MLS-C01 Mock Exam 95 — 25 Questions

Question 1

A company has deployed a machine learning model on a SageMaker endpoint that serves predictions to a web application. The model uses a custom inference container that loads the model artifacts from an ECR repository. After updating the model with new training data, the data scientist creates a new model and updates the endpoint. However, some users report that they still get predictions from the old model. The data scientist confirms that the endpoint configuration points to the new model. What is the most likely cause?

Accepted Answer

DNS caching on the client side is resolving to the old endpoint IP address. Option B is correct because DNS caching at the client side may resolve the endpoint's DNS name to an old IP address, especially if the endpoint's underlying instances have not changed. Option A (incorrect model artifact) would affect all users. Option C (load balancer) is not part of SageMaker. Option D (CloudFront caching) could be an issue if CloudFront is in front, but the question does not mention it.

Answer

The new model artifacts are not correctly uploaded to S3

Answer

The endpoint is behind a load balancer that is not updated

Answer

The inference container is cached and not pulling the new image

Question 2

A company is using Amazon SageMaker to deploy a model for real-time inference. The endpoint uses an ml.c5.xlarge instance. The company wants to reduce costs without affecting performance. The current traffic pattern shows a daily peak of 500 requests per second for 2 hours, and the rest of the day sees fewer than 50 requests per second. The model has a cold start time of about 30 seconds. What should the company do?

Accepted Answer

Configure an auto scaling policy that scales down during low traffic and keep a minimum of 1 instance.. Option D is correct because adding a scaling policy to scale down during low traffic reduces cost, and keeping a minimum instance ensures low latency during low traffic without cold starts. Option A is wrong because serverless endpoints have cold starts and may not handle 500 TPS. Option B is wrong because Batch Transform is not for real-time. Option C is wrong because one instance during peak may cause latency.

Answer

Switch to a serverless inference endpoint.

Answer

Use a single ml.c5.xlarge instance and rely on it.

Answer

Use SageMaker Batch Transform for all predictions.

Question 3

A company uses Amazon SageMaker to train machine learning models. The data science team has developed a training script that uses TensorFlow. They want to run the training job on a GPU instance (ml.p3.2xlarge) and store the model artifact in Amazon S3. The training job completes successfully, but the model artifact is not saved to S3. The team has confirmed that the S3 bucket policy allows write access from the SageMaker execution role. The training script uses the TensorFlow estimator with the following configuration:

```
tensorflow_estimator = TensorFlow(
    entry_point='train.py',
    role='arn:aws:iam::123456789012:role/SageMakerExecutionRole',
    instance_count=1,
    instance_type='ml.p3.2xlarge',
    output_path='s3://my-bucket/output',
    framework_version='2.3',
    py_version='py37',
)
```

The train.py script saves the model using `model.save('/opt/ml/model')`. What is the MOST likely reason the model artifact is not being saved to S3?

Accepted Answer

The SageMaker execution role does not have the s3:PutObject permission for the S3 bucket.. The TensorFlow estimator's output_path specifies where the model artifact should be uploaded after training. However, SageMaker automatically uploads the contents of /opt/ml/model to S3 at the end of training. The script is saving to the correct directory. The issue is likely that the training script is not saving the model correctly or the training fails before saving. But given the job completes successfully, the most common cause is that the SageMaker execution role does not have permission to write to the S3 bucket. The bucket policy allows write access, but the IAM role may lack the necessary S3 permissions. Option C is correct because the role needs s3:PutObject permission on the bucket. Option A is incorrect because the output_path is correctly specified. Option B is incorrect because the script saves to the right directory. Option D is incorrect because the estimator does not have a 'model_dir' parameter that overrides the default; the default is /opt/ml/model.

Answer

The training script must save the model to /opt/ml/model/saved_model instead of /opt/ml/model.

Answer

The output_path parameter is incorrectly formatted; it should include a trailing slash.

Answer

The TensorFlow estimator requires the model_dir parameter to be set to the S3 output path.

Question 4

A company is using Amazon SageMaker to host a real-time inference endpoint for a natural language processing model. The endpoint is configured with an ml.m5.large instance. After deployment, the company observes that the inference latency is higher than expected, and the endpoint is experiencing CPU utilization near 100% during peak hours. The model is a PyTorch model that uses a transformer architecture. The company wants to reduce latency without increasing cost significantly. Which approach should the company take?

