MLS-C01 Machine Learning Implementation and Operations Practice Test 27 — 15 Questions

Question 1

A company uses Amazon SageMaker to train a deep learning model for image classification. The training job is using a single p3.2xlarge instance and is taking too long. The dataset is stored in Amazon S3 and consists of 1 million images. The company wants to reduce training time without changing the model architecture. Which of the following should be done to achieve this?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Enable SageMaker distributed training using the Horovod framework and multiple instances.. Distributed training with Horovod across multiple instances parallelizes the training process, directly reducing the time required for each epoch. Since the model architecture is fixed, data parallelism is appropriate. Other options like Pipe mode or volume size may offer minor improvements but do not address the core compute bottleneck. Converting to RecordIO with File mode does not add parallelism.

Answer

Convert the dataset to the RecordIO format and use File mode.

Answer

Increase the size of the training volume attached to the p3.2xlarge instance.

Answer

Use SageMaker Pipe mode instead of File mode to stream data directly from S3 to the training container.

Question 2

A data scientist is training a deep learning model using Amazon SageMaker. The training data is stored in Amazon S3 and is approximately 2 TB in size. The training job is taking a long time to start because the data is being downloaded from S3 to the training instance's Amazon EBS volume. The data scientist wants to reduce the startup time and improve training performance. Which two actions should the data scientist take? (Choose two.)

Accepted Answer

Use Amazon FSx for Lustre as a data source and mount it to the training instance.. Using Pipe mode streams data directly from S3, eliminating the initial download delay, while Amazon FSx for Lustre provides a high-performance file system that allows immediate access to data without full download. Both actions significantly reduce startup time and improve training performance for large datasets. They are complementary and can be used together for optimal results.

Answer

Enable Amazon SageMaker Managed Spot Training to reduce cost.

Answer

Compress the training data into a single .tar.gz file and use File mode.

Answer

Increase the size of the Amazon EBS volume attached to the training instance.

Question 3

A machine learning engineer is using Amazon SageMaker to train a model. The training job is configured to use a Spot instance to reduce costs. The engineer wants to ensure that the training job can resume from the last saved checkpoint if the Spot instance is interrupted. Which SageMaker feature should the engineer use?

Accepted Answer

Enable managed spot training and specify an S3 bucket for checkpoints.. Managed spot training is the SageMaker feature that enables training on Spot instances with automatic checkpointing and resumption. By specifying an S3 bucket for checkpoints, the training job can save progress and resume from the last checkpoint after an interruption. Other options do not provide this capability.

Answer

Use SageMaker Debugger to save model checkpoints during training.

Answer

Configure the training job to use a persistent volume for checkpoint storage.

Answer

Set the training job's StoppingCondition to a longer duration to prevent interruption.

Question 4

A machine learning engineer is using Amazon SageMaker to train a model on a dataset that contains sensitive financial information. The company requires that the data be encrypted at rest and in transit, and that the training job uses a specific AWS KMS customer managed key for encryption. The engineer has configured the training job to use the KMS key for the output volume. What additional step must the engineer take to ensure that the training data in Amazon S3 is encrypted with the same KMS key?

Accepted Answer

Ensure that the S3 bucket objects are encrypted with the desired KMS key by specifying it during upload or by setting the bucket's default encryption to that key.. The encryption of S3 objects is determined at upload or by bucket default encryption. To use a specific KMS key, the data must be uploaded with that key or the bucket must default to it. The SageMaker execution role then needs kms:Decrypt permission for that key. Specifying the key in the training job input or using SSE-S3 does not achieve the requirement.

Answer

Specify the KMS key in the SageMaker training job's input data configuration using the S3DataSource parameter.

Answer

Attach an IAM policy to the SageMaker execution role that allows s3:GetObject with the condition that the object is encrypted with the KMS key.

Answer

Enable default encryption on the S3 bucket using SSE-S3 (AES-256) instead of a KMS key.

Question 5

A machine learning engineer must retrain a model nightly on data that grows by roughly 4 GB per day. The current training job reads 90 GB of CSV files from Amazon S3 and the input pipeline is the bottleneck, with GPU utilization below 20%. The engineer wants to reduce training time without changing the algorithm. Which action should the engineer take?

Accepted Answer

Convert the CSV data to the RecordIO protobuf format and use Pipe mode so training data streams directly from S3 to the training instances instead of being downloaded to local storage first.. When GPU utilization is low and the input pipeline is the bottleneck, the fix is to improve data delivery rather than add compute. Pipe mode with RecordIO streams shards from S3 to the training container on demand, removing the download-first pattern that delays training start and saturates local disk. This reduces startup time and keeps accelerators fed, which is precisely the stated goal.

Answer

Enable SageMaker Training Compiler for the training job so the framework compiles the model graph and the data loader into a more efficient representation that consumes input faster.

Answer

Increase the number of training instances and enable distributed training with the parameter server strategy so that each instance reads a smaller shard of the dataset from S3.

