835 questions total · 12pages · All types, answers revealed
A machine learning engineer is using SageMaker Debugger to detect if a neural network has dead ReLU units during training. Which built-in rule should they enable?
DeadRelu rule specifically detects dead ReLU units.
Why this answer
The 'DeadRelu' rule in Debugger monitors the fraction of ReLU activations that are zero and alerts if too many neurons are dead.
A machine learning engineer is building a regression model to predict house prices. The feature 'square_footage' has values ranging from 500 to 10,000, while 'num_bedrooms' ranges from 1 to 10. Which preprocessing step is most critical before training a model that uses gradient descent?
Standardization brings features to a common scale, crucial for gradient descent.
Why this answer
Gradient descent is sensitive to the scale of features because it updates weights proportionally to the feature values. With 'square_footage' (500–10,000) and 'num_bedrooms' (1–10), the large range difference causes the loss function's contours to be elongated, leading to slow or unstable convergence. Standardizing both features to zero mean and unit variance ensures each feature contributes equally to the gradient updates, enabling faster and more reliable optimization.
Exam trap
AWS often tests the distinction between scaling for gradient-based optimizers versus other preprocessing steps like encoding or transformation, trapping candidates who confuse feature scaling with handling outliers or categorical data.
How to eliminate wrong answers
Option B is wrong because applying a logarithmic transformation is not the most critical step for gradient descent; it is used to handle skewed distributions or multiplicative relationships, not to address feature scale differences. Option C is wrong because one-hot encoding is for categorical features, and 'num_bedrooms' is ordinal (integer-valued), not nominal; encoding it would create unnecessary sparsity and lose the natural ordering. Option D is wrong because imputing missing values is a general data cleaning step, but the question does not mention any missing data; the core issue here is feature scaling for gradient descent, not missingness.
A company has 200 small PyTorch models that are each used infrequently but need to be available for real-time inference. To minimize costs, they want to host all models on a single endpoint. Which SageMaker feature should they use?
Why this answer
Multi-model endpoints allow hosting hundreds of models on a single endpoint, automatically loading/unloading models based on traffic. Multi-container endpoints are for different containers, not multiple models. Batch and asynchronous are not real-time.
An organization wants to automate ML retraining using an event-driven architecture. Which THREE services should they combine? (Select THREE.)
SageMaker executes the actual retraining.
Why this answer
Amazon SageMaker provides the training jobs and pipelines that execute the ML retraining workflow. Amazon EventBridge acts as the event bus that triggers retraining based on events such as new data arrival or model drift detection. AWS Lambda serves as the lightweight compute layer that can preprocess events, invoke SageMaker APIs, or orchestrate conditional logic before starting a training job.
Exam trap
The trap here is that candidates often confuse AWS Glue as a compute trigger for ML retraining, but Glue is designed for batch ETL and lacks the event-driven, low-latency invocation capabilities required for this architecture.
A data scientist is using Amazon SageMaker Data Wrangler to prepare a dataset. The dataset contains a column with missing values, a column with outliers, and a column with text data. The scientist wants to use built-in transforms to handle these issues. Which THREE transforms are available in Data Wrangler for these tasks? (Select THREE.)
Data Wrangler includes transforms to impute missing values using mean, median, etc.
Why this answer
Data Wrangler provides built-in transforms for handling missing values (e.g., imputation), handling outliers (e.g., clipping), and processing text (e.g., tokenization). SMOTE is available for class imbalance, and one-hot encoding is for categorical features.
A machine learning engineer is building a real-time fraud detection pipeline using Amazon Kinesis Data Streams. The data must be prepared (e.g., feature engineering, normalization) before being fed into a SageMaker endpoint. Which TWO steps should the engineer implement to ensure low-latency data preparation?
Lambda can run custom code (e.g., Python) with low latency on each Kinesis record.
Why this answer
AWS Lambda can process records from Kinesis in near real-time for lightweight transformations like normalization and feature engineering. For streaming data, SageMaker batch transform is not real-time. Glue ETL is batch-oriented and adds latency.
Amazon Kinesis Data Analytics can perform SQL-based transformations in real-time. SageMaker Processing jobs are designed for offline processing.
A company wants to deploy a trained XGBoost model for batch inference on a large dataset stored in S3. The inference job should be cost-effective and does not require real-time responses. Which SageMaker inference option should they use?
Batch Transform is designed for batch inference on S3 data, cost-effective and no real-time requirement.
Why this answer
SageMaker Batch Transform is designed for batch inference on large datasets stored in S3, processing data in chunks and writing results to S3. It is cost-effective for non-real-time scenarios. Real-time endpoints are for low-latency inference.
Serverless is for on-demand, not batch. Asynchronous is for near-real-time with S3 input/output but still not ideal for large batch jobs.
A company is building a recommendation system and has trained a matrix factorization model using SageMaker. They want to evaluate the model's performance using precision at k (P@k) and recall at k (R@k). They have a test set of user-item interactions. The data scientist implements a custom evaluation script that computes these metrics, but the precision values are consistently zero. What is the most likely cause?
Correct: Without negative examples, precision is undefined or zero if no test items are in the recommendation list.
Why this answer
If the test set contains only positive interactions (i.e., every user-item pair in the test set is a ground-truth positive), then precision at k will be zero unless the model recommends exactly those items. Since the model's top-k recommendations are unlikely to perfectly match the test set's positive items for every user, precision (the fraction of recommended items that are relevant) will be zero. This is a known pitfall when evaluating implicit feedback models without negative samples.
Exam trap
The trap here is that candidates assume precision at k can be computed directly from a test set of positive interactions, overlooking that without negative labels, the metric becomes meaningless because the denominator (k) will always yield zero unless the model's top-k exactly matches the test positives.
How to eliminate wrong answers
Option A is wrong because even if the model outputs are not ranked correctly, precision at k would not be consistently zero—it would be some non-zero value if any relevant items appear in the top-k, just potentially lower than expected. Option B is wrong because overfitting would typically cause high training performance and poor generalization, but it would not force precision to be exactly zero; some relevant items could still appear in recommendations. Option D is wrong because a k value that is too large would increase recall (more items considered) but would not cause precision to be zero; precision would still be non-zero if any relevant items are among the top-k.
A machine learning engineer is troubleshooting a model that is producing unexpectedly low accuracy in production. The engineer examines the model's training data and finds that the distribution of the target variable in production is significantly different from the training set. What type of drift is the model experiencing?
Concept drift is a change in the statistical properties of the target variable.
Why this answer
Concept drift refers to any change in the statistical relationship between input features and the target variable, including changes in the target variable distribution. The scenario describes a change in the target variable distribution, which is a form of concept drift, specifically prior probability shift. Option A (Prior probability shift) is indeed a subtype of concept drift, but it is more specific; the question asks for the general drift type, making concept drift the best answer.
Option C (Data drift) refers to changes in the distribution of input features, not the target. Option D (Covariate shift) is a form of data drift where the input distribution changes while the conditional distribution P(Y|X) remains unchanged.
An e-commerce company uses Amazon SageMaker to deploy a real-time inference endpoint for product recommendations. The endpoint receives bursty traffic, with occasional spikes. The company wants to minimize cost while ensuring that latency remains under 100 ms. Which approach should the company take?
Multi-model endpoint with scaling and warm-up can handle bursts cost-effectively.
Why this answer
A multi-model endpoint with automatic scaling allows multiple models to share a single endpoint, reducing cost while handling bursty traffic. Configuring a warm-up period ensures new instances are fully initialized before receiving traffic, preventing cold-start latency spikes and keeping inference under 100 ms.
Exam trap
The trap here is that candidates confuse latency optimization techniques (like elastic inference) with scaling strategies, overlooking that bursty traffic requires dynamic scaling with warm-up to prevent cold-start latency spikes.
How to eliminate wrong answers
Option A is wrong because elastic inference accelerators reduce per-inference latency but do not address the need to scale out during traffic spikes; they add cost without solving the bursty traffic problem. Option B is wrong because scheduled scaling based on historical patterns cannot react to unpredictable spikes, leading to either over-provisioning or latency violations during unexpected bursts. Option C is wrong because deploying on one large instance creates a single point of failure and is cost-inefficient for bursty traffic; it either underutilizes resources during low traffic or fails to handle peak load without latency degradation.
A company is fine-tuning a large language model using LoRA with a Hugging Face estimator in SageMaker. They want to reduce memory usage during training. Which instance type is most cost-effective for this workload?
G5 instances are cost-effective for fine-tuning with LoRA, providing good performance at lower cost.
Why this answer
LoRA reduces the number of trainable parameters, allowing training on smaller GPUs. ml.g5 instances are optimized for machine learning inference and training with a good price-performance for fine-tuning.
A data scientist is preparing a dataset with a categorical feature that has 20 levels. The target variable is continuous. Which THREE encoding methods are appropriate for this scenario? (Select THREE.)
One-hot encoding creates binary columns for each category; works for any categorical feature.
Why this answer
One-hot encoding, ordinal encoding (if order exists), and target encoding are all applicable for categorical features with a continuous target. Label encoding is similar to ordinal but usually implies arbitrary order, but still acceptable; however, the question expects three of the listed. The three most directly appropriate are one-hot, ordinal, and target.
A media company uses SageMaker endpoints to serve a model that predicts video engagement. They have two production variants: Variant A (ml.c5.large) for regular traffic and Variant B (ml.c5.xlarge) for burst traffic. They use weighted routing (90% to A, 10% to B). Recently, during peak hours, Variant A's latency increase causes many requests to time out. The metrics show that both variants are under similar CPU load, but the number of concurrent requests to Variant A is very high. The team wants to ensure that burst traffic is handled properly without manual intervention. What should they do?
