AIAI Agent Dev Labaiagentdev-lab.com
Sign in to save progressSign in
Menu

SAP-C03 · LESSON 1

Architecture Design and Implementation Built for the Cloud25% · 4 topics

Separate throughput, latency, data classification and tenant boundaries. Design RAG ingestion, index refresh, authorized retrieval and response paths, plus inference capacity and failure behavior. Connect AI to existing messaging, workflow and container choices; model-training implementation is a separate concern.

1.1Integrate AI into an existing architecture

Concept

Separate throughput, latency, data classification and tenant boundaries. Design RAG ingestion, index refresh, authorized retrieval and response paths, plus inference capacity and failure behavior. Connect AI to existing messaging, workflow and container choices; model-training implementation is a separate concern.

How it works / steps

  1. Write requirements and success criteria first.
  2. Use a test environment and retain success and failure evidence.
  3. Review the result, then check your decisions with domain practice below.

Workplace example

Use the same query for two tenants and verify documents never cross their boundaries.

Common confusions and exam tips

Mini quiz

Q1. Lambda processes order messages from an SQS standard queue. Occasionally the same order is registered twice. What is the most appropriate measure?

  1. Increase Lambda concurrency to the maximum so all messages are processed at once in one invocation
  2. Delete the SQS standard queue, accept orders by email and have staff register them manually
  3. Record processed order IDs in DynamoDB so the same message received twice is not processed twice
  4. Set message retention to the minimum of 1 minute so old messages disappear before being redelivered
Show answer and explanation

Answer: C Record processed order IDs in DynamoDB so the same message received twice is not processed twice

SQS standard queues provide at-least-once delivery, with occasional duplicates. Make processing idempotent (record processed IDs) or use FIFO queue deduplication.

Q2. Which statement about AWS Step Functions is incorrect?

  1. Workflows (Standard Workflows) can run for only up to 15 minutes
  2. Retries and error handling for each step can be specified in the state machine definition
  3. It can directly call many AWS services such as Lambda, ECS and DynamoDB
  4. It can wait for external processing such as human approval using task tokens
Show answer and explanation

Answer: A Workflows (Standard Workflows) can run for only up to 15 minutes

Standard Workflows can run for up to one year. The 15-minute limit is the maximum duration of a single Lambda function invocation.

Reference links

1.2Design a network connectivity strategy

Concept

Large organizations connect many VPCs, VPNs and Direct Connect with Transit Gateway as the hub, and separate paths with route tables (splitting production and development, routing through an inspection VPC and so on). For name resolution with on-premises, use Route 53 Resolver inbound and outbound endpoints; to expose services, use PrivateLink (which works even with overlapping CIDRs). Follow the Direct Connect resiliency recommendations (multiple locations, multiple connections).

How it works / steps

  1. Plan VPC CIDRs so they do not overlap (use PrivateLink where they do)
  2. Separate paths per environment with Transit Gateway route tables
  3. Make Direct Connect redundant across two locations and add Site-to-Site VPN as a backup
  4. Resolve names in both directions with on-premises using Route 53 Resolver endpoints and rules
  5. Share common networking with other accounts through RAM (Resource Access Manager)

Workplace example

Consolidate 200 VPCs on Transit Gateway and send all outbound traffic through Network Firewall in an inspection VPC. Resolve the internal on-premises domain with outbound endpoint forwarding rules.

Common confusions and exam tips

Mini quiz

Q1. You want to send a large number of files from on-premises servers to S3 over Direct Connect without going through the internet. Which configuration is most appropriate?

  1. Create an S3 gateway VPC endpoint and point to it from the on-premises router's route table
  2. Create a NAT gateway in the VPC and send on-premises traffic over Direct Connect to the NAT gateway
  3. Create an S3 interface VPC endpoint and connect to it from on-premises over Direct Connect
  4. Assign public IPs to the on-premises servers and send to the S3 public endpoint over the internet
Show answer and explanation

Answer: C Create an S3 interface VPC endpoint and connect to it from on-premises over Direct Connect

Gateway endpoints can only be used through route tables inside the VPC and cannot be reached from on-premises. To use one from on-premises, use an interface endpoint with private IPs over Direct Connect or VPN.

