MLS-C01 Practice Test 3 — 25 Questions

Question 1

A company is using Amazon SageMaker to train a deep learning model. The training job is failing with an error 'CUDA out of memory'. The training instance is an ml.p3.2xlarge with 16 GB GPU memory. The model architecture and batch size are appropriate for this instance size. What is the most likely cause of this error?

Accepted Answer

Enable automatic mixed precision (AMP) training to reduce memory usage.. Option C is correct because enabling automatic mixed precision (AMP) training reduces GPU memory usage by storing tensors in half-precision (FP16) where possible, while keeping critical operations in full precision (FP32). This directly addresses the 'CUDA out of memory' error on an ml.p3.2xlarge instance (16 GB GPU memory) without changing the model architecture or batch size, which are already appropriate.

Answer

Reduce the number of epochs.

Answer

Increase the number of GPUs by using a distributed training instance type.

Answer

Use a smaller instance type to force lower memory usage.

Question 2

A company is using Amazon SageMaker to train a deep learning model on a large dataset stored in S3. The training job is failing with an OutOfMemory error. The data scientist wants to minimize cost while resolving the issue. Which action should the data scientist take?

Accepted Answer

Change the input mode from 'File' to 'Pipe'.. The OutOfMemory error occurs because the 'File' input mode downloads the entire training dataset to the instance's local storage before training begins, consuming significant memory. Switching to 'Pipe' mode streams data directly from S3 to the training algorithm, reducing memory footprint and avoiding the need for larger instances. This minimizes cost by using the existing instance type while resolving the memory issue.

Answer

Increase the instance type to one with more memory.

Answer

Use the 'auto' setting for the input mode.

Answer

Reduce the batch size hyperparameter.

Question 3

A company is using Amazon SageMaker to train a XGBoost model on a large dataset. The training job is taking a long time. The data scientist wants to reduce training time without sacrificing model accuracy. The dataset is 100 GB in CSV format stored in S3. What is the most effective approach?

Accepted Answer

Use Pipe mode to stream data from S3 instead of downloading it first.. Option B is correct because SageMaker's Pipe mode streams data directly from S3 to the training algorithm without writing it to disk, eliminating the I/O bottleneck of downloading the full 100 GB dataset. This reduces training time significantly by overlapping data loading with computation, while preserving model accuracy since the entire dataset is still used.

Answer

Reduce the number of instances to avoid communication overhead.

Answer

Use random sampling to reduce the dataset size to 10 GB.

Answer

Use SageMaker Managed Spot Training to reduce cost, but training time may increase due to interruptions.

Question 4

A data scientist is deploying a model using Amazon SageMaker. The model endpoint needs to handle real-time inference requests with low latency. The model is a large ensemble of 10 deep learning models, each approximately 500 MB. What is the most cost-effective deployment strategy that meets the low-latency requirement?

Accepted Answer

Use a SageMaker multi-model endpoint to host all models on one or more instances.. A SageMaker multi-model endpoint (MME) allows hosting multiple models on a single or few instances, dynamically loading them from Amazon S3 into memory as needed. This is the most cost-effective option for a large ensemble of 500 MB models because it avoids the expense of separate endpoints or multiple instances per model, while still supporting low-latency real-time inference by keeping frequently used models cached.

Answer

Deploy each model to a separate endpoint and use a load balancer.

Answer

Use a single endpoint with multiple instances behind it.

Answer

Use a SageMaker batch transform job to process inference requests in batches.

Question 5

An IAM policy is attached to a SageMaker execution role. A data scientist tries to create a training job using a custom algorithm stored in an ECR repository. The training job fails with an 'AccessDenied' error when pulling the Docker image from ECR. What is the missing permission?

Accepted Answer

ecr:GetDownloadUrlForLayer and ecr:BatchGetImage on the ECR repository. When SageMaker pulls a custom Docker image from ECR during training job creation, the execution role needs permissions to download the image layers. The required actions are ecr:GetDownloadUrlForLayer (to generate pre-signed URLs for each layer) and ecr:BatchGetImage (to retrieve image metadata and layer manifests). Without these, the 'AccessDenied' error occurs because SageMaker cannot authenticate or fetch the container image from the ECR repository.

