Google Cloud Professional Machine Learning Engineer practice questions

Question 1

A developer has decided to consult an MLOps engineer when to reuse Google Cloud Pipeline Components rather than write custom Kubeflow Pipelines components from scratch for Vertex AI Pipelines. Which option matches the documented Google Cloud Pipeline Components SDK?

Accepted Answer

The Google Cloud Pipeline Components SDK provides a set of prebuilt Kubeflow Pipelines components that are production quality, performant, and easy to use. You can use Google Cloud Pipeline Components to define and run ML pipelines in Vertex AI Pipelines and other ML pipeline execution backends conformant with Kubeflow Pipelines.

Answer

Google Cloud Pipeline Components are research-only and not suitable for production pipelines. The documentation describes them as production quality, performant, and easy to use.

Answer

Google Cloud Pipeline Components cannot create datasets or run AutoML training and only do logging. The documented components include creating datasets, training AutoML models, running custom training jobs, batch prediction, and creating endpoints.

Answer

Vertex AI Pipelines rejects Google Cloud Pipeline Components and requires every component to be hand-built. Google Cloud Pipeline Components are explicitly built for Vertex AI Pipelines.

Question 2

An MLOps engineer is migrating a senior ML lead which Vertex AI capability lets many small, sparsely-trafficked models share a single VM to reduce serving cost, and what the trade-off is regarding resource isolation. Which option matches the documented Vertex AI model cohosting design?

Accepted Answer

Model cohosting support introduces the concept of a DeploymentResourcePool, which groups model deployments that share resources within a single VM. Cohosting works well when you have many deployed models with sparse traffic. Models in a single deployment resource pool aren't isolated and can be in competition for CPU and memory.

Answer

Vertex AI cohosting fully isolates each model from every other on the shared VM so there is no risk of competition. Cohosted models are not isolated and can be in competition for CPU and memory. The documented behavior in the published Google Cloud reference is the option that matches the verbatim reference snippet attached to this item, so the other choices are eliminated by direct contradiction with that snippet.

Answer

Vertex AI cohosting is forbidden for any model and every deployed model must own its VM. Cohosting via DeploymentResourcePool is the documented mechanism for sharing a VM.

Answer

Cohosting on Vertex AI is best when each model receives sustained heavy traffic. The documentation states cohosting works well when you have many deployed models with sparse traffic.

Question 3

A developer is migrating a hyperparameter search onto a Vertex AI hyperparameter tuning job and must understand how the platform delivers candidate hyperparameter values to the training application and how each trial's effectiveness is evaluated. Which statement is correct?

Accepted Answer

Vertex AI passes hyperparameter values as command-line arguments, and the application must report its metrics for each trial.

Answer

Vertex AI writes hyperparameter values into a mounted config file and reads the accuracy from the saved model checkpoint automatically.

Answer

Vertex AI injects hyperparameter values as environment variables, and Cloud Logging parses trial effectiveness from stdout.

Answer

Vertex AI sets hyperparameter values through the WorkerPoolSpec, and the parameter server reports trial metrics for the app.

Question 4

A team needs to guarantee that the same Dataform workflow definition always compiles to identical SQL, and the data engineering group is debating whether their tool may call an external API during compilation to enrich its config. Which statement matches Dataform's documented hermetic compilation model?

Accepted Answer

Compilation is hermetic and runs in a sandbox with no internet access, so identical code always yields the same SQL and external API calls are unavailable during it.

Answer

Compilation runs against the live BigQuery endpoint, so external API calls are permitted and the emitted SQL can vary between runs as the upstream source data shifts over time.

Answer

Compilation needs internet access to pull the newest versions of project dependencies, which is why two runs of identical code can produce different SQL output.

Answer

Compilation executes the workflow directly in production, and the resulting plan cannot be inspected as an interactive graph before it actually runs against data.

Question 5

A company is migrating a call-center pipeline to Cloud Speech-to-Text and must transcribe both sub-minute voice prompts and full hour-long recorded calls. Select the two statements that match the documented synchronous and asynchronous recognition limits.

Accepted Answer

Synchronous recognition returns results only after all audio is processed and handles requests of one minute or less.

Accepted Answer

Asynchronous recognition runs as a long-running operation that you poll and supports audio of any duration up to 480 minutes.

Answer

Synchronous recognition streams partial results live and accepts audio of any duration up to the 480-minute ceiling.

Answer

Asynchronous recognition returns the entire transcript in a single immediate response, so no result polling is ever required.