Autoscaling adjusts capacity based on load, preventing timeouts.
Why this answer
Changing to target tracking scaling based on the number of concurrent requests (or InvocationsPerInstance) ensures each variant scales based on its load. Option A (swap weights) doesn't fix scaling. Option C (p99 latency alarm) might trigger too late.
Option D (separate endpoint) is not necessary.
A company has a SageMaker endpoint that was deployed successfully and is in service. However, when the team sends test inferences using the InvokeEndpoint API, they receive a 500 internal server error. The endpoint logs in CloudWatch show a stack trace indicating 'OutOfMemoryError: Java heap space'. The model is a large XGBoost model (2 GB) and the endpoint is using an ml.m5.large instance with 8 GB of memory. What is the MOST likely cause and solution?
A 2 GB model plus runtime overhead (e.g., Java heap for XGBoost) can exceed 8 GB. Increasing instance memory resolves the out-of-memory error.
Why this answer
The OutOfMemoryError in Java heap space indicates that the model (2 GB) plus the runtime overhead of the XGBoost container and Java-based inference code exceed the available memory on the ml.m5.large instance (8 GB total, but not all is available for the Java heap). The most direct fix is to use a larger instance type, such as ml.m5.xlarge (16 GB) or ml.m5.2xlarge, to provide sufficient heap space for the model and inference operations.
Exam trap
The trap here is that candidates may incorrectly attribute the OutOfMemoryError to batch size or container compatibility, rather than recognizing that the instance's memory is insufficient for the model size and Java heap overhead.
How to eliminate wrong answers
Option A is wrong because batch size configuration is not applicable to real-time InvokeEndpoint requests (which are single inference calls), and reducing batch size would not resolve a Java heap space error caused by model size and overhead. Option C is wrong because the model is explicitly stated as XGBoost, not a Transformer model, and XGBoost runs efficiently on CPU instances; GPU instances are not required. Option D is wrong because SageMaker provides native support for XGBoost via built-in containers, and the error is a memory issue, not a compatibility issue with the container.
A company uses AWS Glue to run ETL jobs that prepare data for machine learning. The data is stored in Amazon S3 in Parquet format. A data engineer notices that the Glue job is running slowly and consuming a lot of resources. What is the MOST cost-effective way to improve the performance of the Glue job?
G.1X offers more memory, reducing memory-related bottlenecks without increasing DPU count.
Why this answer
Increasing the number of DPUs (Data Processing Units) in AWS Glue can improve parallelism and reduce job runtime, but it increases cost. Using G.1X worker type with more memory per worker can improve performance without increasing DPU count, offering better resource utilization. Switching to CSV may degrade performance.
Using partition pruning on the source data can reduce data scanned but may not address resource consumption.
A company deploys a SageMaker model using AWS KMS for encryption at rest. They have a compliance requirement to rotate the KMS key every year without causing downtime for the inference endpoint. Which approach should they take?
Automatic rotation rotates the key material without changing the key ID, causing no downtime.
Why this answer
AWS KMS supports automatic key rotation, which creates new backing keys annually while retaining the same key ID and metadata. This ensures that the SageMaker endpoint continues to use the same KMS key alias and configuration, so no endpoint update or downtime is required. Automatic rotation satisfies the compliance requirement without any manual intervention or endpoint recreation.
Exam trap
The trap here is that candidates may think rotating a KMS key requires creating a new key and updating the resource (Option B), or that manual recreation is necessary (Option C), when in fact AWS KMS automatic key rotation handles the rotation seamlessly without any endpoint modification or downtime.
How to eliminate wrong answers
Option A is wrong because AWS Certificate Manager (ACM) is for managing SSL/TLS certificates, not for encryption at rest of SageMaker model data; it does not provide KMS key rotation capabilities. Option B is wrong because creating a new KMS key and updating the endpoint configuration would require a deployment update, which can cause a brief interruption or require a rolling update, and it does not leverage the simpler automatic rotation mechanism. Option C is wrong because manually rotating the key by recreating the endpoint would cause downtime during the recreation process, violating the no-downtime requirement.
Which TWO data storage options are commonly used by Amazon SageMaker Feature Store for offline and online storage?
S3 is the default offline store for large historical feature data.
Why this answer
Amazon SageMaker Feature Store uses Amazon S3 as the default offline storage layer because it provides durable, scalable, and cost-effective object storage for large volumes of historical feature data. Amazon DynamoDB is used as the default online storage layer because it offers low-latency, single-digit millisecond read/write performance required for real-time inference serving.
Exam trap
The trap here is that candidates often confuse Amazon ElastiCache (a caching layer) with the primary online storage service, or assume Amazon Redshift is used for offline storage due to its analytical capabilities, but SageMaker Feature Store specifically integrates DynamoDB for online and S3 for offline storage as first-class options.
A data science team is building a model to predict fraudulent transactions. The dataset has 1 million legitimate transactions and only 1,000 fraudulent ones. They plan to use Amazon SageMaker to train a model. Which data preparation technique should they apply to address the severe class imbalance before training?
SMOTE creates synthetic examples by interpolating between existing minority instances, reducing overfitting risk.
Why this answer
SMOTE (Synthetic Minority Oversampling Technique) is the correct choice because it generates synthetic fraudulent samples by interpolating between existing minority class instances in feature space, rather than simply duplicating records. This creates more diverse and realistic training data, reducing overfitting risk while addressing the severe 1:1000 class imbalance. Amazon SageMaker's built-in algorithms and data processing capabilities can easily integrate SMOTE-applied datasets for training.
Exam trap
AWS often tests the misconception that simple random oversampling (Option B) is sufficient, but the trap is that it causes overfitting, whereas SMOTE's synthetic generation provides better generalization for imbalanced datasets.
How to eliminate wrong answers
Option A is wrong because data augmentation using image transformations (e.g., rotations, flips) is specific to image data and does not apply to tabular fraud detection datasets; it introduces irrelevant noise and breaks feature relationships. Option B is wrong because randomly oversampling the fraudulent class by duplicating existing records leads to overfitting, as the model simply memorizes the exact same fraud patterns without learning generalizable features. Option D is wrong because randomly undersampling the legitimate class to 1,000 samples discards 999,000 legitimate transactions, causing massive information loss and severely degrading model performance on the majority class.
A data scientist is trying to create a SageMaker endpoint configuration with 6 instances of ml.c5.large for a production variant. The creation fails with the error shown in the exhibit. Which action should the data scientist take to resolve this issue?
Increasing the quota allows provisioning the requested number of instances.
Why this answer
The error indicates that the requested number of instances exceeds the service quota for ml.c5.large for real-time endpoints. AWS enforces default limits on instance counts per instance type per region. Requesting a quota increase via the Service Quotas console is the correct action to raise the limit and allow the deployment of 6 instances.
Exam trap
The trap here is that candidates may confuse service quotas with resource availability, thinking that deleting unused endpoints or splitting configurations will free up capacity, when in fact the quota is a hard limit that must be explicitly increased.
How to eliminate wrong answers
Option A is wrong because creating two separate endpoint configurations does not bypass the service quota; the total instance count across all endpoints still counts against the same quota. Option C is wrong because using a different instance type like ml.m5.large does not inherently have a higher limit; each instance type has its own default quota, and the limit for ml.m5.large may also be insufficient or unknown without checking. Option D is wrong because deleting unused endpoints does not increase the quota for ml.c5.large; it only frees up currently used instances, but the quota itself remains unchanged.
A machine learning team is setting up Model Monitor for a deployed model. Which THREE factors should they consider when configuring the monitoring schedule? (Select three.)
SNS notifications can alert teams when violations are detected.
Why this answer
Amazon SageMaker Model Monitor can be configured to send notifications via Amazon SNS when monitoring violations are detected. This allows the team to proactively respond to data drift or quality issues without manually polling the monitoring results.
Exam trap
The trap here is that candidates assume monitoring must run daily (Option B) because of common best practices, but the exam tests that the schedule should be based on data arrival patterns, not a fixed minimum frequency.
A machine learning engineer observes that model performance on a SageMaker endpoint has degraded over the past week. Ground truth labels are available with a 2-day delay. The engineer wants to automatically trigger a retraining pipeline when prediction quality drops below an acceptable threshold. Which approach is most appropriate?
Model Quality Monitor evaluates predictions against ground truth; CloudWatch alarm on quality metric triggers retraining.
Why this answer
SageMaker Model Monitor's Model Quality Monitor is specifically designed to compare model predictions against ground truth labels (available with a 2-day delay) and track metrics like accuracy, precision, recall, or F1 score. You can configure a CloudWatch alarm on a metric such as 'accuracy' dropping below a threshold, which triggers an AWS Lambda function to start the retraining pipeline. This automates the detection of prediction quality degradation and the retraining response without manual intervention.
Exam trap
The trap here is that candidates confuse Data Quality Monitor (which monitors input data drift) with Model Quality Monitor (which monitors prediction accuracy against ground truth), leading them to choose Option C incorrectly.
How to eliminate wrong answers
Option B is wrong because manually evaluating the model weekly is not automated and does not meet the requirement to automatically trigger retraining when prediction quality drops; it introduces latency and human error. Option C is wrong because Data Quality Monitor detects drift in input data distribution (e.g., feature skew), not in prediction quality against ground truth labels, so it cannot directly measure model performance degradation. Option D is wrong because SageMaker Clarify is used for bias detection and explainability, not for monitoring prediction quality or triggering retraining based on performance metrics.
A data scientist is performing feature engineering for a dataset with both numerical and categorical features. The data scientist wants to apply transformations that preserve the interpretability of the features. Which TWO transformations should the data scientist use? (Select TWO)
Log transformation reduces skewness while keeping feature order.