Answer

ecr:PutImage on the ECR repository

Answer

s3:GetObject on the ECR repository

Answer

sagemaker:CreateTrainingJob on the ECR resource

Question 6

A company is using Amazon SageMaker to deploy a model for real-time inference. The model endpoint is behind an Application Load Balancer (ALB) for A/B testing. The data scientist notices that the endpoint is returning HTTP 503 errors intermittently. The CloudWatch metrics show that the endpoint's Invocations metric is within limits, but the ModelLatency metric has high variance. What is the most likely cause?

Accepted Answer

The endpoint instance type does not have enough memory for the model.. Option C is correct because high variance in ModelLatency combined with intermittent 503 errors strongly indicates that the model container is running out of memory under load. When memory is insufficient, the inference process may be killed by the kernel (OOM killer) or the container may be throttled, causing sporadic failures that manifest as 503s even though the Invocations metric (request count) appears within limits. The latency spikes occur because the container struggles to allocate memory for each request, leading to timeouts or crashes.

Answer

The model container is using a custom inference code that has a bug.

Answer

The ALB health check is misconfigured and marking instances unhealthy.

Answer

The endpoint is configured with too few instances; increase the instance count.

Question 7

A DevOps engineer created a SageMaker notebook instance using the Terraform configuration shown. The notebook instance is in a VPC with a public subnet. However, the notebook instance cannot access the internet. What is the most likely cause?

Accepted Answer

The subnet does not have a route to an internet gateway.. Option C is correct because a SageMaker notebook instance in a VPC with a public subnet requires a route to an internet gateway (IGW) in the subnet's route table to access the internet. Without that route, traffic from the notebook cannot reach the internet, even if `direct_internet_access` is enabled. The Terraform configuration likely omitted the route to the IGW, causing the connectivity failure.

Answer

The role_arn is incorrect or missing permissions.

Answer

The instance type ml.t2.medium does not support internet access.

Answer

The direct_internet_access parameter is set to 'Enabled' but should be 'Disabled'.

Question 8

A data scientist is trying to create a training job named 'test-model' using an IAM role with the attached policy. The creation fails with an AccessDenied error. What is the most likely cause?

Accepted Answer

The Deny statement uses a wildcard '*' in the condition value, which is not supported for StringNotEquals.. The Deny statement uses 'StringNotEquals' with a wildcard '*' in the condition value, which is not supported for the 'StringNotEquals' condition operator in IAM policies. The 'StringNotEquals' operator requires exact string matching and does not support wildcards; using '*' will cause the condition to never match, effectively making the Deny statement non-functional or causing unexpected behavior. This mismatch leads to an AccessDenied error because the policy evaluation fails to properly deny or allow the action.

Answer

The Resource is set to '*' and should be specific.

Answer

The Deny statement uses 'StringNotEquals' which should be 'StringEquals'.

Answer

The IAM role does not have permission to assume the SageMaker execution role.

Question 9

A company is using AWS Glue to run ETL jobs that transform data for machine learning. The jobs are failing with 'Out of Memory' errors. The data size is growing, and the company needs a cost-effective solution. Which approach should be taken?

Accepted Answer

Increase the number of workers in the job configuration.. Increasing the number of workers in the AWS Glue job configuration distributes the data processing load across more Spark executors, directly addressing the 'Out of Memory' error by providing more aggregate memory without changing the worker type. This is a cost-effective approach because it scales horizontally, often at a lower cost than moving to a larger worker type, and it leverages the existing Glue infrastructure without migrating to EMR.

Answer

Switch to Spark on Amazon EMR.

Answer

Optimize the job by filtering data earlier.

Answer

Use a larger worker type like G.2X.

Question 10

A company is using Amazon SageMaker to build a machine learning pipeline. The pipeline includes data preprocessing, training, and evaluation steps. The company wants to ensure that the pipeline is reproducible and that artifacts are versioned. Which TWO actions should be taken? (Choose TWO.)

Accepted Answer

Use SageMaker Pipelines to create the pipeline and enable versioning on the pipeline artifacts.. SageMaker Pipelines provides a native way to define, orchestrate, and version machine learning pipelines. By enabling versioning on pipeline artifacts (e.g., via the `Pipeline` object's `version` parameter or by using SageMaker Model Registry), each pipeline run is tracked with a unique version, ensuring reproducibility. This directly addresses the requirement for reproducible pipelines and versioned artifacts.

Answer

Use a naming convention for training jobs that includes the date.

Answer

Use AWS CodePipeline to trigger the pipeline on code changes.

Answer

Store the training dataset in a versioned S3 bucket.

