Professional Notes: design the constraint you were given

Module 1: Serverless and real-time data paths

A serverless data pipeline starts from an event. An object landing in Amazon S3, a message on a stream, or an EventBridge rule starts the next step. Do not leave a server polling when the stem asks for the least infrastructure to operate.

Clean Rooms is the cross-account analysis path that does not hand the other party the underlying table. Copying the bucket does the opposite of that constraint.

Module 2: Durable workflows

AWS Step Functions standard workflows are the durable executions. They keep history, run for a long time, and fit a legal approval that might take days. Express workflows are the short, high-volume path. They are the wrong choice when the stem needs an audit of a week-long wait.

A task token pauses the execution until something calls back. That is the human-oversight pattern: the model or the agent proposes, the workflow waits, a person approves, and only then does the next state run. Polling a database from a Lambda loop is the design you replace. A callback is not the same thing as an Express workflow that must finish quickly.

Module 3: Service networking

Amazon VPC Lattice is application networking across VPCs and accounts. Services can be Lambda, Amazon ECS, Amazon EKS, or EC2, and the auth is IAM. Use it when the stem crosses an account boundary or mixes compute types.

Amazon ECS Service Connect is service discovery and connectivity for ECS services. It is built for that cluster, not for a multi-account mesh that also includes Lambda. A security group between two tasks is not a service identity.

Module 4: Container image security

Amazon ECR can scan an image on push. Enhanced scanning uses Amazon Inspector. The pipeline fails the deploy when the finding is critical and the policy says so. Disabling the scan is not remediation. AWS Signer signs the image. The deploy stage verifies the signature before it pulls. An unsigned image from a laptop is the supply-chain gap the stem is describing.

Module 5: Multi-tenant architectures

A noisy neighbor in a pool is moved toward a bridge or a silo for that tenant. Giving every tenant a dedicated account because one tenant is noisy spends the isolation budget on the tenants that did not ask for it.

Module 6: Generative AI in the design

Amazon Bedrock is the managed place to call a foundation model. A knowledge base retrieves from a data source you connect, then the model answers from that retrieval. That is the RAG path. Use it when the corpus changes and you are not training a model. Fine-tuning changes weights. It is the wrong first step for “read yesterday’s policy PDF.”

Amazon Bedrock AgentCore is the agent platform the announcement names: the agent runs, uses tools, and needs an identity. AgentCore Identity is how the agent gets access. A long-term access key in the prompt is the failure. The agent does not become the IAM administrator because it can call a tool.

Module 7: Controls on the model and the pipeline

Amazon Bedrock Guardrails filter content: denied topics, sensitive information, and word filters on the way in and the way out. A guardrail is not a knowledge base, and a bucket policy does not read the completion.

Human oversight is a workflow, not a hope. Step Functions waits on a task token until a person approves. The announcement lists that pattern with the agent work. Skipping the approval because the model was confident is the defect.

A deployment pipeline scans for vulnerabilities before the artifact is released. An image finding, a dependency finding, or an unsigned artifact stops the stage. Logging the finding and deploying anyway does not meet a stem that says “must not release.”

Module 8: The account boundary

The security domain is named security, compliance, and governance. The announcement does not reprint every IAM task. The ceiling is still the design: an identity policy can allow, a resource policy can allow, and both sides must allow a cross-account read. A service control policy sets the maximum for member accounts and does not grant. It does not constrain the management account the way it constrains members. A pipeline assumes a role into the target account. It does not store a long-term key for that account.

Extra The SAP-C02 guide put frontend mobile development, 12-factor methodology, and in-depth operating-system knowledge out of scope. SAP-C03 has not published a new out-of-scope list. Do not treat this paragraph as a SAP-C03 objective.

Module 9: Post-quantum cryptography

The announcement names AWS KMS with ML-DSA and ML-KEM. ML-DSA is a digital signature. ML-KEM is a key-encapsulation mechanism: it establishes a shared secret. Sign the artifact with ML-DSA. Encapsulate the key with ML-KEM. Swapping them fails the stem that asked for a signature or for a secret, because those are different jobs.

