Professional Machine Learning EngineerSample questions
Independent: not affiliated with Google Cloud. Free to try; the study bundle adds a study guide and two practice exams.
5 original questions, none of them in the practice exams.
Pick an answer. Show answer explains every option and links its Google Cloud source.
5 questions / Explanations link Google Cloud sources / Last updated Oct 26, 2026
Question 1
Multiple choice
Section 1 / Task 1.1 Develop ML models with BigQuery ML or AutoML
A payments analyst trains a BigQuery ML logistic regression model from a query that standardizes the transaction amount with ML.STANDARD_SCALER and buckets the hour of day with ML.BUCKETIZE before the CREATE MODEL statement. When the fraud team runs ML.PREDICT on the raw transactions table, the predictions are poor because the model receives unscaled, unbucketed values. The team wants ML.PREDICT callers to pass raw rows with no preprocessing of their own. What should the analyst change?
Show answerHide answer
Answer: A
Why
The TRANSFORM clause of CREATE MODEL holds all preprocessing at model creation and applies it automatically during prediction and evaluation, including the statistics of ML analytic functions such as the scaler's mean and standard deviation. Callers pass raw columns and BigQuery ML applies the training transformations.
Why not the others
BContradicts the requirement, and every caller's query becomes a place where the preprocessing can diverge from training.
CML.EVALUATE reports metrics for a model; it does not change the model or apply preprocessing to inputs.
DChanges only the training table; rows passed to ML.PREDICT are still raw, so the model still receives untransformed values.
Section 2 / Task 2.2 Prototype models in notebooks
A platform team supports 40 data scientists who use Colab Enterprise. Each scientist configures a runtime by hand, and the results differ: some runtimes have no GPU, some have public internet access that the security policy forbids, and some have idle shutdown turned off, so they keep running for days after the last cell executes. The team wants every scientist to start from the same approved VM configuration, stay in Colab Enterprise, and not depend on the team creating compute for each person. What should the platform team do?
Show answerHide answer
Answer: B
Why
A Colab Enterprise runtime template is a VM configuration: machine type, accelerators, network, public internet access and idle shutdown (on by default at 180 minutes). A template can be shared with other users, and each creates their own runtime from it, so all 40 run the same approved configuration.
Why not the others
ALeaves the configuration to each scientist, which is what produced the inconsistent runtimes.
CRuntimes are exclusive to a user: the page states that you can't share your runtimes with others and can't use their runtimes.
DAbandons Colab Enterprise and has the team create 40 separate Workbench instances, one per person, which the team wants to avoid.
A serverless training job (CustomJob) on Agent Platform runs a PyTorch script in a custom container on four GPUs. After about two hours each run slows to a crawl without failing, and the logs show nothing unusual. The ML engineer wants to watch GPU utilization and sample the Python call stack of the live process during the slow phase, without changing the training code. What should the engineer do?
Show answerHide answer
Answer: C
Why
The interactive shell is enabled by setting enableWebAccess to true (enable_web_access in the Python SDK) when creating the CustomJob. It runs debugging utilities in the live container: nvidia-smi for GPU utilization, py-spy to profile a running Python program without modifying it, and perf.
Why not the others
AObserves a different environment rather than the managed job that misbehaves, and leaves the engineer managing a GPU VM.
BChanges the training code, which the requirement rules out, and requires guessing in advance where to instrument.
DThe shell is available only while the job is running; access is lost as soon as the job ends, and files created in the container do not persist.
A retailer serves 12 TensorFlow models for online inference on Agent Platform, one endpoint per model, each on its own n1-standard-4 node. Every model receives a few requests per hour, so most node time is idle and the monthly bill is dominated by the 12 nodes. All 12 models use the same prebuilt TensorFlow inference container image, and all must stay available for online requests. What should the ML engineer do to reduce the cost?
Show answerHide answer
Answer: D
Why
A deployment resource pool is a group of model deployments that share the resources of a single VM; Google lists cost-effective serving as a benefit and says cohosting suits many deployed models with sparse traffic. The pool takes models that use the same container image, and TensorFlow support is generally available.
Why not the others
ASharing an endpoint does not share a VM: each deployed model still gets its own VM by default, with its own minimum node count billed even without traffic, so 12 deployed models still need at least 12 nodes.
BRemoves online inference, which the models must keep providing.
CThe endpoint type changes networking and request limits, not the billing: each deployed model still runs on its own nodes, billed per node hour even when there is no traffic.
Section 5 / Task 5.1 Develop end-to-end ML pipelines
A nightly Agent Platform Pipelines run retrains a pricing model. Twice this month a step failed at 3 a.m. and nobody noticed until the stale model was spotted the next afternoon. The team wants an email to three on-call engineers whenever a run finishes, whether it succeeded, failed or was cancelled, without building a separate alerting system. Which two changes to the pipeline definition meet this requirement? Choose two.
Show answerHide answer
Answers: B and C
Why
Google's email notification setup imports VertexNotificationEmailOp from google_cloud_pipeline_components.v1.vertex_notification_email and places the pipeline tasks inside a dsl.ExitHandler whose exit task is that component. When the pipeline exits, Google Cloud emails the addresses with the status (SUCCEEDED, FAILED or CANCELLED), the job ID and a console link.
Why not the others
AThe component accepts a comma-separated list of up to three email addresses.
DA task at the end of the pipeline does not run when an earlier task fails, so the failures that matter send nothing.
ECaching controls whether a task's earlier output is reused; it has no connection to notifications.
The study bundle has two practice exams with 120 original questions, an answer key and an explanation for every answer, plus the 59-page study guide. Practice exam explanations are shorter than the ones on this page: a short paragraph per question, with no source links.
Study bundle
$19
120 practice questions / Study guide / Answer keys
Every bundle includes 6 months of access: download the PDFs, and any new editions, until 6 months after your purchase date. Your receipt has a license key. Enter it at overqual.com/download for the three PDFs, stamped with your email and order number. They are yours to keep.