AWS Certified Machine Learning - Specialty - (MLS-C01) Exam Questions

Consider the integration points between EMR Spark and SageMaker. Choose based on where your processing and model training will primarily occur. The simplest architecture is one that minimizes maintenance and leverages AWS SageMaker's built-in features to the fullest.

Be aware that you can utilize the SageMaker Spark library to invoke SageMaker from an EMR Spark cluster, or alternatively, use Sparkmagic or Livy to access EMR Spark from a SageMaker notebook. The decision on which approach to use hinges on whether your workflow involves an EMR batch pipeline requiring integration with SageMaker, or vice versa.

Regarding model selection, several options are available:

• SageMaker Spark offers an XGBoostEstimator

• SageMaker features the SageMaker XGBoost algorithm

• XGBoost PySpark Estimator

Correct Choice: Use SageMaker Spark to preprocess data, train with XGBoostSageMakerEstimator, and host on SageMaker.

The SageMaker Spark library facilitates the execution of Spark jobs as part of the machine learning pipeline within SageMaker, without the user needing to set up and manage an EMR cluster or deal with the intricacies of Spark cluster configuration and scaling.

Incorrect Choice: Preprocess data on EMR Spark, save in S3, use SageMaker to train XGBoost, and host for inference.

Valid solution; however, it overlooks AWS integration between EMR and SageMaker, and adds an unnecessary intermediary step via S3.

Incorrect Choice: Configure Sparkmagic in SageMaker, preprocess data on EMR Spark via SageMaker notebook, train with SageMaker XGBoost, and host on SageMaker for inference.

Valid solution; however, it necessitates provisioning an EMR Spark cluster.

Incorrect Choice: Configure Livy in SageMaker, preprocess data on EMR Spark via SageMaker notebook, train XGBoost PySpark Estimator, and host on SageMaker for inference.

Valid solution; however, it necessitates provisioning an EMR Spark cluster.

Review SageMaker's data input modes (File, Pipe, FastFile) and understand their impact on EBS usage to minimize costs.

Correct Choice: Set TrainingInputMode to Pipe

Using Pipe mode streams data directly from S3 to the algorithm, reducing the need to use and store data on EBS volumes, hence lowering costs associated with EBS. This mode is ideal for large datasets.

Incorrect Choice: Set TrainingInputMode to File

This setting downloads the entire dataset from S3 to the EBS volume attached to the training instance, which would increase the EBS usage and costs, contrary to the objective. Valid when data needs pre-processing locally.

Incorrect Choice: Set TrainingInputMode to FastFile

While similar to 'File' mode and potentially reducing some I/O overhead, FastFile still uses EBS storage, incurring costs. Best for scenarios needing faster data access but less concerned about storage costs. Enable `FastFile` in configurations where rapid file access is more critical than minimizing storage costs. PCA on SageMaker does not support FastFile mode; only File or Pipe modes are supported.

Incorrect Choice: Set EnableSageMakerMetricsTimeSeries to false

Disabling time series metrics reduces monitoring detail but doesn't impact EBS costs directly. Relevant if optimizing monitoring costs, not storage costs. Since EnableSageMakerMetricsTimeSeries is set to false by default, changing this setting does not affect the solution.

Review EFS performance metrics (IOPS, throughput, connections) and relate them to the question.

Correct Choice: Increase the provisioned throughput of the EFS file system if it is in the provisioned mode.

1. For Bursting Throughput Mode: If `PercentIOLimit` is approaching 100%, increasing the total storage size will automatically raise the baseline performance and burstable IOPS capacity. This option leverages the natural scaling feature of Bursting Throughput mode.

2. For Provisioned Throughput Mode: Alternatively, if the file system is already in Provisioned Throughput mode or if a more immediate and predictable performance enhancement is needed, manually adjust the `ProvisionedThroughput` setting. This direct intervention ensures performance does not degrade as the `PercentIOLimit` approaches its maximum.

Incorrect Choice: Decrease the number of files stored in the EFS file system to reduce IOPS usage.

Reducing the number of files in EFS to decrease IOPS usage is not typically effective because IOPS limits are more directly influenced by the nature of the file operations and the throughput mode, rather than simply the quantity of files. This action might not significantly impact IOPS utilization if the remaining operations continue to be read/write intensive.

Incorrect Choice: Convert the EFS file system from General Purpose to Max I/O performance mode.

Switching from General Purpose to Max I/O performance mode in EFS is intended for file systems that require high levels of aggregate throughput and IOPS across multiple connections, but it does not directly address the issue of nearing the IOPS capacity limit of a system under heavy load. This switch might improve performance under certain circumstances but doesn't directly manage or alleviate reaching the IOPS capacity. Additionally, Max I/O performance mode may introduce higher latencies for file operations.

