Model Registry<\/span><\/p>\nAnswer: 3) Online store<\/b><\/p>\n
Explanation:<\/b><\/p>\n
The SageMaker Feature Store online store is designed for low-latency retrieval of features during real-time inference. When an application needs current feature values quickly, the online store can provide access without requiring a large analytical query. The offline store serves different purposes, such as historical data analysis and model training. Using the appropriate store depends on the workload requirements. For real-time applications, latency is particularly important because delays in retrieving features can directly affect the overall inference response time.<\/span><\/p>\nQuestion 62. A company wants to prevent unauthorized users from changing a production model’s deployment configuration. Which AWS security mechanism should be used?<\/b><\/h3>\n
1)<\/b> IAM policies and permissions<\/span>
\n<\/span>2)<\/b> CloudWatch metrics<\/span>
\n<\/span>3)<\/b> S3 lifecycle rules<\/span>
\n<\/span>4)<\/b> Model Monitor constraints<\/span><\/p>\nAnswer: 1) IAM policies and permissions<\/b><\/p>\n
Explanation:<\/b><\/p>\n
AWS Identity and Access Management provides permissions that determine which users, roles, and services can perform actions on AWS resources. By applying least-privilege IAM policies, an organization can restrict who is allowed to modify SageMaker endpoints or deployment configurations. This helps protect production machine learning resources from unauthorized changes. CloudWatch is primarily used for monitoring, while Model Monitor focuses on production data and model monitoring. Proper IAM design should separate development and production permissions where appropriate and should avoid granting broad administrative access when narrower permissions are sufficient.<\/span><\/p>\nQuestion 63. Which metric is especially useful when the cost of false positives and false negatives must both be considered for a binary classification model?<\/b><\/h3>\n
1)<\/b> Mean absolute error<\/span>
\n<\/span>2)<\/b> F1 score<\/span>
\n<\/span>3)<\/b> Mean squared error<\/span>
\n<\/span>4)<\/b> Root mean squared error<\/span><\/p>\nAnswer: 2) F1 score<\/b><\/p>\n
Explanation:<\/b><\/p>\n
The F1 score combines precision and recall into a single classification metric. Precision reflects the proportion of predicted positive cases that are actually positive, while recall measures how many actual positive cases the model successfully identifies. The F1 score is useful when both types of classification errors matter and a balance between precision and recall is required. It can be especially informative when class distributions are uneven. However, teams should select evaluation metrics according to business requirements because a single metric may not capture every operational consequence of false positives and false negatives.<\/span><\/p>\nQuestion 64. Which SageMaker deployment strategy allows a new model version to receive a small portion of production traffic while the previous version continues serving most requests?<\/b><\/h3>\n
1)<\/b> Batch deployment<\/span>
\n<\/span>2)<\/b> Canary deployment<\/span>
\n<\/span>3)<\/b> Offline processing<\/span>
\n<\/span>4)<\/b> Data preprocessing<\/span><\/p>\nAnswer: 2) Canary deployment<\/b><\/p>\n
Explanation:<\/b><\/p>\n
A canary deployment introduces a new model version to a limited portion of production traffic before expanding its usage. This allows the team to observe metrics such as latency, errors, and model performance while keeping the previous version available for most requests. If problems are detected, the organization can reduce or stop traffic to the new version. Canary deployments are useful when teams want to reduce the exposure associated with a new release. The exact traffic allocation and monitoring criteria should be defined according to the application’s operational requirements.<\/span><\/p>\nQuestion 65. Which metric measures the average absolute difference between predicted and actual values in a regression problem?<\/b><\/h3>\n
1)<\/b> MAE<\/span>
\n<\/span>2)<\/b> Accuracy<\/span>
\n<\/span>3)<\/b> Precision<\/span>