Why this answer
Log transformation is correct because it reduces skewness in numerical features by compressing the scale of large values, making the distribution more normal while preserving the original feature's interpretability (e.g., a log-transformed income value still relates to income). This is a monotonic transformation, so the order of values is maintained, and the feature remains directly understandable.
Exam trap
AWS often tests the misconception that one-hot encoding always preserves interpretability (it does, but the question pairs it with target encoding as a distractor), leading candidates to select one-hot encoding instead of recognizing that standard scaling is the correct second choice for numerical features.
A team wants to monitor the number of requests and latency of their SageMaker endpoint using a unified dashboard. Which AWS service should they use to create a custom dashboard with these metrics?
CloudWatch Dashboards can display real-time and historical metrics from SageMaker endpoints in a customizable layout.
Why this answer
Amazon CloudWatch Dashboards allow you to create custom views of metrics from any source, including SageMaker endpoint metrics like Invocations and Latency. SageMaker itself does not provide a dashboard for these metrics.
A data scientist is using Amazon SageMaker Data Wrangler to create a feature engineering pipeline for a dataset with both numeric and categorical features. The scientist wants to apply transformations that are appropriate for a linear model. Which THREE transformations should the scientist apply? (Choose THREE.)
High correlation between features can cause multicollinearity, making linear model coefficients unstable and hard to interpret.
Why this answer
Linear models assume features are numeric, scaled, and not highly correlated. StandardScaler ensures all numeric features have comparable scales. One-hot encoding converts categorical features into binary columns without imposing ordinality.
Removing highly correlated features reduces multicollinearity which can destabilize coefficient estimates.
A team uses SageMaker real-time endpoints for inference. They want to deploy a new model version and compare its performance with the current version under live traffic without affecting user experience. Which method should they use?
Shadow testing duplicates traffic to a shadow variant without serving it to users, allowing safe comparison.
Why this answer
Shadow testing (or shadow deployment) sends a copy of live traffic to the new model variant while the current variant serves the actual response. The shadow variant's performance can be monitored without impacting the user.
A company uses Amazon SageMaker Ground Truth to create labeled datasets for object detection. The output must be in COCO format for downstream model training. How should the data preparation process be configured?
Ground Truth supports COCO output for object detection tasks.
Why this answer
Amazon SageMaker Ground Truth natively supports outputting object detection labeling jobs in COCO format. When you select 'Object Detection' as the task type, the labeling job configuration includes an option to specify 'COCO' as the output format, which automatically structures the labeled data into the required COCO JSON schema without any post-processing.
Exam trap
The trap here is that candidates assume post-processing is always required for format conversion, overlooking that Ground Truth can directly output COCO format when the correct task type and output format are selected in the labeling job configuration.
How to eliminate wrong answers
Option A is wrong because Ground Truth does not provide a built-in transformation to convert its default JSON output to COCO format; the conversion must be handled externally. Option B is wrong because while AWS Lambda can be used for custom transformations, it is not a pre-built solution for this specific conversion; using a Lambda function would require writing custom code and is not the recommended or simplest approach. Option C is wrong because writing a custom SageMaker Processing script is an unnecessary extra step; Ground Truth can directly output COCO format, eliminating the need for any post-labeling transformation.
A data engineer needs to convert a JSON dataset to Parquet format for efficient querying with Amazon Athena. The JSON files are in an S3 bucket. Which service can perform this conversion with minimal coding?
Glue Studio's drag-and-drop interface enables JSON to Parquet conversion with minimal coding.
Why this answer
AWS Glue Studio with a visual job is the correct choice because it provides a no-code, drag-and-drop interface to create ETL jobs that can read JSON from S3 and write it as Parquet, with built-in schema inference and transformation capabilities. This minimizes coding effort while leveraging Glue's serverless Spark engine for efficient conversion, making it ideal for preparing data for Athena queries.
Exam trap
The trap here is that candidates often confuse AWS Glue Studio with AWS Glue DataBrew or assume that any AWS service with 'processing' in its name (like SageMaker Processing) is suitable for simple ETL tasks, overlooking the specific no-code visual job capability of Glue Studio.
How to eliminate wrong answers
Option A is wrong because Amazon SageMaker Processing is designed for data preprocessing and model training workflows within the ML pipeline, not for simple file format conversion; it requires writing custom processing scripts and managing infrastructure, which adds unnecessary complexity. Option B is wrong because Amazon EMR is a managed Hadoop/Spark cluster that can perform the conversion, but it requires provisioning and configuring a cluster, writing Spark or Hive code, and managing lifecycle, which is far more coding and operational overhead than a visual job. Option C is wrong because AWS Lambda has a maximum execution time of 15 minutes and a deployment package size limit, making it impractical for converting large JSON datasets to Parquet; it also requires custom Python code with libraries like PyArrow or Pandas, which is not minimal coding.
A company uses SageMaker endpoints with auto-scaling. The endpoint is experiencing high latency during peak hours. The metrics show CPU utilization is low but memory is high. What is the most likely cause?
CPU is low so scaling not triggered, but memory high indicates need for more instances.
Why this answer
The auto-scaling policy is based on CPU utilization, which remains low during the memory-bound issue. Since the scaling trigger is not met, the endpoint does not add more instances to handle the increased load, leading to high latency. Memory pressure without CPU spikes indicates the bottleneck is memory, not compute, so a CPU-based metric fails to scale appropriately.
Exam trap
The trap here is that candidates assume high latency always means CPU is the bottleneck, but the exam tests understanding that auto-scaling must be based on the correct metric; memory pressure can cause latency without CPU spikes, and a CPU-based policy will fail to scale.
How to eliminate wrong answers
Option A is wrong because a memory leak would cause memory to increase over time, not specifically during peak hours, and would likely degrade performance gradually rather than cause latency spikes tied to load. Option C is wrong because insufficient network bandwidth would manifest as network-related errors or timeouts, not high memory utilization with low CPU; network metrics would show saturation. Option D is wrong because a VPC without a NAT gateway affects outbound internet access, not inbound inference requests to the endpoint; SageMaker endpoints in a VPC can receive traffic via VPC endpoints or public endpoints without a NAT gateway.
Which TWO actions are recommended best practices when preparing training data for a machine learning model in AWS? (Choose two.)
Missing values can cause errors or bias if not addressed.
Why this answer
Missing values can introduce bias or cause algorithms to fail, so handling them (e.g., via imputation or removal) is a critical data preparation step in AWS SageMaker. Option D is correct because splitting data into training, validation, and test sets allows you to evaluate model performance on unseen data and prevent overfitting, which is a standard practice in SageMaker's built-in algorithms and training jobs.
Exam trap
The trap here is that candidates assume all outliers must be removed (Option A) or that normalization is always required (Option E), but the exam tests nuanced understanding that these steps depend on the algorithm and data characteristics, not blanket rules.
A team is using SageMaker Pipelines to automate retraining and deployment. They want to trigger the pipeline automatically when new training data is available in an S3 bucket. Which approach should they use?
EventBridge can detect S3 events and start pipeline executions.
Why this answer
Amazon EventBridge can directly capture S3 PutObject events and invoke a SageMaker Pipeline execution as a target. This provides a fully event-driven, serverless integration without polling or manual intervention, aligning with best practices for automating ML workflows when new data arrives.
Exam trap
The trap here is that candidates may overcomplicate the solution by choosing Step Functions (Option D) for orchestration, not realizing that EventBridge provides a simpler, event-driven trigger without the need for polling or additional state machines.
How to eliminate wrong answers
Option B is wrong because registering a pipeline as a model package in SageMaker Model Registry is for versioning and managing trained models, not for triggering pipeline executions based on S3 events. Option C is wrong because a cron job runs on a fixed schedule, which is inefficient and may miss data arrivals or run unnecessarily, whereas the requirement is to trigger only when new data appears. Option D is wrong because using AWS Step Functions to poll S3 introduces latency, cost, and complexity compared to the native event-driven approach with EventBridge, which reacts instantly to S3 events.
A data scientist is preparing a dataset for a regression model that predicts house prices. The dataset includes a `neighborhood` feature with 500 distinct categories. The data scientist wants to encode this feature without increasing dimensionality too much and while capturing the target relationship. Which encoding technique should be used?
Target encoding captures target relationship with low dimensionality.
Why this answer
Target encoding (mean encoding) is the correct choice because it replaces each of the 500 neighborhood categories with the mean of the target variable (house price) for that category. This captures the relationship between the neighborhood and the target while adding only one new feature column, thus avoiding the massive dimensionality explosion that would occur with one-hot encoding (which would create 500 binary columns).
Exam trap
AWS often tests the trade-off between dimensionality and information retention, and the trap here is that candidates may choose one-hot encoding out of habit, failing to recognize that 500 categories make it impractical, or choose label encoding because it seems simple, ignoring the ordinal assumption it imposes.
How to eliminate wrong answers
Option B (One-hot encoding) is wrong because it would create 500 binary columns, drastically increasing dimensionality and leading to the curse of dimensionality, sparsity, and overfitting. Option C (Frequency encoding) is wrong because it replaces categories with their count/frequency, which does not capture the relationship with the target variable (house price) and loses predictive signal. Option D (Label encoding) is wrong because it assigns arbitrary integer labels (e.g., 1, 2, 3) that imply an ordinal relationship, which is inappropriate for a nominal feature like neighborhood and can mislead the regression model into assuming a false order.
A company wants to deploy a single model that processes images from a production line. The images are uploaded to an S3 bucket every few minutes, and the inference results must be stored back to S3. The team wants to avoid paying for idle compute and prefers a fully managed, on-demand solution. Which SageMaker inference option should they use?