Module 10: Cost of the design

The cost domain is named. The task list is not published yet, so these are the purchase models a design has to tell apart. A Compute Savings Plan discounts steady EC2, Fargate, and Lambda usage and can move across instance families. An EC2 Instance Savings Plan is tighter: one family in one Region. A Reserved Instance is a commitment to a specific shape. Spot is spare capacity that can stop. Use it for workers that can be interrupted. Do not put the only copy of a stateful primary on Spot because it was cheaper on the spreadsheet.

On-Demand is the flexible rate with no interruption and no commitment. It is the right place for a spiky, unpredictable tier, and the expensive place to leave a flat production fleet.

Module 11: Storage class and data transfer

Amazon S3 Intelligent-Tiering is for objects whose access pattern you do not know. Glacier classes are for archives, and the retrieval time is part of the answer. A 15-minute restore requirement is not an archive class that takes hours. Lifecycle rules move objects and abort incomplete multipart uploads. They do not change the class of an object you must read on every request.

A NAT gateway charges for data processing. A gateway VPC endpoint for S3 or DynamoDB takes that traffic off the NAT path. An interface endpoint is an elastic network interface with an hourly charge and a security group. Cross-AZ traffic is a design cost. Putting both sides of a chatty pair in one AZ removes that charge and also removes the AZ failure story, so the stem has to say which constraint wins.

Module 12: Resilience engineering

AWS Resilience Hub assesses an application against a resilience target and recommends the alarms, SOPs, and tests. It does not inject the failure. AWS Fault Injection Service runs the experiment: stop instances, disrupt the network, or apply the action you defined, inside the guardrails of the experiment. A production experiment without a stop condition is not the service. It is an outage you scheduled.

AWS Application Recovery Controller is the control plane for a recovery: readiness checks tell you whether the other side can take traffic, and routing controls shift it. Zonal shift moves traffic off one Availability Zone. Multi-AZ database failover keeps one database endpoint inside the Region. It is not a zonal shift of the whole stack, and it is not a Regional disaster plan.

Automated recovery steps live in an AWS Systems Manager runbook. The alarm or the recovery control starts the runbook. A wiki page is not the automation the announcement named.

Module 13: Business continuity patterns

RPO is how much data you can lose. RTO is how long you can be down. Write both numbers before you pick a pattern.

Module 14: Migration

The domain name includes migration. Match the thing that moves. AWS Application Migration Service continuously replicates a server for a lift-and-shift cutover. AWS Database Migration Service moves a database and can keep replicating until you cut over. AWS DataSync moves files and objects. AWS Snowball moves data that should not cross the network. A 2 GB database that must stay in sync is not a Snowball job. A petabyte file share with a short network window is not a laptop upload.

Module 15: Multi-account pipelines and runbooks

A multi-account pipeline lives in a deployment account and assumes a role in each target account. The role is the credential. A static key in the pipeline variable store is the pattern to remove. The same pipeline is where vulnerability scanning runs, before the deploy stage.

Systems Manager Automation is the runbook the resilience section starts. State Manager keeps a configuration true. Session Manager is a shell without port 22. Those are different jobs. Do not open SSH because the runbook failed. Fix the role the runbook uses.

Module 16: Signals the design has to include

The announcement adds container monitoring, AI and ML metrics, synthetic monitoring, and real-user monitoring. Amazon CloudWatch Container Insights is the container metric path for ECS and EKS. Model latency and token or invocation errors are application metrics you publish. They are not an EC2 CPU graph.

CloudWatch Synthetics runs a canary on a schedule so you learn the endpoint failed before a customer does. CloudWatch RUM collects what the browser actually saw. A canary is not a real user, and a real-user beacon is not a scheduled probe of a login you do not want to wait for. VPC Flow Logs still tell you accept or reject. They do not explain a slow model response.

Frequently asked questions