Google Professional Machine Learning Engineer 1-10

For a requirement that demands full control over the underlying infrastructure while minimizing inference latency, manually deploying to a GKE cluster with a custom YAML manifest is the best approach.
GKE lets you fine-tune node types, accelerator placement, and scheduling in detail, so you can keep inference latency to a minimum.
Vertex AI’s one-click deploy (Option A) and deployment via Model Garden (Option D) are convenient but offer far less freedom in controlling the infrastructure.
Vertex AI’s custom inference container (Option C) allows some customization, but you cannot command the entire underlying platform as fully as with GKE.
Option B is correct because it achieves both maximum control and low latency.
Google Kubernetes Engine (GKE) Documentation

Traffic splitting is a feature that lets you distribute prediction requests across multiple models or versions within the same Vertex AI endpoint.
Because you can deploy the new model to the existing endpoint and route only 5% of traffic to it, you can perform a production canary test without creating a new endpoint or a separate service.
This minimizes complexity while allowing a gradual rollout as you watch metrics such as listening time.
Option A requires building a new endpoint and a routing service, which is complex.
Option B stays at offline evaluation and does not amount to production validation.
Option D is a wholesale switch that is high-risk and is not a gradual test.
Vertex AI: Model Deployment and Traffic Splitting

Vertex AI’s Gen AI Evaluation Service is a feature designed to automatically evaluate and compare multiple foundation models or LLMs in parallel using your own datasets.
By using the evaluation service API integrated with Model Garden, you eliminate the need to write your own evaluation scripts or manage complex experiments, allowing you to compare the models most efficiently.
Because the question emphasizes using the “curated internal benchmark datasets,” Options B and C, which substitute open-source data, do not meet the requirement.
Option D requires downloading weights and writing scripts, which is labor-intensive and runs counter to the efficiency requirement.
Vertex AI: Gen AI Evaluation Service API

Batch prediction is a process that runs predictions on a large volume of data all at once, and it is well suited to cases that do not require real-time responsiveness and can be processed on a schedule.
For a use case that digitizes a batch of forms at the end of each business day, Option A—submitting a batch prediction job against the input data in Cloud Storage—is optimal.
The results are also written to Cloud Storage, minimizing manual work and scaling automatically.
Options B, C, and D are geared toward real-time/online use and are either overkill or unsuitable for daily batch processing.
Note that AI Platform is now integrated into Vertex AI, but the concept of batch prediction and the correct answer remain valid today.
Vertex AI: Get Batch Predictions

TensorFlow Data Validation (TFDV) is a library for understanding, validating, and monitoring machine learning data.
Because it can automatically detect and report schema anomalies such as missing features, new features, and data-type mismatches, it is ideal for hardening a pipeline against format changes.
It can also generate descriptive statistics and visualizations, and embedding it in the training pipeline lets you continuously ensure data quality.
Option B forcibly coerces values and thereby hides the problem.
Options C and D require custom implementations, lack comprehensiveness, and are inferior to TFDV for schema validation.
TensorFlow Data Validation (TFDV) Guide

To keep code version control while minimizing compute cost and manual work, a configuration that integrates a source repository with CI and automatically launches retraining on a code push is appropriate.
Option C—integrating Cloud Build with the repository and running retraining triggered by a push—is the optimal solution that satisfies both version control and automation.
Option A relies on coarse Storage monitoring for change detection, and Option B requires a manual run every time.
Option D’s daily schedule runs regardless of code changes, creating waste.
Note that Cloud Source Repositories stopped accepting new customers as of June 2024 (existing users can continue to use it), but the concept of triggering Cloud Build on a code push remains valid today.
Cloud Build: Creating and Managing Build Triggers

Model retraining is the process of updating a model with new data to maintain its performance and accuracy.
When accuracy gradually declines even though the model has not changed, the cause is model drift—the distribution of data such as the market shifting over time—and the remedy is periodic retraining.
Option A (data quality), Option C (too few layers), and Option D (incorrect split ratios) are factors that would produce a certain amount of error from the moment of deployment, and they do not explain the “continuous degradation” over time.
Because retraining is essential to keep up with environmental change, Option B is correct.
Overview of Vertex AI Model Monitoring

To deploy a multi-model workflow while balancing version control and low cost, it is efficient to expose each model with Vertex AI Endpoints and implement the integration layer with a lightweight serverless option.
Exposing each model individually with Vertex AI Endpoints makes version control and reuse easy, and integrating the workflow with Cloud Run, which can scale to zero, minimizes resource usage when idle.
Option A requires building the integration layer yourself in a custom container, which carries high operational overhead, and Options B and D load model files directly, undermining per-model version control and reusability.
Overview of Cloud Run

The Occupancy Analytics model in Vertex AI Vision is a pretrained model designed to count people and vehicles within video frames.
It provides standard features such as counting within an active zone, line-crossing counts, and dwell detection, so it can meet the need to detect the number of waiting customers in near real time with minimal effort.
Options C and D require your own annotation and training, which is labor-intensive, and Option B is a separate model specialized in person/vehicle detection and is not as directly suited as occupancy analytics.
Note that Vertex AI Vision has now been renamed Agent Platform Vision, but this model and the correct answer remain valid today.
Agent Platform Vision: Occupancy Analytics Tutorial

Reduction Server is a fast all-reduce mechanism for GPUs that aggregates gradients from workers using a dedicated group of reducers.
Because the reducers are meant to run on lightweight CPU VMs that are cheaper than GPUs, the correct configuration is to give the third worker pool no accelerators and choose a machine type that emphasizes bandwidth.
Reduction Server does not support TPUs, so configurations that use TPUs, as in Options C and D, are inappropriate.
Option A is wrong because it assigns unnecessary GPUs to the reducer pool, increasing cost.
Option B is correct: GPUs for the two worker pools, and CPU plus high bandwidth for the reducer pool.
Vertex AI: Distributed Training (Reduction Server)