Asynchronous inference is ideal for near-real-time, event-driven workloads with S3 input/output and scales to zero when idle.
Why this answer
Asynchronous inference is designed for this use case: it processes images from S3 input, writes results to S3 output, scales to zero when idle, and is fully managed. Real-time endpoints are always running and incur cost when idle. Batch transform is not event-driven.
Serverless inference is event-driven but has a payload limit and cold start that may not be suitable for image payloads.
A data scientist is working on a time series forecasting problem. The dataset contains a column 'sales' with occasional negative values due to returns. The model expects non-negative input. Which data preparation step should be taken?
Sets returns to zero, which is appropriate for sales data.
Why this answer
Clipping negative sales values to zero directly addresses the model's requirement for non-negative input while preserving the data's temporal structure. This approach is appropriate for time series forecasting where returns cause occasional negative values, as it treats returns as zero sales rather than removing or distorting the data points.
Exam trap
AWS often tests the misconception that removing or imputing negative values is safe in time series, but the trap here is that these actions break temporal dependencies and introduce bias, whereas clipping preserves the sequence structure.
How to eliminate wrong answers
Option B is wrong because applying a log transformation after adding a constant does not guarantee non-negative values; it only compresses the scale and can introduce bias, especially with negative values that require arbitrary shifting. Option C is wrong because removing all rows with negative sales values disrupts the time series continuity and can lead to loss of important temporal patterns, such as seasonality or trends. Option D is wrong because imputing negative values with the mean introduces statistical bias and distorts the underlying distribution, which is particularly problematic in time series where data points are sequentially dependent.
Refer to the exhibit. A user is unable to invoke a SageMaker endpoint. The IAM policy shown is attached to the user. Which permission is missing to allow invocation?
InvokeEndpoint is required to send inference requests.
Why this answer
To invoke a SageMaker endpoint, the user needs the `sagemaker:InvokeEndpoint` permission. The IAM policy shown lacks this action, which is required for making real-time inference requests to the endpoint. Without it, any attempt to call the endpoint via the SDK or CLI will fail with an access denied error.
Exam trap
AWS often tests the distinction between read-only permissions (like `DescribeEndpoint` or `ListEndpoints`) and the specific action required to perform an operation, leading candidates to confuse metadata access with actual invocation capability.
How to eliminate wrong answers
Option B is wrong because `sagemaker:DescribeEndpoint` only allows retrieving metadata about an endpoint, not invoking it for inference. Option C is wrong because `sagemaker:CreateEndpoint` is for creating new endpoints, not for sending inference requests to an existing one. Option D is wrong because `sagemaker:ListEndpoints` only lists endpoints in the account, which does not grant the ability to invoke them.
A company uses Amazon SageMaker Feature Store to store features for a real-time recommendation model. The feature data is updated continuously, and the model must use the most recent feature values for each user at inference time. Which type of Feature Store should the company use for serving features to the model?
An Amazon SageMaker Feature Store online store serves features with low-latency access for real-time inference, satisfying the requirement that the recommendation model must use the most recent feature values per user at inference time. Unlike the offline store, which is optimised for batch retrieval and historical analysis, the online store provides a synchronised, low-latency key-value lookup for continuously updated feature data.
Why this answer
Online store provides low-latency access to the latest feature values for real-time inference. Offline store is for batch training and analytics. Point-in-time queries are for historical retrieval, not real-time serving.
A company uses Amazon SageMaker Data Wrangler to create a data flow for a classification model. The dataset contains a high-cardinality categorical feature 'product_id' with 50,000 unique values. The data scientist wants to reduce dimensionality while preserving predictive power. Which approach is most effective?
Target encoding condenses information into a single numerical feature while retaining predictive signals.
Why this answer
Target encoding is the most effective approach for high-cardinality categorical features because it replaces each category with the mean of the target variable, preserving predictive signal while drastically reducing dimensionality. In SageMaker Data Wrangler, this can be implemented using the 'Encode categorical' transform with the 'Target encoding' option, which avoids the explosion of features caused by one-hot encoding and retains the relationship between product IDs and the target.
Exam trap
AWS often tests the misconception that feature hashing is always safe for high-cardinality features, but the trap here is that hash collisions can degrade model performance, making target encoding a better choice when the target variable is available and predictive.
How to eliminate wrong answers
Option A is wrong because one-hot encoding on a feature with 50,000 unique values would create 50,000 binary columns, leading to extreme dimensionality and sparsity, which degrades model performance and increases computational cost. Option C is wrong because feature hashing maps product IDs to a fixed number of buckets (e.g., 100), which can cause hash collisions and loss of information, reducing predictive power compared to target encoding. Option D is wrong because dropping the column entirely discards all predictive information contained in the product IDs, which is likely to harm model accuracy.
A financial services company deploys a credit risk model using an Amazon SageMaker endpoint with data capture enabled. The model uses a custom container. The compliance team requires that all inference requests and responses are logged to an S3 bucket with server-side encryption using AWS KMS. The IAM role for the endpoint has the following policy. What must be added to meet the compliance requirement?
These permissions are necessary to write to a KMS-encrypted bucket.
Why this answer
The IAM role for the SageMaker endpoint needs permissions to generate a data key (kms:GenerateDataKey) for encrypting captured data and to decrypt (kms:Decrypt) the KMS key when writing to the S3 bucket. Without these, the endpoint cannot use the customer-managed KMS key for server-side encryption, even if the bucket policy allows it.
Exam trap
The trap here is that candidates often assume enabling S3 default encryption (Option C) is sufficient, but SageMaker data capture requires explicit KMS permissions in the endpoint's IAM role to use the customer-managed key.
How to eliminate wrong answers
Option B is wrong because s3:PutObjectAcl is not required for server-side encryption with KMS; it is used for managing object-level access control lists, not encryption. Option C is wrong because enabling S3 default encryption on the bucket does not satisfy the requirement for server-side encryption using AWS KMS for data captured by SageMaker; the endpoint must explicitly use the KMS key via the IAM role. Option D is wrong because modifying the container to handle encryption internally would bypass the managed data capture feature and is not necessary; SageMaker data capture already supports KMS encryption natively.
A machine learning team needs to monitor a deployed model for both data drift and concept drift. Which TWO approaches should they implement? (Select TWO.)
Data quality monitoring detects drift in input features.
Why this answer
SageMaker Model Monitor can be configured for data quality (data drift) and model quality (concept drift) monitoring. Data drift monitors input distribution changes, while model quality monitors prediction accuracy against ground truth.
An MLOps team is designing a CI/CD pipeline for deploying machine learning models to production on Amazon SageMaker. They want to ensure that the deployment process is automated and that models are automatically rolled back if performance degrades. Which of the following AWS services or features should they use to achieve this? (Choose THREE.)
Model Registry manages model versions and approvals.
Why this answer
Amazon SageMaker Model Registry is correct because it provides a centralized catalog for managing, versioning, and approving ML models. It enables automated deployment by triggering CI/CD pipelines when a model version is approved, and supports automatic rollback by allowing you to revert to a previous approved version if performance degrades, as detected by monitoring metrics.
Exam trap
The trap here is that candidates may confuse SageMaker Ground Truth (a data labeling service) or CloudTrail (an auditing service) with the core MLOps components needed for automated deployment and rollback, overlooking that Model Registry, Pipelines, and CloudWatch are the precise services that form the CI/CD and monitoring backbone.
A healthcare company is developing a predictive model to identify patients at risk of readmission within 30 days after discharge. The dataset contains electronic health record (EHR) data from multiple hospitals, stored as Parquet files in Amazon S3. The data includes patient demographics, diagnoses (ICD-10 codes), medications, lab results, and length of stay. A data scientist notices that the 'lab_result' column has a high number of null values (over 60%) because some tests are not applicable to all patients. Additionally, the 'diagnosis_code' column has over 10,000 unique ICD-10 codes. The company wants to build a model that complies with HIPAA and performs well. The data scientist must prepare the features efficiently using AWS services. Which combination of steps should the data scientist take? (Assume the company can use any AWS service.)
Predictive imputation leverages other features to estimate missing values, retaining data. Count encoding reduces the cardinality of diagnosis codes.
Why this answer
It uses AWS Glue ETL to impute missing lab results with a predictive model (XGBoost), which is appropriate for high missingness (>60%) where simple imputation would bias the model, and applies count encoding to the high-cardinality diagnosis codes (10,000+ unique values) to avoid the dimensionality explosion of one-hot encoding while preserving frequency information. This approach balances HIPAA compliance (data stays within AWS) with model performance.
Exam trap
The trap here is that candidates often choose simple mean imputation (Option B) or dropping rows (Option C) without considering the impact of high missingness on bias and data loss, or they overcomplicate encoding (Option D) without recognizing that ordinal encoding implies a false order for categorical codes.
How to eliminate wrong answers
Option B is wrong because replacing 60%+ missing lab results with the overall mean ignores the non-random missingness (tests not applicable to all patients) and introduces severe bias, and grouping 10,000+ ICD-10 codes into only 20 categories based on clinical relevance loses granularity and may not reflect readmission risk patterns. Option C is wrong because dropping all records with null lab results would discard over 60% of the data, leading to massive data loss and a non-representative dataset, and one-hot encoding 10,000+ diagnosis codes creates an unmanageable feature space (sparse matrix) that degrades model performance. Option D is wrong because KNN imputation on a dataset with >60% missingness in the same column is computationally expensive and unreliable (neighbors themselves may have missing values), and ordinal encoding based on ICD-10 chapter order imposes an arbitrary ordinal relationship that does not reflect clinical risk or readmission likelihood.
An ML engineer runs the CLI command shown in the exhibit. However, the training job fails immediately with an error: 'Unable to assume role'. What is the most likely cause?
