AWS Certified Solutions Architect – Professional · SAP-C03 · exam live from Nov 17, 2026
SAP-C03 Study Guide
For the AWS (Amazon Web Services) Certified Solutions Architect – Professional exam. Every task and skill from the current official SAP-C03 exam guide, explained simply, with the design trade-offs the questions test. Pick a domain tab or search any service. Yellow boxes are memory hooks; dashed boxes are exam tips.
🔍/
No matches. Try a shorter word, a service name, or an acronym.
Start here
How the exam works
Format, scoring, dates, what changed from SAP-C02, and what's in scope.
The exam at a glance
This is a professional-level exam. Questions are long business scenarios with several constraints. Usually two or more answers would work; you must pick the one that meets every constraint best (cheapest, least operational overhead, most resilient, fastest to deliver). You will not write code.
75questions (65 scored + 10 unscored)
180minutes, so about 2.4 minutes per question
720passing scaled score (out of 1,000)
0penalty for guessing, so answer everything
$300USD (50% off if you hold an active AWS cert)
3 yrscertification validity
Question types
Type
What you do
Multiple choice
Pick 1 correct answer out of 4.
Multiple response
Pick 2 or more correct answers out of 5 or more. All must be right to get credit.
Scoring is compensatory: you only need to pass the exam overall, not each domain. Unscored questions are not marked, so treat every question as real. Delivered at a Pearson VUE test center or online proctored.
The passing score dropped from 750 (SAP-C02) to 720 (SAP-C03). The question count and time are unchanged. The exam uses short service names; a list of short and full names is under the Help button during the exam.
Read the last sentence first. It holds the real question ("MOST cost-effective", "LEAST operational overhead"). Then read the scenario hunting for the constraints that kill wrong answers.
Key dates: SAP-C02 → SAP-C03
Date
What happens
October 27, 2026
SAP-C03 registration opens (all languages); full exam guide published.
November 16, 2026
Last day to take SAP-C02.
November 17, 2026
SAP-C03 general availability (GA) delivery begins.
Most practice exams and courses on the market were written for SAP-C02. They still cover a lot of SAP-C03 (networking, DR, migration, cost), but they miss the new AI, post-quantum, DevSecOps and observability skills. Everything marked New in C03 in this guide is new content.
Where the points are
Domain
Weight
Approx. scored questions
1. Architecture Design and Implementation Built for the Cloud
25%
~16
2. Security, Compliance, and Governance
25%
~16
3. Cost-Optimized Architecture Design
15%
~10
4. Resilience, Migration, and Business Continuity
15%
~10
5. Operational Excellence and Automation
20%
~13
Domains 1 and 2 are half the exam. Security is now its own 25% domain, so IAM policy logic, Organizations, KMS and data perimeters matter more than ever.
How SAP-C02 domains map to SAP-C03
SAP-C02 domain
%
SAP-C03 domain
%
1. Design Solutions for Organizational Complexity
26%
1. Architecture Design and Implementation Built for the Cloud
25%
2. Design for New Solutions
29%
2. Security, Compliance, and Governance
25%
3. Continuous Improvement for Existing Solutions
25%
3. Cost-Optimized Architecture Design
15%
4. Accelerate Workload Migration and Modernization
20%
4. Resilience, Migration, and Business Continuity
15%
—
5. Operational Excellence and Automation
20%
AWS says the reorganization consolidates C02 skills rather than dropping them: domains are now grouped by pillar (security, cost, resilience, operations) instead of by project stage (new vs existing vs migration).
What's new in SAP-C03
AWS's own summary groups the additions into these themes. Each is covered in its domain tab.
Theme
New skills
Where
GenAI and agentic AI
Bedrock integrations (multi-account), AgentCore agents with tools, RAG with Knowledge Bases, SageMaker AI inference, Guardrails, AgentCore Identity, human-oversight approval workflows, AI cost and AI monitoring, Bedrock Evaluations
1.2, 2.5, 3.1, 3.2, 5.2, 5.3
Cloud-native patterns
Serverless data pipelines, durable workflows (Step Functions, Lambda durable functions), service mesh (VPC Lattice, ECS Service Connect), container image security (ECR), ingress patterns, WebSockets, multi-tenant isolation
1.1
Data and analytics
Lakehouse (Lake Formation, Apache Iceberg, Redshift Spectrum), real-time and batch pipelines (EMR, MSK, AppFlow), cross-account data sharing (Clean Rooms)
1.5
Hybrid and network
Hybrid compute (Outposts, Local Zones, EKS/ECS Anywhere), IPv6 dual-stack migration
1.3, 1.4
Identity
Cognito, OAuth 2.0, EKS Pod Identity, IAM Roles Anywhere
Pipeline vulnerability scanning and signing, cross-account pipelines, Container Insights, Synthetics and RUM, cost-effective observability
3.4, 5.2, 5.3
Removed from the blueprint
"Application upgrade path", "Delegating to AWS" and "Migration-tool security". AWS notes these may still be implicit in other skills.