Accepted Answer

Change the endpoint instance type to ml.g4dn.xlarge to use GPU acceleration.. The issue is high CPU utilization causing latency. Using a GPU instance (ml.g4dn.xlarge) can accelerate inference for transformer models due to parallel processing, reducing latency. Option C is correct. Option A (Elastic Inference) may help but is less effective than a full GPU for transformer models; also, it adds complexity. Option B (Auto Scaling) helps with traffic but does not reduce per-request latency. Option D (batch transform) is for offline inference, not real-time.

Answer

Configure the endpoint with Auto Scaling to add more instances during peak hours.

Answer

Switch to batch transform for inference.

Answer

Attach an Elastic Inference accelerator to the existing instance.

Question 5

A data scientist is using Amazon SageMaker to train a model using a built-in algorithm. The training job uses a large dataset stored in Amazon S3, and the scientist wants to use pipe mode to stream the data directly from S3 to the training instance, reducing the time needed to download the data. The training job is configured with 'InputMode' set to 'Pipe'. However, the training job fails with an error indicating that the algorithm does not support pipe mode. What should the scientist do to resolve this issue?

Accepted Answer

Change the 'InputMode' to 'File'. Option A is correct because not all built-in algorithms support pipe mode; the scientist should use a file mode instead. Option B is wrong because the issue is not with the instance type. Option C is wrong because changing to a different algorithm may not be necessary if the current algorithm works with file mode. Option D is wrong because SageMaker does not support using Glue to stream data directly to training jobs.

Answer

Use a different instance type that supports pipe mode

Answer

Use AWS Glue to stream the data to the training instance

Answer

Switch to a different built-in algorithm that supports pipe mode

Question 6

A data scientist is using Amazon SageMaker to train a model using a built-in algorithm. The training job is taking a long time, and the data scientist wants to improve performance by using a larger instance type with more vCPUs. The training job is currently using an ml.m5.large instance. The data scientist changes the instance type to ml.m5.4xlarge and resubmits the training job. However, the training time does not decrease significantly. What is the MOST likely reason?

Accepted Answer

The built-in algorithm is not designed to scale with additional vCPUs.. The built-in algorithm may not be able to utilize additional vCPUs effectively if it is not parallelized. Option B is correct. Option A is incorrect because data size does not prevent parallelism. Option C is incorrect because the algorithm is not inherently single-threaded; it depends on implementation. Option D is unlikely because training on larger instances generally costs more, but that does not affect time.

Answer

The algorithm is single-threaded and cannot use multiple vCPUs.

Answer

The training job is I/O bound, and increasing vCPUs does not help.

Answer

The training dataset is too small to benefit from more vCPUs.

Question 7

A machine learning engineer is deploying a custom XGBoost model for real-time inference on Amazon SageMaker. The model was trained using the SageMaker XGBoost built-in algorithm. The endpoint is deployed with an ml.m5.large instance and is receiving around 50 requests per second. The engineer notices that the endpoint's latency is around 200 ms, but the requirement is under 100 ms. The model's serialized format is a .tar.gz file. The engineer wants to reduce inference latency without modifying the model or retraining. What should the engineer do?

Accepted Answer

Use SageMaker Neo to compile the model for the target instance.. Option C is correct because SageMaker Neo optimizes trained models for target hardware, reducing latency. Option A is wrong because SageMaker Batch Transform is for batch, not real-time. Option B is wrong because SageMaker Debugger monitors training, not inference. Option D is wrong because Elastic Inference attaches GPU acceleration for deep learning, not XGBoost which is tree-based.

Answer

Configure SageMaker Debugger to optimize the inference code.

Answer

Use SageMaker Elastic Inference to attach an accelerator.

Answer

Use SageMaker Batch Transform instead of a real-time endpoint.

Question 8

A company has a SageMaker endpoint that serves predictions for a mobile app. The endpoint is deployed on a single ml.m5.large instance. Recently, users have reported that the app sometimes returns outdated predictions. The data science team has confirmed that the model is updated daily by retraining with new data and creating a new endpoint configuration. However, the endpoint still returns predictions from the old model for some requests. The team has verified that the new endpoint configuration is associated with the endpoint and that the endpoint is in service. What is the most likely cause of this issue?

Accepted Answer

The endpoint has multiple variants and the old variant still has a weight assigned. Option D is correct because if the old endpoint variant still has a weight greater than zero, some traffic will continue to be routed to the old model. Option A is wrong because if the new configuration were not active, the endpoint would not be in service. Option B is wrong because the model artifacts are stored in S3 and are not deleted during updates. Option C is wrong because CloudFront does not cache SageMaker endpoint responses by default.

