AWSAWS Certified Generative AI Developer - Professional2026-100 Prerequisites

Assign infrastructure to each application boundary

Completed by the knowledge check

Learning objective

Understand VPC, endpoints, compute, storage, queues, and events used by GenAI applications.

Core idea

A model invocation is one step in an application. Something runs the caller, something stores its inputs and outputs, a network path connects it to services, and delivery mechanisms carry work between components. Distinguish those roles before choosing infrastructure. A private network path does not execute code; a queue does not preserve an authoritative document revision; an event describing completion is not the completed result itself.

Read an architecture by asking five questions: where does code run, what durable data does it use, how does it reach a service, who owns unfinished work, and which consumers need to learn about a change? These questions expose missing boundaries that a diagram containing only an application and a model conceals.

How it works

Locate the network path

Amazon Virtual Private Cloud (Amazon VPC) provides a logically isolated virtual network. A subnet is an IP address range within that VPC and resides in one Availability Zone. Route tables determine where traffic from subnets or gateways is directed. These are network resources, not model hosting resources.

A VPC endpoint provides private connectivity to a supported AWS service. For Amazon Bedrock, AWS PrivateLink interface endpoints create endpoint network interfaces in the selected subnets. The private connection does not require an internet gateway or NAT device for that Bedrock path, and caller instances do not need public IP addresses to use it.

Choose the endpoint service for the API you call. For AWS-native model inference, the service name is com.amazonaws.REGION.bedrock-runtime, where REGION is the AWS Region. com.amazonaws.REGION.bedrock serves the Bedrock control plane. A control-plane endpoint does not substitute for a runtime endpoint. With private DNS enabled, the service’s default Regional hostname can resolve through the interface endpoint.

The endpoint network interface’s security group must allow the expected traffic, such as inbound HTTPS from the caller. Private DNS also requires VPC DNS hostnames and DNS resolution. Endpoint policies constrain principals, actions, and resources through that endpoint; they do not replace the caller’s IAM authorization. Detailed private-path policy design belongs to 3.2.1.

Assign compute to the work unit

Lambda functions execute a handler in response to events or API calls while Lambda manages the servers and execution environments. They suit a bounded request adapter or event-driven worker. A normal function invocation has a maximum duration of 15 minutes. That limit matters when you place preprocessing, model waits, and output validation in one invocation; see 2.2.4 for invocation budgets.

An Amazon EC2 instance is a virtual server. Its instance type determines the available hardware and its balance of compute, memory, networking, and storage. Choose this level of control when the application’s process or hardware contract requires it, and account for the infrastructure you must operate. An EC2 instance running caller code does not imply that Bedrock inference runs on that instance.

Compute lifetime is not data lifetime: preserve documents and accepted jobs outside the running process. Broader hosting choices belong to 2.2.6.

Identify durable data separately

Amazon Simple Storage Service (Amazon S3) stores objects in buckets. A bucket, object key, and, when applicable, version ID identify an object. For this application, a general purpose bucket can hold source documents and generated result artifacts. S3 Versioning preserves multiple versions so an overwrite need not erase the previous object.

Store a reference to an exact input revision when reproducibility matters. A job that records only a mutable key can consume different bytes after an overwrite. Versioning provides the storage mechanism; your application must preserve and use the selected version identity. Replication and deletion policies remain separate decisions; see 1.2.11 and 3.2.10.

Do not confuse object storage with every application storage need. A result file, a job-status record, and a retrieval index serve different access patterns. S3 holding the original document does not establish that a derived index is fresh; the RAG stage boundaries in 0.9 still apply.

Give unfinished work an owner

Amazon Simple Queue Service (Amazon SQS) decouples producers from workers. A producer uses SendMessage; a worker uses ReceiveMessage; successful processing eventually leads to DeleteMessage. Receiving work is not the same as completing it.

VisibilityTimeout hides a received message from subsequent receives for a period. If the message is not deleted or its visibility extended before that period ends, it can become visible again. The queue’s default visibility timeout is 30 seconds. Choose the configured value from processing duration and recovery requirements rather than assuming this default covers a model call.

Standard queues provide at-least-once delivery and best-effort ordering. Duplicate delivery therefore belongs in your application contract. FIFO queues provide ordering within a message group, not a universal ordering of every application’s work. Neither choice makes an external business effect safe to repeat by itself. Detailed durable job and retry handling belongs to 2.4.2.

Route events to interested consumers

An event communicates that something happened, such as acceptance of a document revision or completion of a result. Producers can publish with PutEvents. Amazon EventBridge Custom Event Bus routes events to subscribers and retains events for a configured period. A subscriber filters events, can transform them, and delivers each matching event to one target. Fan-out uses multiple subscribers.

CreateSubscriber defines this resource. FilterConfiguration.Filters contains filters with a Pattern and Scope; an event must match every filter in the list. A subscriber without FilterConfiguration receives every event on its bus. Use these mechanisms to deliver completion events to the consumers that need them.

AWS distinguishes Custom Event Bus from Custom Event Bus - Classic, which uses rules and targets; both remain available. Avoid mixing subscriber terminology with a Classic rule design. Transport selection and replay details belong to 2.3.4.

Key distinctions and trade-offs

Endpoint versus compute: A VPC endpoint is a network entry point to a service. It is not the process executing your application or a customer-deployed model. SageMaker AI hosting endpoints are a different resource concept; see 0.16.

Queue versus event routing: A queue holds work for consumption. An event bus routes notifications to interested consumers. Combining them can give each consumer a work backlog without making the bus the authoritative job record.

Private path versus permission: DNS resolution and allowed network traffic establish reachability. IAM and applicable endpoint policies still constrain the request. A private connection is not blanket authorization.

Data durability versus processing success: A stored document and a delivered message prove neither that inference completed nor that its output passed validation. Preserve an explicit application outcome.

Architecture / decision scenario

A document-summary service accepts a file into a versioned S3 bucket and records its bucket, key, and version ID. It places a job reference on SQS. A Lambda function processes a bounded job; a separate EC2 worker is an option for work whose process requirements do not fit that function boundary.

For a worker deployed in the VPC, the design provides the Bedrock runtime interface endpoint, working private DNS, and the expected HTTPS connectivity. The runtime role independently permits the approved model request. An S3 gateway endpoint is a route-table target that provides S3 access without an internet gateway or NAT; S3 also supports interface endpoints. The SQS path uses its own interface VPC endpoint. The Bedrock endpoint covers neither service.

The EC2 worker calls DeleteMessage after processing and ChangeMessageVisibility when it must extend visibility. For Lambda, an SQS event source mapping polls the queue and deletes messages after successful batch processing; configure the queue visibility timeout to at least six times the function timeout, plus MaximumBatchingWindowInSeconds when using a batch window. Both consumers validate and store the summary before reporting success; see 2.4.2 for durable job handling. A completion event on Custom Event Bus reaches separate notification and audit subscribers through their filters. The input object, pending work, network path, compute process, and completion notification each retain a distinct owner.

Each hop from the uploaded file to its subscribers hands the work to a different owner.

Common exam traps

Treating a private subnet as automatic connectivity to every AWS service.
Creating a Bedrock control-plane endpoint for runtime inference calls.
Treating ReceiveMessage as successful job completion.
Assuming FIFO ordering makes every external effect exactly once.
Recording a mutable S3 key when an exact revision is required.
Mixing EventBridge subscribers with Classic rules and targets.

What to remember

Assign a mechanism to each boundary: VPC and endpoints for connectivity, compute for execution, storage for durable artifacts, SQS for unfinished work, and EventBridge for event distribution. Verify permissions and accepted processing outcomes independently of network location and message delivery.