Question 6

An engineering director is tasked with explaining an ML lead to summarize the scale of the Vertex AI Model Garden catalog and what underlying accelerator infrastructure backs it. Which option matches the documented Vertex AI overview?

Accepted Answer

Vertex AI provides access to the Model Garden, which features a curated catalog of over 200 models — including Google's foundation models such as Gemini and a comprehensive selection of partner and open models — along with the underlying TPU/GPU infrastructure.

Answer

Vertex AI's Model Garden catalog ships with under 10 models and no underlying accelerator infrastructure. The documentation describes over 200 models and underlying TPU/GPU infrastructure.

Answer

Vertex AI Model Garden has no Google-built foundation models. The documentation explicitly includes Google's foundation models such as Gemini.

Answer

Vertex AI does not offer underlying TPU or GPU infrastructure for Model Garden. The documentation explicitly mentions the underlying TPU/GPU infrastructure as part of Vertex AI.

Question 7

A senior ML engineer is evaluating a developer to define what continuous training means in the MLOps maturity framework, and what Level 1 introduces. Which option matches the documented Level 1 description?

Accepted Answer

Level 1 introduces ML pipeline automation enabling continuous training by automating the ML pipeline through orchestrated experiment steps, automated data and model validation, pipeline triggers, and metadata management.

Answer

Continuous training in the MLOps framework means manually retraining models on a quarterly cadence. The documentation defines continuous training in Level 1 as automated through orchestrated experiment steps and pipeline triggers.

Answer

Continuous training requires teams to skip data validation entirely. Level 1 includes automated data and model validation as part of ML pipeline automation.

Answer

Level 1 of the MLOps framework eliminates metadata management. Level 1 explicitly includes metadata management as part of the pipeline automation.

Question 8

An ML engineer is evaluating an MLOps lead when a Vertex AI custom inference container is the right serving path instead of a prebuilt container, for example to pre- or post-process inference requests or to use an unsupported framework. Which option matches the documented Vertex AI custom inference container reasons?

Accepted Answer

You might want to use a custom container to serve inferences from an ML model trained using a framework that isn't available as a prebuilt container, to preprocess inference requests or postprocess the inferences generated by your model, to run an inference server written in a programming language of your choice, or to install dependencies that you want to use to customize inferences.

Answer

Vertex AI does not support custom inference containers; teams must always use prebuilt containers. Custom inference containers are an explicit, documented serving option.

Answer

Custom inference containers on Vertex AI must always be re-implemented in Go regardless of the team's preferred language. The documentation lets teams choose the inference server's programming language.

Answer

Vertex AI custom inference containers can only host the same set of frameworks as prebuilt containers with no additional capabilities. The documentation explicitly cites frameworks that aren't available as prebuilt containers as a use case. The documented behavior in the published Google Cloud reference is the option that matches the verbatim reference snippet attached to this item, so the other choices are eliminated by direct contradiction with that snippet.

Question 9

A senior ML engineer is tasked with explaining a hardware partner how Cloud TPUs scale memory and compute for large models, including how multiple TPUs are grouped to handle bigger workloads. Which option matches the Cloud TPU introduction?

Accepted Answer

TPUs have on-chip high-bandwidth memory (HBM) letting you use larger models and batch sizes. TPUs can be connected in groups called slices that scale up your workloads.

Answer

Cloud TPUs share host RAM with the controlling VM and cannot hold large models on-chip. The documentation states TPUs have on-chip high-bandwidth memory (HBM) for larger models and batch sizes.

Answer

Cloud TPUs always operate as a single chip and cannot be grouped to scale up workloads. The documentation states TPUs can be connected in groups called slices.

Answer

TPUs share PCIe DRAM directly with GPUs and therefore have no separate on-chip memory. The documentation describes on-chip high-bandwidth memory (HBM).

Question 10

A data science team that mostly writes R is being optimized by an ML engineering lead which Google Cloud surface combines R notebooks with BigQuery for exploratory data analysis at scale on Google Cloud. Which option matches the documented R-on-Google-Cloud architecture?

Accepted Answer

This document focuses on performing exploratory data analysis using Vertex AI Workbench instances and BigQuery.

Answer

Cloud Run is the documented surface for running R notebooks against BigQuery for exploratory data analysis. The documented architecture pairs Workbench instances with BigQuery, not Cloud Run.

Answer

The team must abandon R entirely and switch to Python; the architecture does not support R. The documented architecture is specifically for data science with R on Google Cloud.

Answer

Dataproc clusters are documented as the only surface that can run R for exploratory data analysis. The documented architecture uses Workbench instances and BigQuery.