\n<\/span>4)<\/b> Recall<\/span><\/p>\nAnswer: 1) MAE<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Mean Absolute Error, or MAE, calculates the average absolute difference between predicted values and actual target values. Because it uses absolute differences, positive and negative errors do not cancel each other out. MAE is expressed in the same general units as the target variable, which can make it relatively easy to interpret. It is useful for understanding the typical magnitude of prediction errors. Compared with metrics that square errors, MAE gives less additional weight to very large errors. The appropriate regression metric should depend on the consequences of different error sizes.<\/span><\/p>\nQuestion 66. A model produces predictions with different levels of confidence. Which technique helps determine whether predicted probabilities correspond well to actual outcome frequencies?<\/b><\/h3>\n
1)<\/b> Tokenization<\/span>
\n<\/span>2)<\/b> Binning<\/span>
\n<\/span>3)<\/b> Feature hashing<\/span>
\n<\/span>4)<\/b> Calibration<\/span><\/p>\nAnswer: 4) Calibration<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Calibration evaluates whether predicted probabilities correspond reasonably well to observed outcome frequencies. For example, among predictions assigned a probability of approximately 0.8, a well-calibrated classifier would produce the positive outcome at roughly that frequency over a sufficiently large representative sample. Calibration is particularly useful when prediction probabilities are consumed directly for decision-making, prioritization, or risk estimation. A model can have good classification discrimination while still having poorly calibrated probabilities. Therefore, calibration provides information that is different from metrics such as accuracy, precision, recall, or ROC-AUC.<\/span><\/p>\nQuestion 67. Which SageMaker capability can help explain how individual input features contribute to a model’s predictions?<\/b><\/h3>\n
1)<\/b> SageMaker Clarify<\/span>
\n<\/span>2)<\/b> SageMaker Batch Transform<\/span>
\n<\/span>3)<\/b> SageMaker Processing only<\/span>
\n<\/span>4)<\/b> SageMaker Model Registry<\/span><\/p>\nAnswer: 1) SageMaker Clarify<\/b><\/p>\n
Explanation:<\/b><\/p>\n
SageMaker Clarify provides model explainability capabilities that can help identify the contribution of input features to predictions. Understanding feature contributions can help data scientists investigate unexpected results and communicate how models use available information. Explainability is especially valuable in applications where model decisions require review or justification. Feature attribution does not necessarily prove that a feature causes an outcome; it describes how the model’s prediction is associated with the input features according to the selected explanation method. Teams should therefore interpret attribution results within the context of the model and dataset.<\/span><\/p>\nQuestion 68. Which AWS service can store custom Docker container images used by SageMaker training or inference workloads?<\/b><\/h3>\n
1)<\/b> Amazon S3<\/span>
\n<\/span>2)<\/b> Amazon ECR<\/span>
\n<\/span>3)<\/b> AWS CloudTrail<\/span>
\n<\/span>4)<\/b> Amazon EventBridge<\/span><\/p>\nAnswer: 2) Amazon ECR<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Amazon Elastic Container Registry, or ECR, is a managed registry for storing and managing container images. Machine learning teams can build custom Docker images containing specific libraries, frameworks, dependencies, or application code and push those images to ECR. SageMaker can then use the appropriate container image for supported training or inference workloads. ECR also integrates with AWS identity and access controls. Using a managed container registry helps organizations maintain consistent environments and version their custom machine learning images across development, testing, and production workflows.<\/span><\/p>\nQuestion 69. What is a primary advantage of using SageMaker Asynchronous Inference for suitable workloads?<\/b><\/h3>\n
1)<\/b> It eliminates all inference latency<\/span>
\n<\/span>2)<\/b> It requires every request to complete synchronously<\/span>
\n<\/span>3)<\/b> It supports workloads that may take longer to process<\/span>
\n<\/span>4)<\/b> It is exclusively designed for model training<\/span><\/p>\nAnswer: 3) It supports workloads that may take longer to process<\/b><\/p>\n
Explanation:<\/b><\/p>\n
SageMaker Asynchronous Inference is intended for inference requests that do not require an immediate synchronous response. It can be useful for workloads involving larger inputs or processing operations that may take longer to complete. Instead of forcing the client to wait for a traditional synchronous response, the request can be processed asynchronously and the result delivered after processing. This architecture can be appropriate for applications such as document processing or other computationally intensive predictions. Teams should still evaluate latency requirements and workload characteristics before selecting the inference mode.<\/span><\/p>\nQuestion 70. Which type of drift occurs when the statistical distribution of input features changes over time?<\/b><\/h3>\n