Question 11

A company runs a machine learning pipeline on Amazon SageMaker. The pipeline consists of three steps: data preprocessing (using a custom container), training (using a built-in algorithm), and model evaluation (using a custom container). The pipeline is orchestrated using AWS Step Functions. Recently, the pipeline has been failing intermittently at the model evaluation step with a 'TimeoutError'. The evaluation step runs a Python script that loads the trained model and a test dataset from S3, computes metrics, and writes results back to S3. The step is configured with a timeout of 600 seconds. The test dataset size has grown over time. The data science team suspects that the timeout is due to the increased data size. They want a solution that minimizes changes to the existing infrastructure and avoids increasing the timeout arbitrarily. Which approach should the team take?

Accepted Answer

Modify the evaluation script to process the test dataset in parallel batches, and use multiprocessing to distribute the workload within the same container.. Option C is correct because it addresses the root cause—the evaluation script's inability to process the growing dataset within the 600-second timeout—by parallelizing the workload within the same container. This approach minimizes infrastructure changes (no instance type or timeout increase) and leverages Python's multiprocessing to reduce wall-clock time, directly tackling the 'TimeoutError' without arbitrary timeout extensions.

Answer

Increase the timeout to 1200 seconds and use a larger instance type for the evaluation step.

Answer

Increase the timeout to 1800 seconds to accommodate the larger dataset.

Answer

Switch the evaluation step to use the 'ml.m5.4xlarge' instance type for more memory and compute.

Question 12

A data scientist is performing EDA on a high-dimensional dataset with 500 features. They want to visualize the data in 2D to check for clusters. They first apply PCA and get a 2D projection that shows no clear structure. They suspect that the data lies on a non-linear manifold. Which of the following techniques should they try next?

Accepted Answer

Use t-distributed Stochastic Neighbor Embedding (t-SNE).. Option D is correct because t-SNE is a non-linear dimensionality reduction technique specifically designed for visualization. Option A (ICA) is a linear method for separating independent components, not for capturing non-linear manifolds. Option B (LDA) is a supervised linear method that maximizes class separation, unsuitable for unsupervised non-linear structure. Option C (PCA with more components) remains linear, so adding components does not help with non-linear manifolds.

Answer

Use Independent Component Analysis (ICA).

Answer

Use Linear Discriminant Analysis (LDA).

Answer

Apply PCA again with more components.

Question 13

A machine learning team is building a multi-class image classifier using a pre-trained ResNet-50 model in Amazon SageMaker. The dataset has 10 classes but is highly imbalanced, with one class representing 80% of the samples. The team wants to improve model performance on the minority classes. Which TWO of the following approaches are most likely to help? (Select TWO.)

Accepted Answer

Oversample the minority classes in the training data.. Oversampling the minority classes (Option A) directly addresses class imbalance by replicating samples from underrepresented classes, giving the model more exposure to them during training. This is a standard data-level technique that helps the ResNet-50 model learn discriminative features for minority classes without altering the loss function or model architecture.

Answer

Reduce the batch size to increase the frequency of weight updates.

Answer

Increase the number of layers in the model.

Answer

Switch to a focal loss function.

Question 14

A company is using SageMaker Ground Truth to label images for a computer vision model. After launching the labeling job, they notice that the labeling throughput is lower than expected. What should they do to increase throughput?

Accepted Answer

Use a private workforce with more workers.. SageMaker Ground Truth labeling throughput is primarily limited by the number of workers available to process tasks. Using a private workforce allows you to directly control and scale the number of workers, which increases parallelism and overall throughput. Public or vendor workforces have fixed capacity and may not scale as quickly, so adding more private workers is the most effective way to boost throughput.

Answer

Change the labeling task to use a single annotator per image.

Answer

Reduce the number of workers assigned to each task.

Answer

Increase the time allowed for each labeling task.

Question 15

An ML engineer is troubleshooting why an automated CI/CD pipeline cannot deploy an updated model to an existing SageMaker endpoint. The pipeline uses the IAM role that has the attached policy shown in the exhibit. What is the MOST likely cause of the failure?

Accepted Answer

The pipeline tries to update an existing endpoint, but the sagemaker:UpdateEndpoint action is not allowed.. The pipeline is attempting to deploy an updated model to an existing SageMaker endpoint, which requires the sagemaker:UpdateEndpoint action. The IAM policy shown in the exhibit (not provided here but implied) does not include this action, so the API call fails with an access denied error. Without explicit permission to update the endpoint, the CI/CD pipeline cannot modify the deployed configuration.

