Overview
Deployment and Orchestration of ML and AI Workflows has three task statements. Task 3.1 is about deployment infrastructure: compute targets, multi-model and multi-container endpoints, real-time versus batch inference, FM deployment options, Bedrock Custom Model Import, agent deployment with communication protocols and RAG retrieval settings such as reranking.
Task 3.2 covers provisioning: on-demand versus provisioned capacity, infrastructure as code, containers, endpoints inside a VPC, programmatic deployment with the SageMaker Python SDK or Boto3, auto scaling metrics, Bedrock knowledge bases, agent state and GPU scaling. Task 3.3 covers MLOps automation: deployment strategies and rollback, CodePipeline and related services, retraining triggers, Model Registry and MLflow versioning, Prompt Management, agent pipelines and knowledge base refresh.
How it's tested
Domain 3 scenarios give traffic patterns, payload sizes, latency targets, network rules and release requirements. You pick the inference option, endpoint configuration, scaling metric or pipeline design. AWS favors managed, automated and reversible deployments, so answers with built-in rollback, registries and infrastructure as code usually beat manual or one-off scripts.
-
Pick the inference mode
Traffic is steady, spiky, large-payload or offline, and you match it to real-time, serverless, asynchronous or batch inference on SageMaker AI or Bedrock.
-
Deploy GenAI components
An FM, agent or knowledge base needs hosting, and you choose Bedrock on-demand, provisioned throughput, Custom Model Import, AgentCore or a SageMaker AI endpoint.
-
Automate safe releases
A team wants repeatable retraining and deployment, and you design Pipelines, Model Registry approval, CodePipeline and blue/green or canary rollback.
A model scores 20 million records each night and results are needed by morning. A real-time endpoint would sit idle most of the day and need many instances during the run. Batch transform starts compute for the job, reads input from S3, writes predictions back and shuts down. Asynchronous inference fits large single requests with queueing, not scheduled bulk scoring.
Task statements 3
Official task statements from the Amazon Web Services outline, in outline order.
3.1 Manage deployment infrastructure for ML and AI model typesCompute targets, orchestrators, multi-model hosting, inference strategies, FM options, external models, agents and RAG settings.
- What it means
- This task statement covers selecting compute environments and deployment targets, orchestrators and multi-model or multi-container strategies, real-time and batch inference, FM deployment options, importing models built outside AWS through SageMaker AI or Bedrock Custom Model Import, deploying agents with tools and communication protocols, configuring FM hosting and resources, and applying RAG retrieval strategies such as reranking.
- What to understand
- SageMaker AI offers real-time endpoints for low latency, serverless inference for intermittent traffic with cold starts, asynchronous inference for large payloads and long processing, and batch transform for offline datasets. Multi-model endpoints share one container across many models; multi-container endpoints run different containers behind one endpoint; inference components allocate resources per model. Bedrock serves FMs on demand, through batch inference or with provisioned throughput, and Custom Model Import brings supported open-weight models. Agents can run on Bedrock AgentCore Runtime and use protocols such as Model Context Protocol. Reranking reorders retrieved chunks to improve relevance.
- How it shows up
- A model type, traffic pattern and latency need are given, and you choose the hosting option, endpoint type or FM deployment path.
- Common confusion
- Serverless inference suits idle periods but has cold starts; asynchronous inference queues large requests and can scale to zero.
3.2 Provision and configure resources for ML and AI workloads based on existing architecture and requirementsCapacity choices, IaC, containers, VPC endpoints, SDK deployment, auto scaling, knowledge bases, agent state and GPU scaling.
- What it means
- This task statement covers balancing on-demand and provisioned resources, automating provisioning with stacks and orchestration, building containers, running SageMaker AI endpoints in a VPC, deploying with the SageMaker Python SDK, AWS CLI or Boto3, choosing auto scaling metrics, managing Bedrock knowledge bases with vector stores and indexing, implementing retrieval pipelines and agent state, scaling GPU workloads and deploying agentic workflow infrastructure.