Answer

The old model artifacts are still being cached by the endpoint

Answer

The mobile app is using a CDN that caches the predictions

Answer

The new endpoint configuration has not been deployed to the endpoint

Question 9

A company is deploying a machine learning model using AWS Lambda for real-time inference. The model is a large ensemble model that takes approximately 500 MB of memory. The Lambda function is configured with 1024 MB of memory and a timeout of 15 seconds. The company observes that the function frequently times out during inference. The company wants to keep using Lambda for its serverless benefits. Which solution should the company implement to reduce inference time?

Accepted Answer

Increase the Lambda function memory to 3008 MB to provide more CPU resources.. Lambda has a maximum memory of 10,240 MB and a maximum timeout of 15 minutes. Increasing memory to 3008 MB gives more CPU power and reduces inference time. Option A is correct. Option B (ElastiCache) adds latency and cost. Option C (Step Functions) adds orchestration overhead. Option D (SageMaker) moves away from serverless, which the company wants to keep.

Answer

Deploy the model on Amazon SageMaker hosting instead of Lambda.

Answer

Use AWS Step Functions to invoke the Lambda function asynchronously.

Answer

Use Amazon ElastiCache to cache model predictions and reduce computation.

Question 10

A machine learning team is using Amazon SageMaker to train a PyTorch model on a dataset that is 500 GB in size. The training job runs on a single ml.p3.2xlarge instance, but the training takes over 48 hours, which exceeds the maximum allowed time. The team wants to reduce training time to under 24 hours. They are open to using multiple instances and have budget for up to 4 instances. The dataset is stored in Amazon S3 and can be split into shards by a key. The model architecture must remain unchanged. What should the team do?

Accepted Answer

Use SageMaker distributed data parallelism with 4 ml.p3.2xlarge instances.. Option D is correct because SageMaker's distributed data parallelism library (SMDDP) can efficiently split data across multiple GPUs with minimal code changes. Option A is wrong because increasing instance type alone may not halve training time. Option B is wrong because Pipe mode reduces I/O but not computation time. Option C is wrong because SageMaker Processing is for preprocessing, not training.

Answer

Use SageMaker Processing to split the data and train separate models.

Answer

Change the instance type to ml.p3.16xlarge.

Answer

Switch to Pipe input mode to stream data faster.

Question 11

A company has deployed a machine learning model on Amazon SageMaker for real-time inference. The endpoint uses a single ml.c5.xlarge instance. Recently, the traffic has increased, and the endpoint is returning HTTP 503 (Service Unavailable) errors during peak hours. The CloudWatch metrics show that the CPU utilization is consistently above 90% during peak times, and the Invocations metric shows that requests are being throttled. The data science team has already optimized the model to reduce inference time by 20%, but the errors persist. The company needs to resolve the issue without increasing costs significantly. Which course of action should be taken?

Accepted Answer

Configure auto-scaling for the endpoint to add instances based on CPU utilization. Option A is correct because adding auto-scaling based on CPU utilization or invocations will dynamically adjust the number of instances to handle the load, reducing errors without incurring costs during low traffic. Option B is wrong because increasing to a larger instance type will increase costs even during low traffic. Option C is wrong because switching to batch transform is for offline processing, not real-time. Option D is wrong because using spot instances could lead to interruptions and does not solve the capacity issue.

Answer

Change the instance type to a larger size, such as ml.c5.2xlarge

Answer

Switch to batch transform to process requests in batches

Answer

Use spot instances to reduce costs and add more instances

Question 12

A machine learning engineer is responsible for deploying a model that was trained using a custom algorithm in Amazon SageMaker. The engineer has built a Docker container that includes the inference code and has tested it locally. The engineer now wants to deploy the container to a SageMaker endpoint for real-time inference. The engineer has already created the model in SageMaker by specifying the image URI and the model artifacts location in S3. However, when the engineer tries to create an endpoint configuration, the operation fails with an error indicating that the model is not in an 'Active' state. What should the engineer do to resolve this issue?

Accepted Answer

Check the CloudWatch logs for the container to ensure the inference server starts correctly. Option C is correct because the model must be in an 'Active' state before it can be deployed, and this requires the container to pass SageMaker's health check. The engineer should check the CloudWatch logs for the container to diagnose the health check failure. Option A is wrong because re-creating the model with the same image will not fix the health check issue. Option B is wrong because the model is already created; the issue is the state. Option D is wrong because the endpoint configuration cannot be created if the model is not active.

Answer

Create the endpoint configuration with a different model name

Answer

Delete and re-create the model, then wait for a few minutes

Answer

Re-create the model using a different image URI

Question 13

A company is using Amazon SageMaker to train a deep learning model. The training job uses a script that reads data from Amazon S3 using the SageMaker SDK's `s3_input` method. The training job runs on a single ml.p3.2xlarge instance. The data scientist notices that the GPU utilization is very low during training, often below 20%. The training dataset is large, approximately 50 GB, stored as TFRecord files in S3. What is the MOST likely cause of low GPU utilization?