Answer

Attach an Amazon EFS file system to the training job and copy the CSV dataset there once, then point the training input channel at the EFS mount path for every subsequent nightly job.

Question 6

A machine learning engineer has trained a model using Amazon SageMaker and wants to deploy it for real-time inference. The model artifacts are stored in Amazon S3. The engineer needs to create a SageMaker endpoint that can serve predictions with low latency. Which step is required to deploy the model?

Accepted Answer

Create a SageMaker model, create an endpoint configuration, and create an endpoint.. The correct deployment process involves creating a SageMaker model that points to the model artifacts and inference image, then creating an endpoint configuration that specifies the instance type, and finally creating an endpoint that provisions the resources. This managed approach ensures low-latency, scalable real-time inference with minimal operational effort.

Answer

Use AWS CloudFormation to create an Amazon EC2 instance with the model and an Application Load Balancer.

Answer

Create a SageMaker notebook instance and run the model directly from the notebook.

Answer

Upload the model artifacts to Amazon ECR and create an AWS Lambda function.

Question 7

A company has deployed a real-time inference endpoint on Amazon SageMaker using a custom container. The endpoint is experiencing high latency and occasional timeouts during peak traffic. The model is CPU-bound and the container is configured with one CPU core and 2 GB of memory. The company wants to reduce latency and improve throughput without changing the model. Which action should a machine learning engineer take?

Accepted Answer

Increase the number of CPU cores allocated to the container and enable multi-threaded inference in the model server.. For a CPU-bound model, the best way to reduce latency and improve throughput is to allocate more CPU resources and enable multi-threaded inference. This allows the container to handle multiple requests concurrently. Auto scaling increases capacity but not per-request speed, GPU instances are unnecessary for CPU-bound models, and serverless may add cold start latency.

Answer

Enable auto scaling for the endpoint based on the InvocationsPerInstance metric.

Answer

Configure the endpoint to use a GPU instance type and enable GPU acceleration in the container.

Answer

Deploy the model to a SageMaker Serverless Inference endpoint to automatically scale with traffic.

Question 8

A machine learning engineer has trained a SageMaker model and registered it in the Amazon SageMaker Model Registry. Compliance requires that every production endpoint only serve model versions that have been reviewed and approved, and that no unreviewed artifact can ever be deployed to production. The engineer needs to wire the deployment pipeline so that only approved model versions are deployed. Which approach meets these requirements with the LEAST operational effort?

Accepted Answer

Use the SageMaker Model Registry model package group's approval status; configure the deployment pipeline to deploy only model package versions whose ModelApprovalStatus is Approved.. The SageMaker Model Registry is the designed control point for model governance: each registered model package version carries a ModelApprovalStatus that a CI/CD pipeline can query before deployment. Deploying strictly versions with Approved status creates an auditable, low-maintenance gate that satisfies the compliance requirement without inventing custom approval tracking or file-move rituals.

Answer

Enable SageMaker Model Monitor on the endpoint and configure a baseline that alerts when an unapproved model is detected serving traffic.

Answer

Copy each candidate model artifact to a separate S3 prefix named approved/ only after a reviewer manually re-uploads the tarball, and have the pipeline deploy from that prefix.

Answer

Attach an IAM policy to the endpoint's execution role that denies sagemaker:CreateModel unless the model name contains the string approved.

Question 9

A data scientist has a trained scikit-learn model artifact in Amazon S3 and needs a fully managed HTTPS endpoint that returns predictions in real time with automatic patching of the underlying infrastructure. The team has no Docker expertise and wants the fewest possible steps. Which SageMaker deployment approach should the data scientist use?

Accepted Answer

Use the SageMaker scikit-learn built-in framework container by calling the model deploy method with the framework version and the S3 model artifact path, letting SageMaker create the model, endpoint configuration, and endpoint.. SageMaker ships managed framework containers for scikit-learn, so an existing model artifact can be deployed by supplying the framework version and artifact path to the deploy method. SageMaker then creates the model, endpoint configuration, and endpoint, and handles infrastructure patching. This meets the managed, real-time, low-effort requirements without Docker work or custom serving code.

Answer

Build a custom inference container with the SageMaker Inference Toolkit, push it to Amazon ECR, and create a model that references the image URI and the model artifact in S3.

Answer

Package the model into an AWS Lambda function using the scikit-learn Python library and expose it through an Amazon API Gateway REST API with a proxy integration.

Answer

Create a SageMaker batch transform job with the model artifact and invoke it on demand each time a prediction is requested, using the job output as the response.

Question 10

A data scientist is using Amazon SageMaker to train a model on a large dataset. The training job is taking longer than expected, and the data scientist suspects that the input data pipeline is the bottleneck. The dataset consists of many small files stored in Amazon S3. The data scientist wants to improve the training throughput by optimizing data input. Which action should the data scientist take?