- What to understand
- AWS CloudFormation and AWS CDK define endpoints, roles and networking as code, and stack outputs pass values between stacks. Containers for SageMaker AI are stored in Amazon ECR. VPC-only endpoints use private subnets, security groups and interface VPC endpoints for SageMaker API and runtime access, with network isolation when no outbound access is allowed. Target tracking on InvocationsPerInstance is the usual scaling metric, while GPU or concurrency metrics suit heavier models. Bedrock knowledge bases connect an S3 data source to a vector store and sync on demand or on schedule. AgentCore Memory keeps short-term and long-term agent state.
- How it shows up
- An existing architecture and requirement are described, and you choose how to provision, network, scale or configure the ML or AI resource.
- Common confusion
- CPUUtilization is often a poor scaling signal for inference; invocations per instance or latency track load better.
3.3 Implement automated orchestration and continuous integration and continuous delivery (CI/CD) pipelines for MLOps and AI workloadsDeployment strategies, CI/CD services, jobs, testing, retraining, versioning, prompts, agents and knowledge base refresh.
- What it means
- This task statement covers automated deployment and rollback, configuring CodeBuild, CodeCommit, CodeDeploy, CodePipeline and CodeConnections, setting up training and inference jobs, automated testing for ML and AI, retraining mechanisms, model versioning with Model Registry and MLflow, Bedrock Prompt Management, agent deployment pipelines and versions, prompt testing frameworks, fine-tuned FM versioning and orchestration of RAG updates and knowledge base refresh.
- What to understand
- SageMaker Pipelines chains processing, training, evaluation, a condition step and model registration. Model Registry holds versioned model packages with approval status, and an approval event can trigger CodePipeline to deploy. SageMaker AI deployment guardrails support blue/green with all-at-once, canary or linear traffic shifting and automatic rollback on CloudWatch alarms. Retraining can be triggered on a schedule or by EventBridge rules from Model Monitor alerts. Prompt Management versions prompts so applications reference a fixed version. CodeConnections links GitHub or GitLab repositories. Knowledge base sync jobs can be scheduled after new documents land in S3.
- How it shows up
- A team needs repeatable, tested and reversible releases, and you choose the pipeline step, registry feature, deployment strategy or trigger.
- Common confusion
- SageMaker Pipelines orchestrates ML steps; CodePipeline orchestrates code and release stages; many designs use both.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Real-time inference
Persistent endpoint for low-latency requests.
- Remember
- Pay while instances run.
- Tested as
- Inference mode questions.
- Trap
- Using it for nightly bulk scoring.
-
Serverless inference
Endpoint that scales with requests, including to zero.
- Remember
- Cold starts; payload and memory limits.
- Tested as
- Intermittent traffic questions.
- Trap
- Choosing it for strict low latency.
-
Asynchronous inference
Queues requests and writes results to S3.
- Remember
- Large payloads, long processing, scale to zero.
- Tested as
- Long-running request questions.
- Trap
- Confusing it with batch transform.
-
Batch transform
Offline inference over a dataset in S3.
- Remember
- No persistent endpoint.
- Tested as
- Scheduled scoring questions.
- Trap
- Keeping an endpoint running for it.
-
Multi-model endpoints
Many models behind one endpoint sharing a container.
- Remember
- Loads models on demand to save cost.
- Tested as
- Many-small-models questions.
- Trap
- Using it when models need different frameworks.
-
Bedrock provisioned throughput
Reserved model units for consistent capacity.
- Remember
- Required for some custom models.
- Tested as
- FM capacity questions.
- Trap
- Buying it for light, irregular traffic.
-
Bedrock Custom Model Import
Brings supported open-weight models into Bedrock.
- Remember
- Serverless hosting for external models.
- Tested as
- External model questions.
- Trap
- Rebuilding a model already trained elsewhere.