1)<\/b> Data drift<\/span>
\n<\/span>2)<\/b> Label encoding<\/span>
\n<\/span>3)<\/b> Model serialization<\/span>
\n<\/span>4)<\/b> Hyperparameter tuning<\/span><\/p>\nAnswer: 1) Data drift<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Data drift occurs when the distribution or characteristics of input data change compared with the data used to establish a baseline or train the model. For example, customer behavior, seasonal patterns, or external conditions can cause production feature distributions to change. Data drift does not automatically mean that model accuracy has decreased, but it can be an important warning signal. Monitoring production inputs can help identify significant changes that deserve investigation. Teams may then evaluate whether the model requires updated training data, revised features, or another form of maintenance.<\/span><\/p>\nQuestion 71. A machine learning pipeline should deploy a model only if its evaluation accuracy exceeds a predefined threshold. Which SageMaker Pipelines step is appropriate?<\/b><\/h3>\n
1)<\/b> Processing step<\/span>
\n<\/span>2)<\/b> Condition step<\/span>
\n<\/span>3)<\/b> Training step<\/span>
\n<\/span>4)<\/b> Model step<\/span><\/p>\nAnswer: 2) Condition step<\/b><\/p>\n
Explanation:<\/b><\/p>\n
A Condition step in SageMaker Pipelines allows a workflow to make decisions based on specified conditions. For example, a pipeline can evaluate a model’s accuracy and continue toward registration or deployment only when the result exceeds a defined threshold. This supports automated quality gates within machine learning workflows. Instead of deploying every newly trained model automatically, organizations can use conditions to enforce evaluation requirements. This approach can improve consistency in automated ML operations because deployment decisions become part of the pipeline rather than relying entirely on manual intervention.<\/span><\/p>\nQuestion 72. Which AWS service is commonly used to collect and visualize operational metrics and logs associated with AWS workloads?<\/b><\/h3>\n
1)<\/b> Amazon CloudWatch<\/span>
\n<\/span>2)<\/b> Amazon ECR<\/span>
\n<\/span>3)<\/b> AWS KMS<\/span>
\n<\/span>4)<\/b> SageMaker Feature Store<\/span><\/p>\nAnswer: 1) Amazon CloudWatch<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Amazon CloudWatch provides monitoring and observability capabilities for AWS resources and applications. It can collect metrics, logs, and other operational information that teams can use to understand system behavior. In machine learning deployments, CloudWatch can help monitor endpoint-related metrics such as invocation activity, errors, and resource utilization. Teams can also create alarms based on selected metrics so that operational issues can trigger notifications or automated actions. CloudWatch does not replace model-quality monitoring, but it can complement services such as SageMaker Model Monitor by providing infrastructure and application-level visibility.<\/span><\/p>\nQuestion 73. Which measure is commonly used to evaluate the ranking ability of a binary classification model across different classification thresholds?<\/b><\/h3>\n
1)<\/b> MAE<\/span>
\n<\/span>2)<\/b> Accuracy<\/span>
\n<\/span>3)<\/b> ROC-AUC<\/span>
\n<\/span>4)<\/b> MAPE<\/span><\/p>\nAnswer: 3) ROC-AUC<\/b><\/p>\n
Explanation:<\/b><\/p>\n
ROC-AUC measures the area under the receiver operating characteristic curve and evaluates how effectively a binary classifier separates positive and negative examples across different classification thresholds. A higher AUC generally indicates stronger ranking discrimination between the two classes. Unlike accuracy, ROC-AUC does not depend on selecting one particular classification threshold. However, it should be interpreted according to the dataset and business context. For highly imbalanced problems, teams may also examine precision-recall behavior because ROC-AUC alone may not fully represent the operational importance of positive-class performance.<\/span><\/p>\nQuestion 74. A team wants to perform training using historical feature values without introducing information that would only have become available after the prediction time. What concept should they consider?<\/b><\/h3>\n