Incorrect Choice: Reconfigure attached EC2 instances to use Elastic Block Store (EBS) instead of EFS.

Moving from EFS to EBS involves significant architectural changes and is not a direct remediation for high IOPS usage in EFS. EBS provides block-level storage and is used for different types of workloads compared to EFS, which offers file-level storage.

Explanation:

✅ AWS Glue with PySpark to write to Parquet in S3

• This is a best practice for transforming semi-structured data into an efficient, columnar format.

• Apache Parquet is optimized for analytics, supports schema evolution, and integrates well with SageMaker via training on data in S3.

• PySpark allows custom transformations including flattening nested JSON.

✅ AWS Glue with DynamicFrames to flatten and write Parquet

• Similar to C, but uses DynamicFrames, which are specifically designed for semi-structured data like JSON.

• Writing to S3 in Parquet format with partitioning (e.g., by event date) improves query performance and reduces storage costs.

❌ Store in Redshift as normalized tables

• While Redshift is good for analytics, normalized relational tables can complicate feature extraction due to joins.

• Also, not ideal for SageMaker training jobs, which often work better with flat files in S3.

❌ Stream to Elasticsearch

• Amazon Elasticsearch (now OpenSearch) is good for text search and real-time dashboards, not cost-effective for large-scale historical analytics or model training.

• Storage costs and JSON query limitations make it a poor fit here.

❌ Athena to CSV in RDS

• Athena is suitable for querying, but writing output to CSV and storing in RDS introduces inefficiencies.

• CSV is not optimal for analytics or ML training.

• RDS is expensive for storing large analytical datasets.

Review best practices for hyperparameter tuning, especially regarding resource optimization and search technique effectiveness.

Correct choice: Decrease the number of concurrent hyperparameter tuning jobs

Reducing concurrency minimizes resource usage and costs, allowing for more focused and potentially insightful individual job analyses without overwhelming your compute resources.

Correct choice: Use logarithmic scales on your parameter ranges

Logarithmic scales are effective for exploring parameters that affect model performance exponentially, not linearly, helping to quickly identify optimal values in a broad search space.

Incorrect choice: Disable reverse logarithmic scales on your parameter ranges

If the parameter is highly sensitive near 1, disabling reverse logarithmic scaling could slow the search for an optimal value, especially for parameters within 0<=x<1.0 range.

Incorrect choice: Switch to grid search tuning strategy

Grid search is exhaustive and may increase costs significantly compared to other strategies like random or Bayesian optimization, especially with a large parameter space.

Incorrect choice: Disable the use of a parent job for the warm start configuration.

Removing the parent job disregards potentially valuable insights from previous tunings, missing out on accelerating the convergence to optimal hyperparameters based on past learnings.

Assess each storage option's capacity to integrate diverse datasets efficiently, ensuring data integrity and accessibility for ML model training.

Correct Choice: Store data in a centralized data lake

Centralized data lakes support diverse data formats and large volumes, essential for aggregating disparate health records. This architecture simplifies data management, enhances accessibility for analysis, and is scalable. Implementation involves setting up storage in AWS S3, structuring data by date and patient, and using metadata tagging to objects.

Incorrect Choice: Store data in an on-premise database

On-premise databases might struggle with the scale and variety of data typical in healthcare settings. They can be suitable for localized, smaller-scale applications where real-time access to a centralized cloud solution is not critical. Implementation would involve setting up and maintaining hardware, which can be costly and less scalable.

Incorrect choice: Store data on FTP servers

FTP servers are outdated for handling sensitive, large-scale datasets like patient information due to security and efficiency concerns. They might be used for transferring smaller, less sensitive files within an internal network. Implementation involves setting up FTP access points, which is not recommended for sensitive data.

Incorrect Choice: Store encrypted data in isolated AWS RDS instances

Using isolated RDS instances can introduce unnecessary complexity and hinder data integration across multiple sources. While secure, it's less efficient for tasks requiring holistic data access and analysis. This approach is more suited for structured data requiring high transaction rates. Implementation would involve setting up multiple RDS instances, managing data encryption, and ensuring proper access controls.

Focus on the primary function of each AWS service mentioned and match it to the task of labeling data for training a sentiment analysis model.

Correct choice: Employ Amazon SageMaker Ground Truth to annotate historical tweets with positive or negative sentiments, utilizing the labeled data to train a sentiment analysis model on SageMaker.

Ground Truth is specifically designed for data labeling, providing a direct path to creating accurately labeled datasets for training machine learning models, like sentiment analysis, in SageMaker.