-
Amazon Bedrock AgentCore
Services to run, connect and observe agents: Runtime, Gateway, Memory, Identity, Observability.
- Remember
- Framework-agnostic agent hosting.
- Tested as
- Agent deployment questions.
- Trap
- Building custom session state stores.
-
SageMaker Model Registry
Versioned model packages with approval status.
- Remember
- Approval can trigger deployment.
- Tested as
- Versioning and audit questions.
- Trap
- Deploying unregistered artifacts.
-
Deployment guardrails
Blue/green with canary or linear shifting and auto rollback.
- Remember
- Rollback on CloudWatch alarms.
- Tested as
- Safe release questions.
- Trap
- All-at-once with no alarms.
-
Bedrock Prompt Management
Stores and versions prompts.
- Remember
- Apps call a fixed prompt version.
- Tested as
- Prompt versioning questions.
- Trap
- Hard-coding prompts in application code.
-
Auto scaling metrics
Target tracking on SageMakerVariantInvocationsPerInstance and similar.
- Remember
- Match the metric to the bottleneck.
- Tested as
- Scaling questions.
- Trap
- Scaling inference on CPU alone.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
SageMaker AI inference options
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Real-time | Always-on endpoint | Lowest latency, steady traffic | Paying for idle time |
| Serverless | On-demand compute per request | Spiky or rare traffic, cold starts | Large GPU models |
| Asynchronous | Queued requests, results to S3 | Large payloads, long jobs | Treating it as offline batch |
| Batch transform | Job over a dataset | No endpoint, scheduled scoring | Expecting per-request responses |
Bedrock FM deployment choices
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| On-demand | Pay per token | Variable or low traffic | Expecting reserved capacity |
| Batch inference | Asynchronous jobs over S3 prompts | Lower cost for bulk work | Using it for chat |
| Provisioned throughput | Reserved model units | Steady high volume or custom models | Buying it for tests |
| Custom Model Import | Host your own open-weight model | Bring models trained elsewhere | Assuming any architecture is supported |
Pipeline and release tools
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| SageMaker Pipelines | ML workflow DAG | Processing, training, evaluation, registration | Using it to build application code |
| CodePipeline | Release orchestration | Source, build, test and deploy stages | Expecting it to run training natively |
| Model Registry | Model versions and approvals | Gate between training and deployment | Skipping approval status |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Always real-time
- What candidates pick
- Deploy every model to a real-time endpoint.
- Why it looks right
- It is the most familiar SageMaker AI option.
- Why it's wrong
- Real-time endpoints bill while idle; batch transform, asynchronous or serverless inference fit offline, large-payload or sporadic traffic at lower cost.
Clue to watch for Words such as nightly, occasional or large file.
-
Async equals batch
- What candidates pick
- Use asynchronous inference to score a full dataset each night.
- Why it looks right
- Both run without an immediate response.
- Why it's wrong
- Asynchronous inference handles individual queued requests on an endpoint; batch transform is built for whole datasets without a persistent endpoint.
Clue to watch for A complete dataset in S3 on a schedule.
-
Public endpoint with tight rules
- What candidates pick
- Rely on IAM alone when the endpoint must not reach the internet.
- Why it looks right
- IAM controls who can invoke.
- Why it's wrong
- Network isolation needs VPC configuration with private subnets, security groups and interface endpoints; IAM does not control network paths.
Clue to watch for No internet access allowed.
-
Provisioned throughput for pilots
- What candidates pick
- Buy Bedrock provisioned throughput for a small internal pilot.
- Why it looks right
- Reserved capacity sounds reliable.
- Why it's wrong
- Provisioned throughput is a commitment for steady, high-volume or custom-model use; on-demand fits low and unpredictable pilot traffic.
Clue to watch for Low or uncertain request volume.
-
CodePipeline trains models
- What candidates pick
- Put training logic directly in CodeBuild steps.
- Why it looks right
- CodePipeline already runs the release.