1)<\/b> Point-in-time correctness<\/span>
\n<\/span>2)<\/b> Model serialization<\/span>
\n<\/span>3)<\/b> Container scaling<\/span>
\n<\/span>4)<\/b> Endpoint routing<\/span><\/p>\nAnswer: 1) Point-in-time correctness<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Point-in-time correctness ensures that training data contains only information that would have been available at the time the prediction was supposed to occur. This is important when historical features are retrieved from a feature store or other data source. If future information accidentally appears in the training dataset, the model may receive information that would not have been available during real-world inference. Such leakage can produce overly optimistic evaluation results. Maintaining correct timestamps and feature availability is therefore important when constructing historical training datasets for time-dependent machine learning problems.<\/span><\/p>\nQuestion 75. Which deployment architecture can expose multiple model versions through one endpoint while assigning different traffic percentages to them?<\/b><\/h3>\n
1)<\/b> SageMaker endpoint variants<\/span>
\n<\/span>2)<\/b> S3 lifecycle configuration<\/span>
\n<\/span>3)<\/b> AWS CloudTrail<\/span>
\n<\/span>4)<\/b> SageMaker Processing jobs<\/span><\/p>\nAnswer: 1) SageMaker endpoint variants<\/b><\/p>\n
Explanation:<\/b><\/p>\n
SageMaker endpoint variants allow multiple model configurations to operate behind an endpoint and can support traffic distribution between them. This capability is useful for controlled model testing, staged deployments, and comparing model versions under production conditions. For example, an organization can direct a smaller portion of traffic to a new model while maintaining service through an existing version. The traffic allocation can be adjusted based on observed operational and model metrics. Endpoint variants therefore provide a mechanism for managing multiple inference configurations without requiring every model version to have a completely separate application interface.<\/span><\/p>\nQuestion 76. What is the main purpose of a model card in a machine learning governance process?<\/b><\/h3>\n
1)<\/b> To increase GPU memory<\/span>
\n<\/span>2)<\/b> To document important information about a model<\/span>
\n<\/span>3)<\/b> To replace model training data<\/span>
\n<\/span>4)<\/b> To automatically retrain the model<\/span><\/p>\nAnswer: 2) To document important information about a model<\/b><\/p>\n
Explanation:<\/b><\/p>\n
A model card is used to document important information about a machine learning model, such as its intended use, limitations, evaluation information, and relevant considerations. Documentation supports transparency and helps stakeholders understand how a model was developed and where it may or may not be appropriate. Model cards do not directly improve computational performance or automatically retrain a model. Instead, they contribute to governance and responsible model management by providing structured information that can be reviewed before and during deployment.<\/span><\/p>\nQuestion 77. A model’s input data changes significantly after deployment, but the model’s relationship between inputs and the target also changes. What concept describes the changing relationship?<\/b><\/h3>\n
1)<\/b> Data serialization<\/span>
\n<\/span>2)<\/b> Feature scaling<\/span>
\n<\/span>3)<\/b> Concept drift<\/span>
\n<\/span>4)<\/b> Data compression<\/span><\/p>\nAnswer: 3) Concept drift<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Concept drift occurs when the relationship between input features and the target variable changes over time. This differs from simple data drift, where the distribution of input features changes without necessarily indicating that the underlying relationship with the target has changed. Concept drift can reduce model effectiveness because the patterns learned during training may no longer represent current conditions. Detecting it may require monitoring predictions, outcomes, or other relevant signals over time. When meaningful drift is identified, organizations may need to reassess training data, features, model assumptions, or retraining schedules.<\/span><\/p>\nQuestion 78. Which SageMaker capability is designed to test a new model version against a production model without sending the new model’s predictions to end users?<\/b><\/h3>\n
1)<\/b> Batch Transform<\/span>
\n<\/span>2)<\/b> Shadow testing<\/span>
\n<\/span>3)<\/b> Feature Store<\/span>
\n<\/span>4)<\/b> Model Registry<\/span><\/p>\n