AWS Certified Machine Learning – Specialty

D. SageMaker Neo

SageMaker Neo is a model compilation service that optimizes models from various frameworks for specific hardware platforms (including SageMaker instances with accelerators) to achieve faster inference and lower costs. It produces a compiled, deployable artifact.

B. SageMaker Elastic Inference

SageMaker Elastic Inference allows you to attach GPU acceleration to CPU-based EC2 instances or SageMaker instances at a fraction of the cost of a full GPU instance. Since the model requires GPU resources but inference is infrequent and latency is not extremely critical, Elastic Inference provides a cost-effective way to get GPU acceleration without paying for oversized GPU instances, aligning perfectly with the cost optimization goal for infrequent workloads.

A. ContainerStartupHealthCheckTimeoutInSeconds

The `ContainerStartupHealthCheckTimeoutInSeconds` parameter defines the maximum amount of time SageMaker waits for the container to emit a successful response to the `/ping` health check. If a container has large dependencies or complex initialization logic, increasing this timeout allows more time for it to become ready, preventing premature health check failures and deployment issues.

A. Deploy to a SageMaker Multi-AZ endpoint.

A SageMaker Multi-AZ endpoint automatically distributes inference requests across instances in multiple Availability Zones. If an AZ becomes unhealthy, traffic is automatically routed to healthy instances in other AZs, providing built-in high availability and resilience against AZ outages without manual intervention or additional services like Route 53 for failover management.

A. Amazon SageMaker Pipelines

Amazon SageMaker Pipelines is specifically designed for building, automating, and managing end-to-end machine learning workflows (MLOps pipelines), making it the best choice for orchestrating daily model retraining, evaluation, and deployment.

A. Canary deployment

Canary deployment involves gradually rolling out a new model version to a small subset of users (the 'canary' group) while the majority of traffic still goes to the old version. This allows for real-world testing and monitoring of the new model's performance and stability with minimal risk, as issues affect only a small percentage of users, and a quick rollback is possible by simply removing the canary variant.

B. SageMaker Endpoint Auto Scaling and Multi-AZ deployment.

SageMaker Endpoint Auto Scaling dynamically adjusts the number of inference instances based on load, ensuring capacity during traffic spikes. Multi-AZ deployment distributes these instances across multiple Availability Zones, providing high availability and resilience against single-point failures, which is crucial for consistent low latency and service continuity.

C. SageMaker Real-Time Endpoint with data capture enabled.

SageMaker Real-Time Endpoints can be configured with data capture, which automatically saves a configurable percentage of inference requests and responses (including input data and prediction results) to an S3 bucket. This raw data, combined with CloudWatch logs for model versioning and timestamps, creates an immutable audit trail necessary for regulatory compliance and deep analysis.

A. SageMaker Data Capture

SageMaker Data Capture allows you to configure a real-time inference endpoint to automatically capture and save input and output data, as well as model predictions, to an Amazon S3 bucket. This data can then be used for model monitoring, debugging, auditing, and retraining purposes.

B. SageMaker Feature Store.

SageMaker Feature Store is specifically designed to address the challenges of feature management for ML, providing a centralized repository for features. It supports both online (low-latency retrieval for inference) and offline (point-in-time correctness for training) access, along with versioning and consistent feature definitions.

D. A/B Testing

A/B testing on SageMaker endpoints allows routing a small percentage of live traffic to a new model (variant) while the majority of traffic goes to the current model (production), enabling direct comparison of performance metrics with minimal user impact.

D. SageMaker endpoint

SageMaker endpoints are designed for real-time, low-latency, and high-availability inference, making them ideal for interactive applications like fraud detection.

D. Deploying the model to a multi-instance SageMaker endpoint across multiple Availability Zones.

Deploying a multi-instance SageMaker endpoint across multiple Availability Zones provides the highest level of availability and resilience. If one instance or an entire AZ fails, traffic is automatically routed to healthy instances in other AZs, ensuring continuous operation and low latency.

B. Configuring a Multi-AZ endpoint

Configuring a Multi-AZ endpoint for Amazon SageMaker ensures that the inference endpoint is deployed across multiple Availability Zones. This provides high availability and resilience against single Availability Zone outages or instance failures within a single AZ.

C. Canary Deployment

Canary deployment allows gradually routing a small percentage of live traffic to the new model (Model B) while the majority still goes to the old model (Model A). This enables real-time monitoring of the new model's performance with actual user traffic. If issues arise, traffic can be instantly rolled back to Model A, minimizing impact. This fits the requirement for gradual shifting and quick rollback.

A. SageMaker Neo

SageMaker Neo is designed to compile machine learning models to optimize them for specific hardware targets (like IoT Greengrass devices), resulting in up to 2x faster inference and reduced memory footprint.

B. SageMaker Model Parallelism (Inference)

SageMaker Model Parallelism (Inference) is specifically designed to deploy very large models, like LLMs, by automatically splitting the model across multiple GPU instances or devices to overcome memory limitations and improve inference latency.

B. Amazon SageMaker Feature Store

Amazon SageMaker Feature Store is a purpose-built service for machine learning that provides a centralized repository for features. It supports both online (low-latency) and offline (batch) access for training and inference, and includes capabilities for feature versioning and sharing, directly meeting all stated requirements.

A. Amazon SageMaker Feature Store

Amazon SageMaker Feature Store is a purpose-built service for storing, updating, and serving machine learning features for training and inference. It supports both online (low-latency) and offline (batch processing) access, ensuring consistency and reusability.

C. SageMaker Feature Store - Online and Offline Stores

Amazon SageMaker Feature Store is designed with both an Online Store and an Offline Store. The Online Store provides low-latency access to the latest feature values for real-time inference, while the Offline Store stores historical feature data for model training, batch inference, and analytical purposes. Using both modes simultaneously ensures all requirements are met.

B. Amazon SageMaker Clarify

Amazon SageMaker Clarify is specifically designed to help machine learning developers detect potential bias in their data and models, and to provide explainability for model predictions. It supports various bias metrics and explainability techniques (like SHAP and LIME) to ensure fairness and transparency, which are critical for sensitive applications like credit scoring and regulatory compliance.

C. Regularly retrain the model on periodically collected new data.

Concept drift occurs when the relationship between input features and the target variable changes over time. While Model Monitor can detect drift, the most direct and effective strategy to address concept drift is to regularly retrain the model on the most recently collected data that reflects the new underlying concepts. This ensures the model learns the current patterns and relationships.

B. Amazon SageMaker Pipelines

Amazon SageMaker Pipelines is a purpose-built MLOps service for building, automating, and managing end-to-end machine learning workflows. It allows data scientists and ML engineers to define a sequence of steps (like data processing, training, evaluation, registration, and conditional deployment), track lineage, and automate the entire ML lifecycle, directly addressing the need for a robust, orchestrated MLOps practice within SageMaker.

A. ContainerStartupHealthCheckTimeoutInSeconds

The `ContainerStartupHealthCheckTimeoutInSeconds` parameter defines the maximum time that SageMaker waits for a container to respond to health checks during startup and also sets the maximum duration for a single inference request. If the model initialization is slow or a single inference request takes longer than the default (typically 60 seconds), increasing this value can prevent timeouts under load or during deployment.

C. SageMaker Clarify for bias detection and explainability.

SageMaker Clarify is specifically designed to detect potential bias in ML models and provide explainability for their predictions, directly addressing the requirements for fairness and transparency across demographic groups and regulatory compliance.