Accepted Answer

The data loading pipeline is not optimized, causing the GPU to wait for data.. Low GPU utilization often indicates that the data loading pipeline is a bottleneck. The training script may not be using efficient data loading techniques like prefetching and parallel data extraction. Option A is correct. Option B (batch size) could be a factor but is less likely given TFRecord format. Option C (instance type) is unlikely because ml.p3.2xlarge has a capable GPU. Option D (framework) is not the cause.

Answer

The training script is using a CPU-only version of TensorFlow.

Answer

The batch size is too large, causing the GPU to run out of memory.

Answer

The instance type does not have enough GPU memory for the model.

Question 14

A company uses Amazon SageMaker to deploy a model for real-time inference. The model is a linear regression model that was trained using the SageMaker built-in Linear Learner algorithm. The endpoint is configured with an ml.m5.large instance. After deployment, the company notices that the endpoint returns incorrect predictions. The training data was normalized, but the inference requests send raw feature values without normalization. What should the company do to fix the issue?

Accepted Answer

Create a SageMaker inference pipeline that includes a preprocessing step to normalize the input data before passing it to the model.. The model expects normalized input. The inference pipeline must include a preprocessing step to normalize the data. Using a SageMaker inference pipeline with a preprocessing container (e.g., scikit-learn) before the model container is the correct approach. Option B is correct. Option A (retrain with raw data) is a viable alternative but would require retraining and may reduce model performance. Option C (transform job) is for batch inference, not real-time. Option D (change instance type) does not address the data mismatch.

Answer

Retrain the model using raw data without normalization.

Answer

Change the endpoint instance type to a GPU instance to handle the raw data.

Answer

Use a batch transform job to preprocess the data before sending it to the endpoint.

Question 15

A media company ingests video metadata from multiple sources into an Amazon S3 bucket. Each metadata record is a JSON file about 2 KB. They use AWS Glue ETL jobs to process these files and load them into Amazon Redshift for analytics. The jobs currently run hourly and take about 10 minutes to process all new files. However, the company is growing and expects the number of files to increase 100x. The data engineering team wants to minimize processing time and cost. The Glue job currently reads all files from the S3 bucket using a full scan. What should they do to optimize the pipeline?

Accepted Answer

Use S3 event notifications to trigger Glue jobs only for new files. Using S3 event notifications to trigger Glue jobs on new objects eliminates scanning and reduces latency. Option A is wrong because increasing DPUs costs more. Option B is wrong because consolidating files reduces number of objects. Option D is wrong because converting to Parquet helps but still scanning all files.

Answer

Consolidate the small JSON files into larger files using a scheduled job

Answer

Convert the data to Parquet format and partition it

Answer

Increase the number of Glue DPUs to process files faster

Question 16

A financial services company uses Amazon Kinesis Data Streams with 50 shards to ingest real-time stock trade data. The data is consumed by a custom Java application running on Amazon EC2 instances. Recently, the application has been experiencing high latency, and CloudWatch metrics show that the average iterator age is increasing. The application uses the Kinesis Client Library (KCL) with DynamoDB for lease tracking. The EC2 instances are in an Auto Scaling group with a minimum of 2 and maximum of 10 instances, and the current CPU utilization is below 50%. The team wants to reduce latency without increasing costs significantly. What should they do?

Accepted Answer

Increase the maximum size of the Auto Scaling group and set a scaling policy based on iterator age. Increasing the number of consumers (EC2 instances) allows parallel processing of shards, reducing iterator age. Option A is wrong because reducing shards would increase load per shard. Option C is wrong because DynamoDB provisioned throughput is not the bottleneck. Option D is wrong because enabling enhanced fan-out is for multiple consumers, not the same consumer group.

Answer

Increase the provisioned read capacity of the DynamoDB lease table

Answer

Enable enhanced fan-out on the Kinesis stream

Answer

Increase the number of shards in the Kinesis stream

Question 17

An e-commerce company uses Amazon Kinesis Data Firehose to deliver clickstream data to an Amazon S3 bucket. The data is then queried using Amazon Athena. The marketing team wants to run daily reports that aggregate click events by product ID. However, the reports are slow because Athena scans the entire dataset each time. The data is partitioned by date (e.g., s3://bucket/clickstream/2023/01/01/). The product ID is a column within the data. The data engineering team wants to improve query performance without moving the data to another service. Which approach should the team take?