- Why it's wrong
- Training belongs in SageMaker training jobs or Pipelines with tracked lineage; CodePipeline should trigger and deploy, not replace ML orchestration.
Clue to watch for Requirements mention lineage or approval.
-
Manual knowledge base refresh
- What candidates pick
- Ask an engineer to resync the knowledge base when documents change.
- Why it looks right
- It works for a demo.
- Why it's wrong
- Manual sync is error-prone; schedule or event-trigger ingestion jobs when new objects land in S3.
Clue to watch for Documents change regularly.
Answer framework
Use these steps for Domain 3 questions about deployment, provisioning and MLOps automation.
- Read the traffic pattern, payload size, processing time and latency target, then choose real-time, serverless, asynchronous or batch inference on SageMaker AI.
- For foundation models, choose Bedrock on-demand, batch, provisioned throughput or Custom Model Import, or a SageMaker AI endpoint for full control.
- Apply the stated constraints, such as VPC isolation, GPU scaling, multi-model hosting, agent runtime and memory, or knowledge base vector store and sync settings.
- Automate with SageMaker Pipelines and Model Registry approval, then deploy through CodePipeline or EventBridge triggers defined as code.
- Make every release reversible with blue/green, canary or linear shifting, CloudWatch alarms, versioned prompts and versioned agents.
Practice question
Want more like this? The MLA-C02 practice test has questions for every domain.
Checklist
0 of 14
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Begin with the inference decision, since it shows up in many forms. Write each SageMaker AI option and each Bedrock option on a card with its traffic pattern, latency, cost model and one limit, then quiz yourself on short scenarios until the choice is automatic. Add multi-model endpoints and inference components to the same cards.
Next, deploy something end to end. Train a small model, register it in Model Registry, deploy it with the SageMaker Python SDK, add a target-tracking scaling policy and then redeploy a new version with canary traffic shifting and a CloudWatch alarm. Repeat the deployment in a VPC with private subnets to see which endpoints and permissions are needed.
Then cover the GenAI and agent pieces. Create a Bedrock knowledge base, schedule a sync, and save a prompt version in Prompt Management. Read the AgentCore documentation for Runtime, Memory and Gateway so you can explain where agent state lives and how tools connect.
Close by sketching a full MLOps pipeline: source in a Git repository through CodeConnections, CodeBuild tests, SageMaker Pipelines training and evaluation, registry approval and automated deployment with rollback.
FAQ
How much of MLA-C02 is Domain 3?
Deployment and Orchestration of ML and AI Workflows is 24% of scored content, about twelve of fifty scored questions, across three task statements. It rose from 22% in MLA-C01.
What is the difference between asynchronous inference and batch transform?
Asynchronous inference queues individual requests on an endpoint and writes each result to S3. Batch transform runs a job over an entire dataset without a persistent endpoint.
When should I use a multi-model endpoint?
Use one when you host many models that share a framework and see low or uneven traffic, so they can share instances and load from S3 on demand.
Does MLA-C02 cover AI agents?
Yes. Task statements cover deploying agents, agent communication protocols, agent state management, agentic workflow infrastructure and automated agent deployment pipelines with version management.
Which CI/CD services are named in the exam guide?
AWS CodeBuild, CodeCommit, CodeDeploy, CodePipeline and CodeConnections, alongside SageMaker Model Registry, MLflow on SageMaker AI and Bedrock Prompt Management.
Were SageMaker Neo and bring your own container removed?
The comparison page lists edge optimization with SageMaker Neo and the BYOC skill as deletions, although building and maintaining containers for ML and AI workloads remains a skill.
How should I scale a SageMaker AI endpoint?
Use target tracking on a load metric such as invocations per instance, or GPU and concurrency metrics for heavy models, with cooldowns tuned to traffic.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 3 knowledge
You have covered inference options, FM and agent deployment, provisioning, auto scaling and MLOps pipelines. Check whether you can choose the right deployment for each case.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the MLA-C02 AI tutor Get any concept explained your way