Q2. VPC A and VPC B are attached to a Transit Gateway, and routes for both are propagated to the Transit Gateway route table. However, EC2 in A cannot reach EC2 in B. What is the most appropriate thing to check next?

  1. Whether appliance mode is enabled on the Transit Gateway attachments
  2. Whether static routes are also added to the Transit Gateway route table
  3. Whether DNS resolution and DNS hostnames are enabled in VPCs A and B
  4. Whether A's and B's subnet route tables route the peer CIDR to the Transit Gateway
Show answer and explanation

Answer: D Whether A's and B's subnet route tables route the peer CIDR to the Transit Gateway

Besides routes on the Transit Gateway, each VPC's subnet route tables also need a route sending the other's CIDR to the Transit Gateway. Then check security groups and network ACLs. [Other options] Appliance mode is for inspection appliance paths. Routes are already propagated, so static routes aren't needed. DNS settings don't affect traffic by IP address.

Reference links

1.3Design a strategy to meet reliability requirements

Concept

Reliability design combines shrinking the blast radius (cell-based architecture, shuffle sharding), loosely coupling dependencies (SQS, EventBridge), retries (exponential backoff and jitter) with timeouts and circuit breakers, and monitoring and raising quotas (Service Quotas).

How it works / steps

  1. Identify single points of failure and make them redundant across multiple AZs
  2. Add timeouts and asynchrony so a slow dependency does not stop everything
  3. Retry with exponential backoff and jitter, and make operations idempotent
  4. Alarm when approaching Service Quotas limits and raise them in advance
  5. Remove unhealthy resources automatically with health checks

Workplace example

When an external payment API slows down, timeouts and a circuit breaker keep order intake running, and payments are queued in SQS for later processing.

Common confusions and exam tips

Mini quiz

Q1. You are choosing how to integrate applications. You want to decouple order intake from processing and send messages that fail processing to a dead-letter queue after retries. Which service is the most suitable?

  1. AWS Step Functions
  2. Amazon SQS
  3. Amazon SNS
  4. Amazon EventBridge
Show answer and explanation

Answer: B Amazon SQS

SQS holds messages between sender and receiver and can move messages that repeatedly fail processing to a dead-letter queue.

Reference links

1.4Design a solution to meet performance objectives

Concept

Meet performance objectives (latency, throughput) with caching (CloudFront, ElastiCache, DAX, API Gateway caching), database choice (purpose-built DBs), compute choice (instance types, GPUs, Graviton), asynchrony and global placement (edge, multiple Regions). Confirm the objectives with load testing.

How it works / steps

  1. Define performance objectives numerically (such as p99 latency)
  2. Choose purpose-built DBs that fit the access patterns
  3. Cache repeatedly read data
  4. Make heavy processing asynchronous to speed up responses
  5. Confirm objectives are met with load testing

Workplace example

To respond to users worldwide within 200 ms at p99, serve static content from CloudFront and API reads from each Region with DynamoDB global tables and DAX.

Common confusions and exam tips

Mini quiz

Q1. You are choosing storage that meets performance requirements. You need the fastest possible space for a temporary cache during processing. Data may be lost if the instance stops. Which is the most suitable?

  1. Amazon FSx for Lustre
  2. Amazon EFS
  3. EBS io2 volume
  4. Instance store
Show answer and explanation

Answer: D Instance store

Instance store is disk attached directly to the host and is fast, but data is lost on stop or termination, so use it for temporary purposes.

Reference links

Lessons are written by this site based on public information and do not guarantee exam content. Product specifications change, so check the latest information at the reference links.

Notes (unofficial, original questions, no pass guarantee)