Answer

The pipeline tries to create a new endpoint, but the sagemaker:CreateEndpoint action is denied.

Answer

The pipeline tries to delete the old endpoint, but the sagemaker:DeleteEndpoint action is denied by a Deny statement.

Answer

The pipeline attempts to invoke the endpoint, but the sagemaker:InvokeEndpoint action is denied.

Question 16

A data science team is training a large deep learning model using Amazon SageMaker. The training job is taking a long time because the model has many layers and the dataset is large. The team wants to reduce training time by distributing the training across multiple GPUs on a single instance, as well as across multiple instances. Which TWO actions should the team take? (Choose two.)

Accepted Answer

Configure the training job to use SageMaker's model parallelism (SMP) library for pipeline or tensor parallelism.. Option B is correct because the SageMaker model parallelism (SMP) library is specifically designed to split large deep learning models across multiple GPUs using pipeline or tensor parallelism. This allows the team to train models that are too large to fit on a single GPU and to reduce training time by parallelizing computation across devices within and across instances.

Answer

Use SageMaker's distributed data parallelism (SMDDP) library to shard the model across GPUs.

Answer

Use SageMaker's managed training with a single instance containing multiple GPUs and enable data parallelism.

Answer

Set the instance type to a single GPU instance and rely on automatic model parallelism.

Question 17

A company wants to deploy a machine learning model that performs real-time inference with sub-second latency. The model is a deep neural network with 500 MB of weights. The inference endpoint must scale to zero when not in use to minimize cost. Which AWS service should the company use?

Accepted Answer

Use Amazon SageMaker Serverless Inference to host the model.. Amazon SageMaker Serverless Inference is designed for workloads with intermittent traffic patterns, automatically scaling to zero when idle and scaling up for real-time requests. It supports models up to 1 GB in size and provides sub-second latency for inference, making it ideal for this 500 MB deep neural network. This service eliminates the need to manage underlying infrastructure while meeting the latency and cost requirements.

Answer

Deploy the model as an AWS Lambda function with provisioned concurrency.

Answer

Host the model on Amazon ECS with Fargate and use a target tracking scaling policy.

Answer

Create an Amazon SageMaker real-time endpoint with automatic scaling policies.

Question 18

A data scientist wants to deploy a PyTorch model for real-time inference with low latency. Which AWS service should they use?

Accepted Answer

Amazon SageMaker real-time endpoint. Amazon SageMaker real-time endpoints are specifically designed for low-latency inference on deployed models, including PyTorch models. They provide persistent HTTPS endpoints that autoscale and support custom containers, making them ideal for real-time prediction workloads.

Answer

Amazon Elastic Container Service (ECS)

Answer

Amazon SageMaker batch transform

Answer

AWS Lambda

Question 19

A company wants to use Amazon SageMaker to host a model that was trained using a custom algorithm. The model artifact is stored in Amazon S3. The company wants to ensure that the endpoint can automatically scale based on the number of incoming requests. Which configuration should the company use?

Accepted Answer

Create a SageMaker real-time endpoint and configure automatic scaling using a target tracking policy.. Option B is correct because a SageMaker real-time endpoint with automatic scaling using a target tracking policy allows the endpoint to dynamically adjust the number of instances based on the incoming request load. This configuration is ideal for hosting a custom algorithm model artifact stored in S3, as it provides low-latency inference and can scale out or in based on a target metric like average CPU utilization or request count per instance.

Answer

Create a SageMaker multi-model endpoint with automatic scaling.

Answer

Use SageMaker Serverless Inference which scales automatically.

Answer

Use SageMaker Batch Transform with a scheduled job.

Question 20

A company's ML pipeline uses AWS Step Functions to orchestrate data preprocessing, training, and evaluation. The training step occasionally fails due to a transient error. What is the most robust way to handle this without manual intervention?

Accepted Answer

Implement a retry policy with exponential backoff on the training step in the state machine. Option A is correct because AWS Step Functions natively supports retry policies with exponential backoff, which automatically retries failed tasks after a delay that increases progressively. This handles transient errors (e.g., resource contention, network glitches) without manual intervention, making the pipeline robust and self-healing.

Answer

Configure a CloudWatch alarm to notify the team when the step fails

Answer

Use a parallel state to run multiple training instances simultaneously

Answer

Use a custom Lambda function to catch the error and restart the training step