Accepted Answer

Combine the small files into larger files (e.g., TFRecord or RecordIO format) before training.. Combining many small files into larger files in an optimized format like TFRecord or RecordIO reduces the overhead of file operations and allows for more efficient sequential reads, which directly improves data input throughput. This is a best practice for training on large datasets with many small files.

Answer

Use a larger instance type with more CPU cores to speed up data loading.

Answer

Use the SageMaker Pipe mode with the built-in algorithm to stream data directly from S3.

Answer

Increase the number of epochs to allow the model to learn more from the data.

Question 11

A machine learning engineer is using Amazon SageMaker to train a model. The training job is taking longer than expected, and the engineer suspects that the data loading is the bottleneck. The training data is stored in Amazon S3 as many small files. Which of the following is the most efficient way to improve data loading performance?

Accepted Answer

Use SageMaker Pipe mode to stream data directly from S3.. SageMaker Pipe mode streams data directly from S3 to the training container, eliminating the need to download all data before training begins. This is particularly beneficial for many small files, as it reduces startup latency and improves I/O throughput. Other options either do not address the bottleneck or introduce additional overhead.

Answer

Use a larger instance type with more CPU and memory.

Answer

Increase the number of epochs to allow more time for data loading.

Answer

Download the data to the instance's Amazon EBS volume before training.

Question 12

A company runs a real-time SageMaker endpoint backed by an Auto Scaling policy. Traffic is highly variable: it is near zero overnight and spikes sharply every weekday at 09:00. Users report timeouts during the first few minutes of each spike because the endpoint scales out too slowly. The engineer must reduce cold-start scale-out latency while keeping costs low during idle periods. Which change is MOST effective?

Accepted Answer

Configure a scheduled scaling policy that raises the desired instance count shortly before 09:00 each weekday, combined with the existing target tracking policy.. The spike is predictable, so proactive capacity is the right tool: a scheduled scaling policy in Application Auto Scaling pre-provisions instances before 09:00, while target tracking continues to handle unpredictable variation within the spike. Because the schedule also scales back down overnight, idle costs stay controlled, giving the best latency improvement for this recurring pattern.

Answer

Enable Elastic Inference on the endpoint and attach an accelerator to each instance to absorb the spike.

Answer

Increase the target value of the TargetTrackingScaling policy so the endpoint scales out more aggressively at lower utilization.

Answer

Switch the endpoint to Serverless Inference with a provisioned concurrency of zero and rely on automatic scaling.

Question 13

A company operates a SageMaker real-time endpoint that serves a fraud-detection model. Compliance requires that predictions be explainable for individual transactions, and the data science team wants to detect when incoming feature distributions shift away from the training baseline. The team wants to use SageMaker built-in capabilities rather than custom code. Which two capabilities should they enable? (Choose two.)

Accepted Answer

SageMaker Clarify online explainability on the endpoint, configured with a baseline and a SHAP explainer, so each invocation can return feature attributions alongside the prediction.. Clarify online explainability produces SHAP-based attributions for individual real-time requests, meeting the explainability mandate, while Model Monitor data quality monitoring compares live request features to a training baseline and raises violations when distributions drift. Both are native SageMaker features configured through the SDK, so the team avoids writing custom drift or explanation code.

Answer

SageMaker Debugger with a built-in rule such as VanishingGradient, attached to the endpoint to capture tensors during inference and flag anomalous activation patterns.

Answer

SageMaker Batch Transform with a data capture configuration that stores every request and response in S3 for later offline analysis by the compliance team.

Answer

SageMaker Automatic Model Tuning with a hyperparameter tuning job that periodically retrains the model on recent transactions to keep the fraud-detection model aligned with current behavior.

Question 14

A machine learning engineer is deploying a scikit-learn model to a SageMaker endpoint. The model requires a Python preprocessing function that transforms raw JSON input into a NumPy array before prediction. The engineer wants to minimize inference latency and avoid writing a custom Docker container. Which SageMaker feature should be used?

Accepted Answer

SageMaker Inference Pipeline. Inference Pipeline is the correct choice because it lets you deploy multiple containers behind a single endpoint, with the preprocessing container handling raw input and passing transformed data to the model container. This eliminates the need for a custom container while keeping latency low, as both steps execute within the same endpoint.

Answer

SageMaker Model Monitor

Answer

SageMaker Batch Transform

Answer

SageMaker Neo

Question 15

A machine learning team wants to automatically detect when a deployed SageMaker model's predictions start to drift from the original training data distribution. They need to monitor feature distributions and receive alerts when significant deviations occur. Which SageMaker feature should they use?

Accepted Answer

SageMaker Model Monitor. SageMaker Model Monitor is purpose-built to detect drift by comparing production data to a baseline. It captures statistics and constraints, schedules monitoring jobs, and emits CloudWatch metrics that can trigger alarms. The other services address training debugging, bias explanation, or workflow orchestration, none of which provide continuous drift detection for deployed models.

Answer

SageMaker Pipelines

Answer

SageMaker Clarify

Answer

SageMaker Debugger