In-scope and out-of-scope services
AWS publishes a list of services that can appear as right answers and a short list that should not. Both are "non-exhaustive". C03 lists parent services only, so features count too (for example Bedrock Guardrails, AgentCore, S3 Object Lock, Route 53 Resolver). In the Services A–Z tab each card shows In scope, Out of scope or Not listed (on neither list; still might appear, but less likely to be the answer).
Category
In scope (SAP-C03)
Analytics
AppFlow, Athena, EMR, Kinesis Data Streams, MSK, OpenSearch Service, QuickSight, Redshift, Clean Rooms, Glue, Lake Formation
All IoT services, WorkSpaces and AppStream 2.0, Elastic Transcoder, Kinesis Video Streams, the individual AI services (Comprehend, Rekognition, Textract and so on), CodeGuru, Lightsail, Pinpoint, SES, Amplify, App Runner, AWS Batch, CodeArtifact, Data Exchange, Device Farm, Health Dashboard, License Manager, Proton, Service Catalog and Wavelength.
If an option is on the out-of-scope list, it's almost certainly a distractor. Cross it off first. If two options look equally good, lean toward the in-scope one.
Services closed to new customers
AWS moved these to maintenance on November 7, 2025 (existing customers only). Old practice questions love them; on a 2026 exam, prefer the current replacement when one exists.
Service
Prefer instead
AWS Snowball Edge (Storage and Compute Optimized)
Direct Connect, DataSync over the network, or AWS Data Transfer Terminal (bring your own drives to an AWS facility)
AWS Migration Hub, AWS Application Discovery Service
S3 Object Lambda, Amazon Glacier (the original vault service)
Lambda behind S3 access points or API Gateway; S3 Glacier storage classes (still fully available)
Out of scope (don't over-study)
The guide says you are not expected to have: in-depth operating-system knowledge, AI/ML model training, fine-tuning or data-science workflows, security incident forensics or detailed threat investigation, or application code development and debugging. You design and choose; you don't build models or debug code.
The classic answer-killers
Phrase in the question
What it rules out
"Least operational overhead" / "fully managed"
Anything on EC2 you patch yourself, custom scripts, self-managed Kafka or databases
"Most cost-effective"
Over-provisioning, active/active when pilot light meets the RTO, NAT gateway for S3 traffic
Architecture design and implementation built for the cloud
25% of scored content · Tasks 1.1–1.5
1.1Event-driven architectures and messaging
Services talk through messages and events instead of direct calls, so each part can scale and fail on its own. The exam tests which messaging service fits the delivery guarantee, ordering and fan-out needs.
Existing Kafka apps/tools; lift Kafka with no code change
Amazon MQ
Managed ActiveMQ / RabbitMQ broker
Broker semantics (JMS, AMQP, MQTT, STOMP)
Migrating apps that already use a standard broker protocol
Making consumers safe
Idempotency: processing the same message twice gives the same result. Store a processed-message ID (for example a DynamoDB conditional write, or Powertools for AWS Lambda idempotency) so retries don't double-charge.
Dead-letter queues (DLQs): park messages that keep failing; alarm on DLQ depth; redrive later.
Visibility timeout (SQS) must be longer than processing time, or another consumer picks the message up again.
Partial batch failure reporting for Lambda + SQS/Kinesis so one bad record doesn't retry the whole batch.
Queue = a to-do list one team works through. Topic = a megaphone. Bus = a mail sorter that reads the envelope. Stream = a tape recorder anyone can rewind.
Try it
Pick the messaging service
Cross-account and cross-Region event routing
EventBridge rules can target an event bus in another account or Region. A common pattern: every workload account forwards security or ops events to a central bus in a monitoring account. The target bus needs a resource-based policy allowing the source (use aws:PrincipalOrgID to allow the whole organization). EventBridge also offers global endpoints for Regional failover of event ingestion.
