Automate orchestration and delivery for ML and AI workloads

• Shadow testing evaluates live requests without returning candidate responses to users
• Production variant weights control the traffic used for live model comparisons
• Explicit variant targeting overrides weighted request distribution
• Blue/green deployment retains the old fleet while the new fleet is validated
• Canary and linear deployments differ in how they increase traffic exposure
• Deployment alarms connect validation failures to automatic rollback
• Rolling deployment updates serving capacity in configurable batches
• SageMaker AI deployment guardrails apply only to supported endpoint configurations
• A manual approval action pauses a release until an authorized decision is recorded
• CodePipeline coordinates release stages while actions perform their own work
• A CodeBuild buildspec defines both execution commands and build outputs
• CodeBuild can obtain sensitive values from managed secret stores
• A repository connection must complete provider authorization before it can trigger releases
• A full-clone source reference is intended for downstream CodeBuild actions
• Git source control and S3 object versioning serve different collaboration needs
• CodeDeploy handles application platforms rather than SageMaker AI endpoint updates
• A CodeDeploy AppSpec identifies deployment resources and lifecycle validation hooks
• Unit tests and integration tests catch different defects in an ML release
• SageMaker AI pipeline steps separate processing, training, and batch inference
• A condition step turns evaluation results into a pipeline decision
• EventBridge can start a model-building pipeline when relevant data changes
• Scheduled retraining uses a separate trigger that must be managed with the pipeline
• Pipeline step caching reuses successful matching work within a configured window
• Callback steps wait for external work to report completion
• Batch transform distributes S3 objects rather than splitting one object across instances
• Batch input splitting and output assembly preserve file boundaries
• Model Groups organize related model versions in a production catalog
• Model approval status can gate an explicitly configured deployment workflow
• Cross-account model deployment needs access to the Model Group and its underlying artifacts
• MLflow connects model candidates to the experiments that produced them
• Managed MLflow separates service-managed metadata from customer S3 artifacts
• Prompt variables reuse a tested template with different runtime inputs
• Prompt variants compare alternative messages and inference configurations
• Published prompt versions preserve tested configurations for deployment
• Custom prompt datasets make automated model tests specific to the application
• Judge-model and human evaluations provide different forms of release evidence
• RAG testing must assess retrieval as well as the generated response
• AgentCore Evaluations scores instrumented agent behavior rather than only final text
• MLflow traces connect unexpected AI outputs to the components that produced them
• Updating an AgentCore Runtime creates a new immutable configuration version
• Named AgentCore endpoints support controlled promotion independently of the default endpoint
• Flow aliases let applications change workflow versions without changing the invocation target
• Provisioned custom-model updates must preserve the supported base-model relationship
• A successful update request does not mean the new custom model is already serving
• Knowledge-base source changes require a sync to update the retrieval index
• Some metadata-only refreshes can reuse existing document embeddings
• Refresh orchestration must observe ingestion completion and document results
• Newly ingested vectors may need time to become queryable
• Bedrock agent aliases select immutable versions for application invocation

Unlock with Premium — includes all practice exams and the complete study guide.