Incorrect choice: Employ Amazon SageMaker Blazing Text to annotate historical tweets with positive or negative sentiments, utilizing the labeled data to train a sentiment analysis model on SageMaker.

BlazingText is used for natural language processing tasks such as word embeddings, not for annotating or labeling data, making it unsuitable for initial data annotation.

Incorrect choice: Employ Amazon Mechanical Turk to annotate historical tweets with positive or negative sentiments, utilizing the labeled data to train a sentiment analysis model on SageMaker.

While Mechanical Turk can be used for manual labeling, it lacks the integrated machine learning-based annotation features of Ground Truth, potentially affecting efficiency and scalability.

Incorrect choice: Employ Amazon SageMaker Random Cut Forest algorithm to annotate historical tweets with positive or negative sentiments, utilizing the labeled data to train a sentiment analysis model on SageMaker.

Random Cut Forest is designed for anomaly detection, not for labeling data with sentiments. It does not perform data annotation but identifies outliers within data, making it inappropriate for sentiment labeling.

Identify the workflow: data ingestion, processing, model training, and result storage. Focus on services that offer scalable, cost-effective solutions for each step, especially considering traffic variability and real-time recommendation requirements.

Correct choice: Capture clickstream data using Amazon Kinesis Data Firehose to Amazon S3. Process the data with Amazon SageMaker for model training using Managed Spot Training. Publish results to Amazon DynamoDB for instant recommendation serving.

This choice efficiently manages high-volume data ingestion (Kinesis Firehose to S3), cost-effective processing and model training (SageMaker with Spot Training), and real-time recommendation serving (DynamoDB), aligning with requirements for scalability and cost efficiency.

Incorrect choice: Stream clickstream data into Amazon S3 via Amazon Kinesis Data Streams, then use AWS Glue for real-time ETL processing. Utilize Amazon SageMaker for model training, adjusting capacity with Spot Instances as needed. Store outcomes in DynamoDB for live recommendations.

While this setup uses AWS services effectively, real-time processing with AWS Glue is less suited for the detailed model training scenario described, which benefits more from batch processing and analysis.

Incorrect choice: Direct clickstream data to Amazon S3 using Amazon Kinesis Data Firehose, conducting nightly analysis with AWS Glue DataBrew and Amazon SageMaker using On-Demand Instances for model training. Deploy results to DynamoDB for real-time recommendations.

This approach is valid but less cost-effective due to the use of On-Demand Instances for model training. Spot Instances or Managed Spot Training offer similar capabilities with better cost management.

Incorrect choice: Route clickstream data to Amazon Managed Streaming for Apache Kafka (Amazon MSK), then process in real-time with Amazon SageMaker for predictive modeling. Persist model insights in Amazon Aurora for delivering real-time content recommendations.

Amazon MSK and Aurora introduce complexity and potential over-provisioning for this use case. The initial question suggests a need for simplicity and cost efficiency, which is better served by the direct S3 to SageMaker to DynamoDB pipeline.

Evaluate each data format based on compatibility with MXNet, efficiency in handling large image files, and the potential to reduce training time and I/O overhead.

Correct Choice: RecordIO

RecordIO format is specifically optimized for high throughput and efficient data serialization. It is ideal for large-scale image datasets in MXNet, reducing read times and enhancing overall training efficiency.

Incorrect Choice: ORC

ORC is optimized for storing tabular data, not suitable for image datasets.

Incorrect Choice: Parquet

Parquet is designed for efficient storage of columnar text data, not binary image data.

Incorrect Choice: TFRecord

TFRecord is tailored for TensorFlow applications, not for MXNet-based workflows.

Assess if the architecture saves network round trips and whether a pre-trained street sign classifier is available.

Correct choice: Amazon SageMaker, Neo, Greengrass

SageMaker trains models, Neo optimizes and compiles for edge devices, and Greengrass allows local execution on edge, reducing network latency.

Incorrect choice: Amazon Rekognition, Lambda, IoT Core

Rekognition offers pre-trained models not specialized for street signs, Lambda and IoT Core introduce potential network latency for high-volume, real-time processing.

Incorrect choice: Amazon SageMaker, ECS, EC2 Spot Instances

While feasible, this setup mainly suits cloud-based processing, potentially increasing latency due to network trips and not optimized for edge computing.

Incorrect choice: Amazon SageMaker Ground Truth, Rekognition Custom Labels, Lambda, S3, DeepLens

Ground Truth and Rekognition Custom Labels could tailor models for street signs, but this setup might not handle thousands of images per second efficiently due to hardware limitations and potential network latency in coordinating these services for real-time inference.