Without proper trust policy, SageMaker cannot assume the role, causing immediate failure.
Why this answer
The 'Unable to assume role' error indicates that SageMaker cannot assume the IAM role specified in the CLI command. This is a trust policy issue: the role's trust policy must include SageMaker as a trusted service (i.e., `"Service": "sagemaker.amazonaws.com"`). Without this, SageMaker is not authorized to assume the role, regardless of the role's permissions.
Exam trap
AWS often tests the distinction between IAM role permissions (what the role can do) and trust policies (who can assume the role), leading candidates to mistakenly select a permission-related option when the error is about trust.
How to eliminate wrong answers
Option A is wrong because the error is about assuming the role, not about the role's permissions to create the training job; permission errors would appear as 'AccessDenied' or similar, not 'Unable to assume role'. Option B is wrong because a missing ECR image would cause an error like 'Image not found' or 'RepositoryNotFoundException', not an assume role error. Option C is wrong because a non-existent S3 bucket would result in an error like 'NoSuchBucket' or 'AccessDenied' when SageMaker tries to access it, not an assume role failure.
A company is using Amazon SageMaker to deploy a model for real-time inference. The model requires access to a private S3 bucket that contains reference data. The company wants to ensure that the endpoint can access the S3 bucket without using a public internet connection. Which TWO actions should they take? (Select TWO.)
VPC endpoints allow private connectivity to S3 without internet.
Why this answer
Attaching the SageMaker endpoint to a VPC with a VPC endpoint for S3 (Gateway type) allows the endpoint to access the S3 bucket using AWS's private network, bypassing the public internet. This ensures traffic stays within the AWS backbone, meeting the requirement for no public internet connection. Option C is also correct because the SageMaker execution role must have an IAM policy with s3:GetObject permissions to authorize the read access to the private S3 bucket, which is a prerequisite for any S3 operation.
Exam trap
The trap here is that candidates often confuse VPC endpoints (which keep traffic private) with NAT gateways or internet gateways (which route traffic over the public internet), and they may overlook the mandatory IAM permissions required for S3 access even when using a VPC endpoint.
A team is deploying a model that requires low-latency inference for real-time predictions. They are using a SageMaker endpoint with a single instance. During testing, they observe high latency. Which change would most effectively reduce latency?
Correct: Larger instances provide more CPU/GPU for faster inferences.
Why this answer
Switching to a larger instance type (Option D) directly increases the compute and memory resources available to the SageMaker endpoint, which reduces inference latency by allowing the model to process requests faster. Since the team is using a single instance, scaling up is the most straightforward way to handle the computational load and meet real-time latency requirements.
Exam trap
The trap here is that candidates often confuse scaling up (larger instance) with scaling out (multiple instances) or assume that Elastic Inference always reduces latency, but Elastic Inference adds network latency and is better for cost savings on large models, not for minimizing per-request latency.
How to eliminate wrong answers
Option A is wrong because a multi-model endpoint is designed to host multiple models on a single instance to improve resource utilization, not to reduce latency for a single model; it can actually increase latency due to model loading and unloading overhead. Option B is wrong because Elastic Inference attaches a separate accelerator for deep learning inference, but it adds network round-trip time between the instance and the accelerator, which can increase latency for real-time predictions, especially for small models or low-latency requirements. Option C is wrong because SageMaker Batch Transform is an asynchronous batch processing service that processes large datasets offline, not suitable for real-time, low-latency predictions.
A data engineer is using Amazon SageMaker Ground Truth to create a labeled dataset for an object detection task. The dataset contains millions of images, and the labeling budget is limited. Which approach can reduce labeling costs while maintaining high model accuracy?
Active learning iteratively selects the most valuable samples for human labeling, reducing cost while maintaining model performance.
Why this answer
Active learning in Ground Truth selects the most informative samples (e.g., uncertain predictions) for human labeling, reducing the number of labels needed while maximizing model improvement. This is a built-in feature.
A company wants to deploy a PyTorch model on SageMaker using the NVIDIA Triton Inference Server for GPU acceleration. They have an existing Triton configuration. Which approach should they take?
The SageMaker Triton DLC is pre-configured for Triton and supports PyTorch models.
Why this answer
AWS provides a pre-built SageMaker Triton Inference Server container as part of the Deep Learning Containers (DLCs), which is optimized for GPU acceleration and supports the existing Triton configuration without modification. This container integrates directly with SageMaker hosting endpoints, enabling seamless deployment of PyTorch models with Triton's features like dynamic batching and model concurrency.
Exam trap
The trap here is that candidates may assume SageMaker Neo is a universal compilation tool for any inference server, but Neo is specifically for hardware-specific optimization and does not support Triton's runtime environment, leading them to incorrectly select Option A.
How to eliminate wrong answers
Option A is wrong because SageMaker Neo compiles models for specific hardware targets (e.g., Intel, ARM) and does not support compilation for the NVIDIA Triton Inference Server; Neo is designed for edge devices and does not integrate with Triton's serving architecture. Option B is wrong because while packaging Triton as a custom container is possible, using SageMaker batch transform is not the recommended approach for real-time inference with GPU acceleration; batch transform is for offline, asynchronous processing, not for low-latency serving. Option D is wrong because installing Triton at runtime on the standard PyTorch container is inefficient and error-prone; it adds startup latency, may cause dependency conflicts, and bypasses the pre-optimized, tested Triton container that AWS provides.
A team uses AWS Auto Scaling for a SageMaker real-time endpoint. They notice that when scaling in, the latest instance is always terminated first, causing disruption to recent requests. How can they configure the scaling policy to terminate the oldest instance first?
You can set the termination policy to 'OldestInstance' in the scaling policy configuration.
Why this answer
AWS Auto Scaling for SageMaker endpoints supports a termination policy of 'OldestInstance', which explicitly instructs the scaling process to terminate the instance that has been running the longest. By default, Auto Scaling terminates the newest instance (the default termination policy), which can disrupt recent requests. Configuring the termination policy to 'OldestInstance' ensures that the oldest, most stable instance is removed first, minimizing disruption to in-flight requests.
Exam trap
The trap here is that candidates assume the default termination policy is 'OldestInstance' or that lifecycle hooks can influence instance selection, when in fact the default is 'NewestInstance' and lifecycle hooks only add a delay, not a selection rule.
How to eliminate wrong answers
Option B is wrong because the default behavior of AWS Auto Scaling is to terminate the newest instance first (the 'Default' termination policy), not the oldest, so action is needed to change this. Option C is wrong because lifecycle hooks are used to perform custom actions (e.g., draining connections) before an instance is terminated or launched, but they do not control which instance is selected for termination; they only add a pause in the lifecycle. Option D is wrong because AWS CloudFormation is an infrastructure-as-code service for provisioning resources, not a mechanism to configure the termination policy of an Auto Scaling group; the termination policy must be set directly on the Auto Scaling group or via the SageMaker endpoint configuration.
A company needs to update a model in production without any downtime. They currently have a single real-time endpoint serving traffic. Which approach allows them to deploy a new model version and switch traffic gradually while being able to roll back quickly?
This allows gradual traffic shift and the old variant can be used for rollback if needed.
Why this answer
SageMaker supports production variants with traffic splitting. By creating a new variant with the new model and shifting traffic gradually, the old variant remains available for rollback. Blue/green deployment with a new endpoint and endpoint configuration swap also allows quick rollback.
The key is to have both variants active during the transition.
An ML team wants to use Amazon SageMaker Ground Truth to create a labeled dataset for a multi-class image classification task. They have a large set of unlabeled images and want to minimize labeling costs while maintaining high accuracy. Which Ground Truth feature should they enable?
Active learning selects the most uncertain or informative samples for labeling, minimizing cost while maximizing model improvement.
Why this answer
Active learning in SageMaker Ground Truth automatically selects the most informative unlabeled images for human labeling, reducing the total number of labels needed while maintaining model accuracy. By iteratively training a model on a small labeled subset and then using that model to identify uncertain predictions, the system focuses labeling effort on the data that will most improve the model, directly minimizing labeling costs.
Exam trap
The trap here is that candidates may confuse 'annotation consolidation' (a post-labeling quality step) with a cost-reduction feature, or think that workforce management alone reduces costs, when in fact active learning is the specific feature designed to minimize the number of labels required.
How to eliminate wrong answers
Option B (Annotation consolidation) is wrong because it refers to combining multiple annotations for the same data point to produce a ground truth label, which does not reduce the number of labels needed. Option C (Data labeling workforce management) is wrong because it involves managing human labelers (e.g., public, private, or vendor workforces) but does not inherently reduce labeling volume or cost. Option D (Consolidated labeling) is not a distinct SageMaker Ground Truth feature; it is a generic term that might be confused with annotation consolidation, and it does not address cost minimization through selective labeling.
Refer to the exhibit. A SageMaker execution role has the IAM policy shown. The team attempts to run a training job that writes results to 's3://my-bucket/training/output/model.tar.gz'. What will happen?
The Deny does not affect this resource.
Why this answer
The training job will succeed because the Allow statement in the IAM policy explicitly grants s3:PutObject on the specific object 's3://my-bucket/training/output/model.tar.gz', and the Deny statement only blocks PutObject on objects with a 'training/' prefix in the key. Since the target object key is 'training/output/model.tar.gz', it does not start with 'training/' (the prefix is 'training/output/'), so the Deny does not apply. The Allow is therefore effective, and the model artifact is written successfully.
Exam trap
The trap here is that candidates assume the Deny statement blocks all PutObject actions to the 'training/' directory, but they overlook that the target object's key includes a subdirectory ('output/'), so the prefix 'training/' does not match the full key path 'training/output/model.tar.gz'.
