Google ML Engineer Practice Test

The presence of complex, non-linear interactions between multiple dynamic variables that are difficult to model with manual static rules. is correct because ML is best suited for scenarios where complex, non-linear relationships exist among many dynamic features (like weather, real-time traffic, and historical delays) that static, rule-based systems cannot easily capture. The cited source, Professional ML Engineer Exam Guide, supports this answer for the Problem Framing and Use Case Selection scenario rather than the adjacent distractors.

Covariate shift occurs when the distribution of the input features $P(X)$ changes between training and production (e.g., temperatures spiking due to a heatwave), but the relationship between the features and the target label $P(Y|X)$ remains constant (the physics of rain generation didn't change). Concept drift occurs when the mapping itself changes (e.g., if a temperature spike that once indicated rain now indicates drought). Prior probability shift refers to a change in the frequency of labels $P(Y)$. Training-serving skew is an engineering discrepancy between pipelines, not an environmental event.

In regulated environments, models, prompts, and pipelines all need to be handled as governed production assets with reviewable change history. Ad hoc edits and undocumented approvals create avoidable operational and compliance risk.

Cloud Dataflow is the ideal serverless platform for highly parallelized, heavy batch-preprocessing workloads (such as cleaning and tokenizing millions of text files). Apache Beam supports reading raw text files from GCS, distributing the compute across worker nodes, and writing directly to the TFRecord format (the optimized binary storage format recommended for high-performance training with TensorFlow/TPUs). A single Workbench notebook cannot scale to 100 million files, BigQuery is optimized for structured tables rather than raw text file parsing, and Cloud SQL is a relational database not suited for large-scale unstructured NLP preprocessing.

Supervised fine-tuning is the intended choice when you have a well-defined task and labeled data that should shape the model's behavior. Merely renaming an endpoint or skipping evaluation would not adapt the model for the domain-specific task.

When a team has no specialized ML expertise, tight time constraints, and no proprietary training data for a standard task (like translation), a pre-built Google Cloud API (Cloud Translation API) is the ideal choice. It offers low-latency, state-of-the-art machine translation out-of-the-box, allowing integration within days. Custom model development on Vertex AI or BQML would require significant time, expertise, and labeled parallel text datasets.

In a shadow serving deployment, the new candidate model (Model v2) receives the exact same production traffic as the active model (Model v1) in real-time. However, Model v2's predictions are logged silently in the background and are not returned to the end users (who still receive Model v1's predictions). This allows testing Model v2's performance, latency, and system stability under real production load with zero user impact. Routing 10% of users to Model v2 where they actually see the results is a Canary deployment, not Shadow serving. Immediate deletion (blue-green without overlap) is highly risky, and running unmonitored at night is not a valid pattern.

Cloud Build is GCP's serverless CI/CD platform. It integrates natively with GitHub repositories. When code changes are pushed, Cloud Build triggers can execute defined steps: running Python unit tests, building custom container images via Docker, pushing them to Artifact Registry, and compiling the KFP pipeline code into JSON/YAML specifications. Using CE cron jobs, raw scheduler functions without build capabilities, or Workbench notebooks is complex, brittle, and violates standard DevOps practices.

Apply downsampling to the majority class (or SMOTE), evaluate using Precision-Recall AUC (PR-AUC), and impute missing income values using the median value (or a predictive imputer). is correct because For highly imbalanced data (0.1% fraud), class balancing (such as downsampling the majority class or synthetic oversampling like SMOTE) is required during training to prevent the model from ignoring the minority class. The cited source, Handling imbalanced datasets | Google Cloud, supports this answer for the Data Preparation and Feature Management scenario rather than the adjacent distractors.

When Vertex AI runs a custom training job, it automatically injects several environment variables into the container environment. `AIP_MODEL_DIR` contains the Cloud Storage path where the training script should save its final model artifacts. `AIP_TENSORBOARD_LOG_DIR` specifies where TensorBoard logs should be written. Hyperparameters should be passed as command-line arguments to the container entrypoint, which can be parsed using standard Python libraries like `argparse`. Hardcoding paths or values, using local `/tmp` without cloud sync, or making manual HTTP uploads violates standard containerized ML design on Vertex AI.