Accepted Answer

Repartition the data by product ID in addition to date. Partitioning by product ID would allow Athena to prune partitions for queries filtering by product ID. Option A is wrong because converting to Parquet alone may improve but does not eliminate full scan. Option C is wrong because creating a view does not change physical storage. Option D is wrong because Redshift Spectrum still requires scanning.

Answer

Convert the data from JSON to Parquet format

Answer

Use Amazon Redshift Spectrum to query the data

Answer

Create a view in Athena that filters by product ID

Question 18

A company runs a real-time recommendation system that uses Amazon SageMaker endpoints for inference. The system ingests user activity data from a mobile app via Amazon API Gateway and AWS Lambda, which writes events to an Amazon Kinesis Data Stream. A second Lambda function consumes the stream, calls a SageMaker endpoint to generate recommendations, and stores the results in Amazon DynamoDB. The system has been working well, but recently the team noticed an increase in latency from the time a user action occurs to when the recommendation is stored. The SageMaker endpoint shows increased invocation latency but no throttling. CloudWatch metrics show that the Kinesis stream's IteratorAgeMilliseconds is increasing, indicating the consumer is falling behind. The Lambda consumer's duration is within limits, but the number of invocations is lower than expected. The team suspects the issue is with the event source mapping. Which course of action should the team take to reduce the latency?

Accepted Answer

Increase the batch size in the event source mapping to process more records per invocation.. Option B is correct because the consumer is falling behind despite adequate Lambda duration, suggesting that the batch size or parallelization factor is too low. Increasing the batch size allows each Lambda invocation to process more records, increasing throughput. Option A is wrong because increasing shards increases cost and may not help if the consumer is the bottleneck. Option C is wrong because reducing concurrency would worsen the situation. Option D is wrong because the Lambda function is already consuming from Kinesis; using Firehose would not directly solve the consumer lag.

Answer

Increase the number of shards in the Kinesis data stream to increase parallelism.

Answer

Decrease the Lambda function's reserved concurrency to force it to scale down.

Answer

Replace the Lambda consumer with an Amazon Kinesis Data Firehose delivery stream.

Question 19

A data engineering team is building a real-time fraud detection pipeline. The pipeline ingests transaction data from an Amazon Kinesis Data Stream with 10 shards. Each shard produces about 500 records per second, each record is 2 KB. The data is processed by a Lambda function that runs for about 200 ms and then writes results to an Amazon DynamoDB table. The team notices that the Lambda function is experiencing a high number of throttles, and there are increasing numbers of records being retried. The Lambda function's reserved concurrency is set to 100. The DynamoDB table has 100 read capacity units and 100 write capacity units. Which change would most effectively reduce throttling and improve processing throughput?

Accepted Answer

Increase the Lambda function's reserved concurrency to 1000.. The Lambda function is throttling because the concurrent executions needed exceed the reserved concurrency. Each shard invokes Lambda with batches, and with 10 shards and a batch size of 100 (default), the number of concurrent invocations can be high. Increasing reserved concurrency to a higher value (e.g., 1000) allows more concurrent executions, reducing throttling. However, if DynamoDB write capacity is also a bottleneck, increasing it might help. But the primary issue is Lambda throttling.

Answer

Decrease the Lambda function's batch size to 10.

Answer

Increase the DynamoDB write capacity units to 1000.

Answer

Increase the number of shards in the Kinesis stream to 100.

Question 20

A startup is building a data pipeline that ingests data from multiple sources into an Amazon S3 data lake. The data includes CSV files from legacy systems, JSON from web APIs, and Avro from mobile apps. The data must be transformed into Parquet format and cataloged for querying with Amazon Athena. The pipeline must be serverless and minimize operational overhead. The team has decided to use AWS Glue for ETL and cataloging. However, they are concerned about the cost of running Glue jobs continuously. The data arrives in small batches every 10 minutes. Which approach should the team use to minimize cost while meeting the requirements?

Accepted Answer

Use scheduled Glue jobs to process the data every hour, consolidating multiple batches. Triggering Glue jobs on schedule (e.g., every hour) to process accumulated data reduces the number of job runs and cost, while still meeting near-real-time needs. Option A is wrong because continuous streaming with Firehose may not handle all source formats. Option C is wrong because using Lambda for transformation is limited by timeout and memory. Option D is wrong because running a single daily job may introduce too much latency.

Answer

Use AWS Lambda functions to transform each file upon arrival and store as Parquet

Answer

Use Amazon Kinesis Data Firehose to stream data directly into S3 and use Glue to catalog it

Answer

Use a single daily Glue job to process all data at once