   • BA single checkpoint at the very end offers no protection against a preemption that occurs before training finishes.
   • CRestarting from scratch after each preemption wastes all prior progress and defeats the purpose of fault-tolerant training.
6. Question 6Serving and Scaling Models

   A startup has a small scikit-learn model wrapped in a FastAPI container that receives only a handful of requests per hour, with long stretches of no traffic at all. Which serving option minimizes cost while keeping operations simple?

   • ADeploy the container to Cloud Run, which scales to zero between requestsCorrect
   • BDeploy the container to a Compute Engine VM that runs continuously
   • CDeploy the model to a Vertex AI endpoint with two minimum replicas
   • DDeploy the container to GKE with a dedicated continuously running node pool
   ✓ Correct answer: A

   For sparse, unpredictable traffic, paying for continuously running infrastructure is wasteful. Cloud Run's scale-to-zero model fits a FastAPI container that only occasionally receives requests, while GKE node pools, Vertex AI minimum replicas, and a standing Compute Engine VM all incur cost even when no requests arrive.

   Why the other options are wrong

   • BA continuously running VM is billed around the clock regardless of how rarely it is used.
   • CMinimum replicas on a Vertex AI endpoint keep nodes warm and billed even with no incoming requests.
   • DA dedicated GKE node pool keeps compute running continuously, which is unnecessary for sparse traffic.
7. Question 7Serving and Scaling ModelsSelect all that apply

   Security at Glacier Bank wants to ensure a Vertex AI prediction endpoint is never reachable from the public internet. Choose two features that help achieve this.

   • APrivate Service Connect for private accessCorrect
   • BA private endpoint using VPC peeringCorrect
   • CA dedicated public endpoint configuration
   • DTraffic splitting across deployed models
   • EExplainable AI feature attribution reports
   ✓ Correct answer: A, B

   VPC Network Peering based private endpoints restrict access to a peered VPC network, while Private Service Connect extends private access to other projects or organizations without requiring peering. A dedicated public endpoint still serves traffic over the public internet, just with more isolation and a higher payload limit, and traffic splitting or explainability features do not affect network reachability at all.

   Why the other options are wrong

   • CA dedicated public endpoint still communicates over the public internet, it only adds isolation and a higher payload limit.
   • DTraffic splitting controls routing between deployed models and has no effect on network reachability.
   • EExplainable AI attribution reports describe prediction reasoning and have no effect on network exposure.
8. Question 8Automating and Orchestrating ML Pipelines

   A retailer's ML team needs to compute rolling 10 minute purchase counts per user from a live clickstream, applying custom session windowing logic in Python, to use as a real time feature. Which approach should they use to build this transformation?

   • AAn Apache Beam streaming pipeline on Dataflow with custom windowingCorrect
   • BA Cloud Composer DAG that executes a SQL aggregation every 10 minutes
   • CA BigQuery scheduled query that re runs its aggregation every 10 minutes
   • DA BigQuery materialized view that refreshes its aggregation continuously
   ✓ Correct answer: A

   Dataflow executes Apache Beam pipelines that natively support both batch and streaming data with custom windowing, triggers, and stateful processing written in general purpose code such as Python. BigQuery based approaches operate on discrete query executions or view refreshes and are not designed to maintain custom per user session state across a continuous stream. Cloud Composer is an orchestrator for scheduling tasks, not a stream processing engine itself.

   Why the other options are wrong

   • BCloud Composer orchestrates when tasks run but delegates the actual data processing elsewhere, so it does not itself provide stream processing capability.
   • CA scheduled query re run every 10 minutes only approximates streaming and cannot maintain custom session state between runs.
   • DA materialized view keeps a SQL aggregation fresh but cannot express custom Python windowing logic.
9. Question 9Automating and Orchestrating ML Pipelines

   A team needs to call three Cloud Functions in sequence, with a conditional branch that skips the third function when the second returns an empty result, and they want a serverless option with no persistent infrastructure and no need for artifact lineage tracking. Which tool fits best?

   • ACloud Composer, heavier than needed for three simple calls
   • BVertex AI Pipelines, focused on ML specific artifact lineage
   • CCloud Workflows, suited to short conditional call chainsCorrect
   • DCloud Scheduler, limited to firing a single scheduled target
   ✓ Correct answer: C

   Cloud Workflows is designed exactly for this kind of short sequence of service calls with branching, executing steps defined in YAML or JSON without provisioning any servers. Cloud Composer and Vertex AI Pipelines both require more setup and are aimed at larger DAGs or ML specific artifact tracking, which is unnecessary overhead here.

   Why the other options are wrong

   • ACloud Composer could run this as a DAG, but its persistent Airflow environment is unnecessary overhead for three simple sequential calls.
   • BVertex AI Pipelines is built for ML specific steps with artifact lineage, which this simple function chaining does not need.
   • DCloud Scheduler only fires a single target on a schedule and has no way to express conditional branching between calls.
10. Question 10Monitoring AI Solutions

    A logistics analytics team is getting frequent alerts on a feature that fluctuates naturally with weekly seasonality, using the monitoring job's default alerting threshold. What adjustment would most directly reduce these false alarms for that one feature?

    • ASwitch the entire monitoring job from prediction drift detection to training-serving skew detection instead
    • BRaise the alert threshold for that specific feature so smaller fluctuations stop triggering alertsCorrect
    • CLower the alert threshold configured specifically for that feature so the job becomes more sensitive overall
    • DIncrease the sampling rate for the entire monitoring job so more requests are captured for that feature
    ✓ Correct answer: B

    Each feature can have its own threshold value representing the distance score that must be crossed before an alert fires, and the default value, commonly around 0.3, is only a starting point. Raising that threshold for the seasonal feature specifically means normal fluctuation no longer triggers a notification, while other features keep their own settings. Changing the sampling rate or the monitoring mode would not target the specific noisy feature the way a per-feature threshold change does.

    Why the other options are wrong

    • ASwitching between skew and drift changes the comparison baseline, not the sensitivity threshold that governs false alarms.
    • CLowering the threshold would make the job more sensitive and produce more alerts, the opposite of what is needed here.
    • DA higher sampling rate improves statistical confidence but does not change the sensitivity setting that determines when an alert fires.

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
• Google Cloud Professional Cloud Security Engineer practice examRelated
• Google Cloud Professional Cloud Network Engineer practice examRelated
• Associate Google Workspace Administrator practice examRelated

Google Cloud Professional Machine Learning Engineer practice exam FAQ