How to eliminate wrong answers
Option A is wrong because the Deny statement does not block all PutObject actions; it only denies PutObject on objects whose key starts with 'training/', and the target object key 'training/output/model.tar.gz' does not match that prefix. Option C is wrong because the Deny statement does not override the Allow in this case; the Deny only applies when the condition (key starting with 'training/') is met, which it is not. Option D is wrong because there is no mention of encryption keys in the policy; the policy only controls access permissions, not encryption behavior.
Which SageMaker feature automatically generates model cards, feature importance, and bias reports without requiring manual coding?
Autopilot automatically creates model cards, feature importance, and bias reports.
Why this answer
SageMaker Clarify provides bias detection and feature importance, and it can generate reports. SageMaker Autopilot generates model cards and explanations. SageMaker Experiments tracks experiments.
SageMaker Model Monitor is for monitoring. Autopilot is the correct answer because it automates the entire pipeline including model cards and explanations.
A data scientist is using Amazon SageMaker Data Wrangler for feature engineering on a large dataset stored in S3. The dataset has a column 'ProductCategory' with 1000+ unique values. To reduce dimensionality, they want to group categories that appear less than 1% of the time into an 'Other' category. Which Data Wrangler transform should they use?
This built-in transform can group categories below a frequency threshold into an 'Other' value.
Why this answer
The 'Handle rare values' transform in SageMaker Data Wrangler is specifically designed to group infrequent category values into a single 'Other' bucket based on a frequency threshold (e.g., less than 1%). This directly addresses the need to reduce dimensionality by consolidating rare categories without requiring custom code or manual grouping.
Exam trap
The trap here is that candidates may confuse the 'Handle rare values' transform with the 'One-hot encode with threshold' transform, mistakenly thinking the threshold in one-hot encoding serves the same purpose as grouping rare categories, when in fact it limits the number of one-hot columns created, not the grouping of infrequent values.
How to eliminate wrong answers
Option A is wrong because 'Group similar categories' is a manual grouping transform that requires the user to explicitly define which categories to combine, not an automated threshold-based grouping of rare values. Option B is wrong because while a custom Python transform could technically achieve this, it is unnecessary and less efficient when a built-in, optimized transform ('Handle rare values') exists for this exact purpose. Option D is wrong because 'One-hot encode with threshold' applies to one-hot encoding (creating binary columns) and its threshold controls the maximum number of one-hot features, not the grouping of rare categories into an 'Other' bucket.
A data scientist deploys a model and wants to monitor the endpoint's invocation latency. They notice that the CloudWatch metric 'ModelLatency' is high, but 'OverheadLatency' is low. Which statement correctly interprets these metrics?
High ModelLatency indicates inference time is the issue.
Why this answer
The 'ModelLatency' metric measures the time taken by the SageMaker model container to process a single request, including inference and any preprocessing/postprocessing within the container. 'OverheadLatency' measures the time spent on SageMaker infrastructure (e.g., network I/O, request queuing, and response handling). When ModelLatency is high and OverheadLatency is low, the bottleneck is clearly the model inference time itself, not the infrastructure overhead. Therefore, optimizing the model (e.g., quantization, pruning) or upgrading to a faster instance type (e.g., GPU vs.
CPU) is the correct remediation.
Exam trap
The trap here is that candidates confuse 'ModelLatency' with overall endpoint latency and assume any high latency is due to infrastructure or scaling issues, when in fact the metric explicitly isolates the model's own inference time from overhead.
How to eliminate wrong answers
Option A is wrong because high ModelLatency with low OverheadLatency indicates the delay is inside the model container, not in SageMaker's infrastructure overhead; checking endpoint configuration would not address the model's own inference time. Option C is wrong because endpoint overload typically manifests as increased OverheadLatency (due to request queuing) or increased Invocations and 5xx errors, not as isolated high ModelLatency with low OverheadLatency. Option D is wrong because network latency is captured within OverheadLatency, not ModelLatency; moving the endpoint closer to clients would reduce OverheadLatency but would not affect the model's inference computation time.
A data engineer is preparing a large dataset of 10 TB for ML training on Amazon SageMaker. The data is stored in Amazon S3 as CSV files. To reduce training time and cost, the engineer wants to use a columnar format that is optimized for analytical queries. Which format should the engineer convert the data to?
Parquet is a columnar format that speeds up data access and reduces storage costs.
Why this answer
Parquet is a columnar storage format that is highly optimized for analytical queries and is natively supported by Amazon SageMaker for efficient data loading. By converting the 10 TB of CSV data to Parquet, the data engineer can reduce I/O and storage costs because columnar formats allow SageMaker to read only the columns needed for training, rather than scanning entire rows. This directly addresses the goal of reducing training time and cost for ML workloads.
Exam trap
AWS often tests the distinction between columnar formats (Parquet vs. ORC) by making both appear correct, but the trap here is that ORC is tightly coupled with Hive and less commonly used with SageMaker, while Parquet is the de facto standard for AWS-native ML and analytics services.
How to eliminate wrong answers
Option A (XML) is wrong because XML is a verbose, row-oriented text format that is not optimized for analytical queries; it would increase storage size and I/O overhead, making training slower and more expensive. Option C (ORC) is also a columnar format optimized for analytical queries, but it is primarily designed for and tightly integrated with the Apache Hive ecosystem, whereas Parquet is the more universally supported and recommended format for Amazon SageMaker and AWS analytics services. Option D (JSON Lines) is wrong because it is a row-oriented, text-based format that lacks the compression and columnar pruning benefits of Parquet, leading to higher storage costs and slower data access for ML training.
A company deploys a real-time inference endpoint and wants to be alerted if the number of 4XX errors exceeds 10 per minute over a 5-minute period. Which steps should they take?
Correct metric, threshold, and action.
Why this answer
A CloudWatch alarm on the `4XXError` metric with a threshold of 10 and an evaluation period of 5 minutes directly monitors the rate of HTTP 4XX errors from the SageMaker real-time inference endpoint. When the alarm state transitions to ALARM (i.e., the average 4XX errors per minute exceeds 10 over the 5-minute window), it triggers an SNS notification to alert the team. This is the standard approach for real-time metric-based alerting in AWS.
Exam trap
The trap here is that candidates confuse metric-based alerting (CloudWatch alarms on `4XXError`) with monitoring services (Model Monitor) or scaling mechanisms (auto-scaling), leading them to pick options that address different operational concerns.
How to eliminate wrong answers
Option B is wrong because the `Invocations` metric counts total requests, not 4XX errors, so it cannot detect error rate thresholds. Option C is wrong because endpoint auto-scaling with a target tracking policy adjusts capacity based on a target metric (e.g., Invocations per instance), not on error counts, and it does not generate alerts. Option D is wrong because SageMaker Model Monitor is designed for data quality, bias, and drift detection on captured payloads, not for real-time HTTP error rate monitoring; it cannot directly trigger alerts on 4XX error counts per minute.
A data scientist is using SageMaker to train a linear regression model. After training, they evaluate the model on the test set and get an R² of 0.95. However, when they deploy the model to a SageMaker endpoint and run predictions on new data, the predictions are far off. What is the most likely cause?
Correct: Data drift causes model to perform poorly on new data despite good test metrics.
Why this answer
A high R² of 0.95 on the test set indicates the model fits the test data well, but if the test set was drawn from the same distribution as the training data and does not reflect the real-world production data, the model will fail to generalize. In SageMaker, the endpoint serves predictions on live data that may have different statistical properties, leading to poor performance despite high test-set metrics. This is a classic case of dataset shift, not a model training or deployment configuration issue.
Exam trap
The trap here is that candidates confuse high test-set R² with model generalization, overlooking that the test set itself may be non-representative of production data, which is a core concept in the MLA-C01 exam under 'Model Evaluation and Validation'.
How to eliminate wrong answers
Option A is wrong because a different inference script would cause runtime errors or incorrect preprocessing, not systematically poor predictions on new data; SageMaker endpoints use the same inference code as the training container unless explicitly changed. Option C is wrong because using a wrong algorithm would typically result in poor training metrics (e.g., low R² on the test set), not a high R² of 0.95; the model converged well on the given data. Option D is wrong because overfitting would manifest as a large gap between training and test set performance (e.g., R² near 1.0 on training but much lower on test), but here the test R² is 0.95, suggesting the model generalizes to the test set; the issue is with production data differing from the test set.
A machine learning engineer notices that the latency of a SageMaker endpoint has increased over time. They need to identify which component (model inference vs. pre/post-processing) contributes most to the latency. Which CloudWatch metrics should they examine?
ModelLatency shows inference time inside the container; OverheadLatency shows SageMaker overhead. Comparing them pinpoints the latency source.
Why this answer
SageMaker endpoints emit CloudWatch metrics that break down total latency into model inference time (ModelLatency) and the time spent in pre/post-processing (OverheadLatency). By comparing these two metrics, the engineer can pinpoint whether the bottleneck is in the inference code or in the custom preprocessing/postprocessing logic. Option D directly provides both metrics needed for this root-cause analysis.
Exam trap
The trap here is that candidates confuse the total Latency metric with a breakdown metric, assuming it alone can identify the bottleneck, when in fact only the pair of ModelLatency and OverheadLatency provides the necessary decomposition.
How to eliminate wrong answers
Option A is wrong because Latency is the total end-to-end response time, and ModelLatency alone only covers inference; together they do not isolate the pre/post-processing component. Option B is wrong because Invocations and 4XXError track request count and client-side errors, not latency breakdown. Option C is wrong because 5XXError indicates server-side failures and MemoryUtilization shows resource pressure, but neither metric decomposes latency into inference vs. overhead.
Refer to the exhibit. A team receives an error when running a SageMaker Model Monitor schedule for data quality. What should they do to resolve this issue?