Question 11

A developer is being asked to choose between Cloud Vision API features for two distinct workloads: extracting short text from street-sign photos, and extracting dense text from scanned multi-page documents with paragraph and word structure. Which Vision API mapping matches the documented OCR feature pair?

Accepted Answer

Cloud Vision API provides two annotation features that support optical character recognition: TEXT_DETECTION detects and extracts text from any image, and DOCUMENT_TEXT_DETECTION also extracts text from an image, but the response is optimized for dense text and documents, with the JSON including page, block, paragraph, word, and break information.

Answer

Use TEXT_DETECTION for both workloads because DOCUMENT_TEXT_DETECTION is deprecated and routes the same way internally. Both features remain documented OCR options, and DOCUMENT_TEXT_DETECTION targets dense-text documents.

Answer

Use Document AI Form Parser for the street-sign photos and Vision API DOCUMENT_TEXT_DETECTION only for handwriting. Vision API itself documents TEXT_DETECTION for arbitrary images including street signs.

Answer

Vision API has a single LANDMARK_DETECTION feature that returns OCR text as a side effect, so no separate OCR feature is needed. The Vision API documents TEXT_DETECTION and DOCUMENT_TEXT_DETECTION as the OCR features, not LANDMARK_DETECTION.

Question 12

An organization needs to harden the audit posture of its Vertex AI deployment and is reviewing which Agent Platform audit log types are always recorded and which require explicit enablement. Select the two options that match the documented Agent Platform audit logging behavior.

Accepted Answer

Admin Activity audit logs record admin write operations on metadata or configuration and cannot be disabled.

Accepted Answer

Data Access audit logs include data read and data write of user data and must be explicitly enabled to receive.

Answer

System Event audit logs record the admin read operations that users perform against metadata or configuration.

Answer

Data Access audit logs are recorded by default and capture the automated Google Cloud configuration changes.

Question 13

An MLOps engineer is rolling out a developer how a Vertex AI pipeline written with the Kubeflow Pipelines SDK gets from local Python code to a runnable artifact on Vertex AI Pipelines. Which option matches the documented Vertex AI pipelines build flow?

Accepted Answer

After defining the pipeline workflow by connecting component outputs as inputs to downstream components, you compile it into YAML format using the Kubeflow compiler, then submit the compiled pipeline to Vertex AI Pipelines using the Python client library, specifying the project ID, region, and pipeline root path for artifact storage.

Answer

Vertex AI Pipelines reads Python source files directly without any compilation step. The documented flow compiles the pipeline into YAML before submission.

Answer

The Vertex AI Pipelines service requires a hand-written JSON pipeline definition with no SDK-driven compilation. The Kubeflow Pipelines SDK compiles pipelines into YAML using the Kubeflow compiler.

Answer

Vertex AI Pipelines requires the engineer to upload Python source files directly to Cloud Functions before any pipeline can run. The documented submission path is the Python client library against the compiled YAML. The documented behavior in the published Google Cloud reference is the option that matches the verbatim reference snippet attached to this item, so the other choices are eliminated by direct contradiction with that snippet.

Question 14

A production team is migrating a high-throughput model to a public Vertex AI endpoint and requires larger request payloads, longer request timeouts, and isolation from other customers' traffic. Which public endpoint type does the documentation recommend they choose as a best practice?

Accepted Answer

A dedicated public endpoint, which provides production isolation, larger payload sizes, and longer request timeouts.

Answer

A shared public endpoint, because shared endpoints reserve isolated per-tenant network capacity and larger payloads.

Answer

A private endpoint over VPC peering, because public endpoints can never isolate one tenant's traffic from another's.

Answer

Either public endpoint type, since shared and dedicated public endpoints expose identical timeouts and payload limits.

Question 15

A developer is tasked with inspecting the file system of a running custom training container on Vertex AI to analyze GPU usage and confirm the Google Cloud permissions available to the container. Which Vertex AI capability supports this?

Accepted Answer

Use an interactive shell to browse the container file system and run debugging utilities for GPU and permission checks.

Answer

Use Cloud Profiler to open a remote interactive terminal into the running container and edit the training process in place.

Answer

Use Cloud Logging to stream stdout, which lets you browse the live container file system and inspect installed GPU drivers.

Answer

Use the WorkerPoolSpec containerSpec.args field to attach a shell and inspect container permissions after the job ends.

Google Cloud Professional Machine Learning Engineer

Sample questions

Sources

Similar exams