A baseline must be generated from training data to compare inference data against.
Why this answer
The error occurs because SageMaker Model Monitor requires a baseline to compare against live data. Without a baseline job created from the training dataset, the monitoring schedule fails. Option D resolves this by generating the necessary statistics and constraints that define expected data quality.
Exam trap
The trap here is that candidates often assume the error is a permissions or configuration issue (S3 access or data capture), but the root cause is the mandatory prerequisite of a baseline job before a monitoring schedule can run.
How to eliminate wrong answers
Option A is wrong because the IAM role likely already has S3 access if the endpoint and model artifacts are deployed; the error is not a permissions issue. Option B is wrong because restarting the monitoring schedule does not address the missing baseline; the schedule will fail again. Option C is wrong because data capture must already be enabled on the endpoint for Model Monitor to collect inference data; the error indicates a missing baseline, not a missing data capture configuration.
A data scientist is preparing a dataset for a linear regression model. The dataset has a few missing values in a numerical feature with a normal distribution and no outliers. Which imputation method is most appropriate?
Mean is appropriate for normally distributed numerical data without outliers.
Why this answer
For a numerical feature with a normal distribution and no outliers, the mean is the most appropriate imputation method because it preserves the central tendency of the data without introducing bias. In linear regression, mean imputation maintains the expected value of the feature, which is critical for unbiased coefficient estimates when data are missing completely at random (MCAR).
Exam trap
The trap here is that candidates often confuse the median with the mean for normal distributions, but the median is actually less efficient and can lead to biased variance estimates, while the mean is the maximum likelihood estimator for normally distributed data with no outliers.
How to eliminate wrong answers
Option A is wrong because the mode is intended for categorical data, not for a normally distributed numerical feature, and it would distort the distribution by replacing missing values with the most frequent value rather than the central tendency. Option C is wrong because the median is robust to outliers, but since the dataset has no outliers and a normal distribution, the median is less efficient than the mean and would slightly underestimate the variance, reducing statistical power. Option D is wrong because dropping rows with missing values reduces sample size and can introduce bias if the missingness is not completely random, whereas imputation with the mean is a standard technique for MCAR data in linear regression.
A company wants to deploy a machine learning model that makes real-time predictions for a mobile app. The model is a deep neural network with a large model size (500 MB). Which SageMaker endpoint configuration is most cost-effective while meeting low-latency requirements?
Ensures low latency and is cost-effective for a single model with sustained traffic.
Why this answer
A real-time endpoint with a single instance provides the lowest latency for a 500 MB deep neural network model, as it keeps the model loaded in memory and ready for inference without cold starts or multi-model overhead. This configuration is also cost-effective for consistent traffic patterns, as you pay for the instance uptime rather than per-invocation or for multiple model loads.
Exam trap
The MLA-C01 exam often tests the misconception that multi-model endpoints are always more cost-effective for large models, but the trap here is that multi-model endpoints introduce significant latency from disk I/O for models over 100 MB, making them unsuitable for real-time inference despite lower instance costs.
How to eliminate wrong answers
Option A is wrong because multi-model endpoints are designed to host multiple smaller models on a single instance, but they incur latency overhead from loading/unloading models from disk, which is unsuitable for a 500 MB model requiring real-time predictions. Option B is wrong because serverless inference has a maximum payload size of 6 MB and experiences cold starts, making it incompatible with a 500 MB model and low-latency requirements. Option D is wrong because batch transform is an asynchronous, offline inference method that does not provide real-time predictions, and it is designed for large-scale batch processing, not low-latency mobile app requests.
A machine learning engineer is deploying a PyTorch model for real-time inference on SageMaker. The model requires GPU for low-latency predictions. The deployment fails with the error: 'The primary container does not support the requested instance type.' The instance type is ml.p3.2xlarge. Which action should the engineer take to resolve the issue?
Older PyTorch versions may not support GPU; using a supported version resolves the error.
Why this answer
The error 'The primary container does not support the requested instance type' typically occurs when the specified PyTorch framework version in the SageMaker estimator does not include GPU support for the chosen instance type (ml.p3.2xlarge). SageMaker's prebuilt PyTorch containers are version-specific and only certain versions are compiled with CUDA and GPU libraries; using a version that lacks GPU support causes the container to reject GPU instance types. Verifying and selecting a PyTorch version that explicitly supports GPU instances resolves the mismatch.
Exam trap
The trap here is that candidates often assume the error is due to resource limits (quota) or hardware incompatibility (Neo), rather than recognizing it as a framework version and container image mismatch specific to GPU support.
How to eliminate wrong answers
Option A is wrong because SageMaker Neo compiles models for edge devices or optimized inference on specific hardware, but it does not fix a container-instance type compatibility error; the error occurs before model compilation. Option B is wrong because a service quota increase addresses insufficient capacity or account limits for the instance type, not a container-level compatibility error; the error indicates the container rejects the instance type, not that the instance is unavailable. Option D is wrong because creating a custom inference container is unnecessary when the issue is simply a version mismatch in the prebuilt container; the error can be resolved by selecting a supported PyTorch version without custom container overhead.
A company wants to build a customer service chatbot that answers questions about their internal policy documents. The documents are updated monthly, and the team cannot afford to retrain a model each time. Which approach is MOST appropriate?
RAG retrieves relevant document chunks at query time, ensuring the chatbot always answers from the latest uploaded documents without any model retraining.
Why this answer
RAG (Retrieval-Augmented Generation) allows the LLM to retrieve relevant document sections at inference time, so knowledge stays current without retraining. The other options either require expensive retraining for each update or lack document grounding.
A data scientist is training a regression model in Amazon SageMaker. The dataset contains missing values in several features. The scientist wants to handle missing values as part of the training pipeline to ensure consistency between training and inference. Which approach should the scientist use?
Consistent preprocessing in pipeline.
Why this answer
It uses a scikit-learn container within SageMaker to create a preprocessing step that imputes missing values, then includes that step in the inference pipeline. This ensures the same imputation logic (e.g., mean, median, or custom strategy) is applied consistently during both training and inference, preventing data drift and maintaining reproducibility. SageMaker Pipelines or the built-in scikit-learn container allow the preprocessing to be serialized as part of the model artifact, so inference requests automatically undergo the same transformation.
Exam trap
The trap here is that candidates often assume SageMaker Autopilot (Option B) is the correct choice because it automates preprocessing, but they miss that the question specifically requires a custom, reproducible pipeline that ensures consistency between training and inference, which Autopilot does not expose for custom control.
How to eliminate wrong answers
Option A is wrong because handling missing values in a separate Jupyter notebook and saving the cleaned data breaks the training-inference consistency; the imputation logic is not captured in a reusable pipeline, leading to potential mismatch when new data arrives during inference. Option B is wrong because SageMaker Autopilot is an automated machine learning service that handles missing values internally during model selection, but it does not allow the data scientist to control the imputation method or integrate a custom preprocessing step into a production inference pipeline. Option C is wrong because dropping all rows with missing values can discard valuable data, reduce model performance, and is not feasible when missing values appear in inference-time data, as the pipeline would have no strategy to handle them.
A data scientist is training a model using SageMaker and wants to use spot instances to reduce costs. The training job is checkpointed every 5 minutes. However, the job gets interrupted frequently and never completes. What is the MOST likely cause?
If interruptions occur more often than checkpoints, progress is lost and job may never complete.
Why this answer
Spot instances can be reclaimed with little notice. If the job checkpoint interval is longer than the average interruption notice, progress may be lost. Using a smaller instance type reduces cost but not interruption frequency.
Incorrect checkpoint path causes save failures. Too few max retries causes job to stop after few interruptions.
A machine learning engineer is deploying a TensorFlow model for real-time inference. The model has high latency on CPU. Which TWO actions can reduce inference latency? (Choose two.)
Elastic Inference adds GPU acceleration, reducing latency.
Why this answer
Compiling with SageMaker Neo optimizes the model for the target hardware. Attaching Elastic Inference provides GPU acceleration without moving to a full GPU instance.
A data scientist is using Amazon SageMaker Ground Truth to create a labeled dataset for an object detection model. The dataset contains 1 million images, and the team wants to reduce labeling cost by labeling only the most informative samples. Which feature of Ground Truth should they use?
Active learning selects samples where the model is uncertain, maximizing labeling efficiency.
Why this answer
Ground Truth offers active learning, which automatically selects the most informative samples to label, reducing cost. Option A is correct. Options B, C, and D do not provide automatic sample selection for labeling.
A data scientist is running a SageMaker training job with a custom PyTorch image. The training script loads a large dataset into memory, and the job fails with an out-of-memory error after a few minutes. The instance type is ml.m5.xlarge (16 GB RAM). What should the data scientist do to resolve this issue without changing the instance type?
Smaller batch sizes reduce memory consumption per step, helping to fit within the available RAM.
Why this answer
Reducing the batch size decreases the amount of data loaded into memory at once, directly addressing the out-of-memory error without changing the instance type. Since the training script loads a large dataset into memory and fails after a few minutes, a smaller batch size reduces peak memory consumption per iteration, allowing the job to fit within the 16 GB RAM of ml.m5.xlarge.
Exam trap
The trap here is that candidates confuse streaming data (Pipe mode) with reducing in-memory data loading, not realizing that the script's explicit load into memory bypasses any streaming benefit.
How to eliminate wrong answers
Option A is wrong because SageMaker Managed Spot Training provides cost savings via discounted spare EC2 capacity but does not free or reduce memory usage; it can even cause interruptions that require checkpointing. Option B is wrong because increasing the number of workers with multiprocessing increases memory overhead due to data duplication across processes, exacerbating the out-of-memory issue. Option D is wrong because SageMaker Pipe mode streams data from S3 directly to the training algorithm without writing to disk, but the training script still loads the dataset into memory, so the memory footprint remains unchanged.
A company has deployed a SageMaker real-time endpoint for a model that predicts customer churn. The endpoint uses a single ml.m5.large instance. After deployment, the team notices that during peak hours, the endpoint returns 5xx errors for about 20% of requests. The endpoint has not been configured with any scaling policy. The team needs to resolve this issue with minimal cost increase. Which solution should the team implement?
Auto Scaling adds instances only when needed, minimizing cost while handling peak load.
Why this answer
Enabling Auto Scaling with a target tracking policy based on the average InvocationsPerInstance metric dynamically adjusts the number of instances in response to traffic spikes, preventing 5xx errors during peak hours without over-provisioning. This approach minimizes cost by scaling only when needed, unlike manual instance upgrades or batch transforms that either increase baseline cost or introduce latency.
Exam trap
The trap here is that candidates often confuse 'scaling up' (increasing instance size) with 'scaling out' (adding more instances), and overlook that Auto Scaling with a target tracking policy is the most cost-effective way to handle variable traffic, as it matches capacity to demand in real time.
How to eliminate wrong answers
Option A is wrong because deploying to a multi-model endpoint reduces resource utilization by sharing a single container across multiple models, but it does not address the root cause of insufficient capacity for a single model under peak load; it may even exacerbate contention. Option C is wrong because increasing the instance type to ml.m5.xlarge provides more compute per instance but incurs a fixed higher cost regardless of traffic, failing the 'minimal cost increase' requirement and not dynamically adapting to variable load. Option D is wrong because SageMaker batch transform is designed for asynchronous, offline inference on large datasets, not for real-time requests; it would introduce unacceptable latency and cannot serve interactive predictions, thus not resolving the immediate 5xx errors during peak hours.
An MLOps engineer is designing a CI/CD pipeline for deploying machine learning models to a production SageMaker endpoint. The pipeline should include automated testing, approval gates, and rollback capability. Which THREE components should be included in the pipeline? (Select THREE.)
Infrastructure as code allows precise rollback by redeploying a previous CloudFormation stack.
Why this answer
Using a CloudFormation template to deploy the SageMaker endpoint infrastructure enables rollback via stack update. If a deployment fails, CloudFormation can automatically roll back the stack to the previous known good state, ensuring infrastructure consistency and reducing downtime.
Exam trap
The trap here is that candidates confuse model registry steps (Option A) or training monitoring tools (Option E) with deployment pipeline components, but the question specifically asks for components that enable automated testing, approval gates, and rollback capability in the CI/CD pipeline for deploying to a production SageMaker endpoint.
A machine learning engineer is deploying a pre-trained NLP model on Amazon SageMaker for real-time inference. The model expects input sequences of variable length, and performance is critical. The engineer wants to minimize latency while handling the variable-length inputs efficiently. Which approach should the engineer choose?
Dynamic batching reduces padding and latency.
Why this answer
Dynamic batching with a custom inference script that groups requests by sequence length minimizes padding overhead and maximizes hardware utilization. By batching similar-length sequences together, the model avoids excessive padding to the maximum length in the batch, which reduces wasted computation and latency. This approach is particularly effective for variable-length NLP inputs on SageMaker, where the inference container can be customized to implement the grouping logic.
Exam trap
AWS often tests the misconception that padding to the maximum length is always necessary or efficient, but the trap here is that dynamic batching with length-based grouping is a more sophisticated technique that balances batching efficiency with minimal padding overhead.
How to eliminate wrong answers
Option A is wrong because pruning and quantization reduce model size and can improve latency, but they do not address the core issue of efficiently handling variable-length input sequences; they are orthogonal optimizations. Option B is wrong because padding all sequences to the maximum length in the batch introduces significant wasted computation and memory, especially when sequence lengths vary widely, leading to higher latency. Option D is wrong because processing each request individually eliminates batching benefits, resulting in lower throughput and higher per-request latency due to underutilized hardware accelerators.
A machine learning engineer is using SageMaker Autopilot for AutoML. Which TWO outputs does Autopilot produce?
An ML team at a financial services company has developed a fraud detection model using Amazon SageMaker. The model is currently deployed to a production endpoint with a single variant using the previous model version. The team wants to deploy a new model version with a canary deployment where 10% of traffic goes to the new version and 90% remains on the old version for 30 minutes before shifting all traffic to the new version if no issues are detected. Which step is essential to achieve this safe rollout?
This configuration uses SageMaker's blue/green deployment with canary traffic shifting, which is the correct approach.
Why this answer
It uses the SageMaker endpoint update with a new production variant and sets 'InitialVariantWeight' to 10 for the new model and 90 for the old model, which routes 10% of traffic to the new version. Additionally, specifying a 'BlueGreenUpdatePolicy' with a 'TrafficRoutingConfiguration' for canary enables the automatic shift of all traffic to the new variant after 30 minutes if no issues are detected, achieving the desired safe rollout.
Exam trap
The trap here is that candidates may think canary deployments require manual traffic splitting or separate endpoints, but SageMaker's native 'BlueGreenUpdatePolicy' with 'TrafficRoutingConfiguration' automates the entire process, including traffic shifting and rollback, without needing custom code or DNS manipulation.
How to eliminate wrong answers
Option A is wrong because the 'Deploy' method on a model object does not have a 'mode' parameter set to 'canary'; SageMaker's built-in XGBoost container does not support canary deployment via a 'mode' parameter, and canary deployments are managed at the endpoint configuration level, not within the algorithm container. Option C is wrong because hosting the endpoint on at least two instances is not a requirement for canary deployments, and manually adjusting the endpoint's DNS to split traffic is not a supported or reliable method in SageMaker; traffic splitting is done via variant weights in the endpoint configuration. Option D is wrong because deploying the new model as a separate endpoint and using a SageMaker predictor to randomly route 10% of inference requests is not a built-in feature of SageMaker; it would require custom code and does not provide the automatic traffic shifting after 30 minutes, nor does it integrate with SageMaker's native deployment monitoring and rollback capabilities.
A data scientist is working with a dataset containing customer demographics and purchase history. The dataset includes categorical variables with high cardinality (e.g., ZIP code, product ID). The data scientist wants to perform feature engineering to improve model performance. Which THREE feature engineering techniques should the data scientist consider? (Choose three.)
PCA can reduce noise and multicollinearity.
Why this answer
Principal Component Analysis (PCA) is a dimensionality reduction technique that transforms correlated numerical features into a smaller set of uncorrelated principal components, capturing the maximum variance in the data. This is correct because the dataset includes numerical features (e.g., purchase amounts, age) where PCA can reduce noise and multicollinearity, improving model performance without losing critical information.
Exam trap
AWS often tests the distinction between techniques that are universally applicable (like PCA for numerical features) versus those that are specifically designed to handle high-cardinality categorical variables (like target encoding and frequency encoding), tempting candidates to choose one-hot encoding without considering its impracticality for high cardinality.
A company uses SageMaker Pipelines to automate their ML workflow. They notice that the pipeline reruns all steps even when the input data has not changed. Which feature should they enable to avoid unnecessary recomputation?
Caching stores step outputs and reuses them when inputs are identical, preventing unnecessary reruns.
Why this answer
Pipeline caching in SageMaker Pipelines automatically reuses the output of a step if its inputs (including parameters, data, and code) have not changed since the last successful execution. This avoids recomputation by comparing a hash of the step's dependencies against previous runs, making it the correct feature to prevent unnecessary reruns when input data remains identical.
Exam trap
The trap here is that candidates confuse caching with conditional branching or parallel execution, assuming that skipping steps via conditions or running steps in parallel will avoid recomputation, when in fact only caching directly reuses prior outputs based on input immutability.
How to eliminate wrong answers
Option B is wrong because a Lambda step is used for custom processing or integration (e.g., invoking external APIs), not for detecting input changes or caching step outputs; it would add complexity without solving the core caching requirement. Option C is wrong because a Conditional step evaluates a condition to branch the pipeline (e.g., skip a step based on a metric), but it does not automatically detect unchanged inputs or cache results; it requires manual logic and still incurs overhead for the condition check. Option D is wrong because setting the pipeline execution mode to 'Parallel' controls whether steps run sequentially or concurrently, but it does not prevent recomputation of steps whose inputs have not changed; it only affects execution order, not caching.
A company uses SageMaker endpoints for real-time inference. They want to automatically scale the number of instances based on the number of outstanding requests. Which auto-scaling policy type should they choose?
Target tracking automatically adjusts capacity to keep the specified metric at the target value.
Why this answer
Target tracking scaling adjusts the instance count to maintain a target metric value (e.g., average invocation count per instance). Step scaling uses predefined scaling adjustments based on alarm breaches but does not directly track a target. Simple scaling is not recommended for production.
Scheduled scaling is for predictable patterns, not dynamic.
A data science team is deploying a PyTorch model for real-time inference with sub-second latency requirements. They need to minimize cost while handling variable traffic. Which TWO approaches should they consider? (Choose TWO.)
Neo optimizes the model for the target hardware, reducing inference latency and often allowing a smaller instance type.
Why this answer
Serverless inference auto-scales to zero when not in use and charges per request, minimizing cost for variable traffic. SageMaker Neo compiles the model for optimal hardware performance, achieving low latency. Multi-model endpoints (MME) are for hosting multiple models, not single-model optimization.
Elastic Inference adds GPU acceleration at lower cost than a full GPU instance, but with Neo compilation the team may not need it. Batch transform is for offline, not real-time.
Practice MLA-C01 by domain
Target a specific domain to shore up weak areas.