"Decouple and process in order per customer" → SQS FIFO with customer ID as the message group ID. "Fan out one order event to billing, shipping and analytics" → SNS → SQS (or EventBridge with three rules). "Replay last week's clickstream into a new consumer" → Kinesis Data Streams (or EventBridge archive and replay for events).
1.1Stateless apps, serverless pipelines and durable workflows
Externalize state
A stateless server keeps nothing it can't lose, so any instance can serve any request and Auto Scaling can kill instances freely. Put session data in ElastiCache or DynamoDB (with TTL), files in S3, shared POSIX files in EFS. Avoid sticky sessions as the main design: they pin users to instances that might disappear.
Serverless data pipelines New in C03
Producers→Kinesis Data Streams→Lambda (transform) or Amazon Data Firehose→S3 data lake (Parquet)→Glue crawler / Data Catalog→Athena SQL→QuickSight
No servers anywhere: you pay per GB or per request. Firehose buffers, converts to Parquet and partitions on the way into S3. Athena charges per TB scanned, so columnar formats and partitions cut cost.
Durable multi-step workflows New in C03
Option
Max duration
Execution model
Pick it for
Step Functions Standard
Up to 1 year
Exactly-once per step; full visual execution history; priced per state transition
Long-running, auditable business processes; human approvals
Step Functions Express
Up to 5 minutes
At-least-once; priced per request and duration
High-volume, short event processing (IoT, streaming transforms)
Lambda durable functions
Up to 1 year
Checkpoint and replay inside your Lambda code: steps, waits (no compute charge while waiting), callbacks
Developers who want a workflow as ordinary sequential code rather than a state-machine definition
Plain Lambda
15 minutes
Single invocation
One short task
Step Functions Standard = a registered letter (slow, tracked, exactly once). Express = bulk mail (fast, cheap, at least once).
Any workflow longer than 15 minutes rules out a single Lambda. Need a person to approve a step → Step Functions with a callback task token (.waitForTaskToken) or a Lambda durable function callback.
1.1Containers, service mesh and image management
Option
What it is
Pick it when…
Amazon ECS
AWS-native container orchestration; simpler
No Kubernetes requirement; least operational overhead
Amazon EKS
Managed Kubernetes control plane
Team already uses Kubernetes, needs its APIs, Helm, portability, or hybrid with EKS Anywhere
AWS Fargate
Serverless compute for ECS or EKS tasks/pods
No EC2 nodes to patch or scale
EC2 launch type / managed node groups / Karpenter
You run the nodes
GPUs, special instance types, daemonsets, maximum cost control with Spot
Service mesh and service-to-service networking New in C03
ECS Service Connect: built-in service discovery, load balancing, retries and traffic metrics between ECS services by short name. No sidecar config to manage yourself.
Amazon VPC Lattice: an application-layer network that connects services across VPCs and accounts (EC2, ECS, EKS, Lambda) with IAM auth policies, without peering or Transit Gateway, and even with overlapping CIDRs. Share service networks with AWS RAM.
AWS App Mesh was retired (end of support September 30, 2026); prefer Service Connect or VPC Lattice.
Container images in ECR New in C03
Basic scanning (on push, OS CVEs) vs enhanced scanning (Amazon Inspector: continuous, OS and language packages).
Lifecycle policies expire old or untagged images to cut storage cost.
Tag immutability stops a tag like v1.4 being overwritten; replication copies images cross-Region / cross-account (needed for DR).
Sign images (AWS Signer with Notation) and verify at deploy time.
Auto scaling for serverless and containers
Workload
Scaling control
Lambda
Reserved concurrency (cap and guarantee a slice), provisioned concurrency (pre-warmed, removes cold starts; can be scheduled with Application Auto Scaling)
ECS
Service auto scaling (target tracking on CPU, memory or ALB requests per target) via Application Auto Scaling; capacity providers scale the EC2 cluster
EKS
Horizontal Pod Autoscaler (more pods), Cluster Autoscaler or Karpenter (more nodes)
Network calls, bigger code, body access, origin selection
Functions = flip a header.Lambda@Edge = do real work.
Real-time with WebSockets and pub/sub New in C03
API Gateway WebSocket APIs keep a two-way connection open. Store connection IDs in DynamoDB; backends push to clients through the @connections management API. Combine with SNS/EventBridge to publish one event to many connected clients.
1.1Multi-tenant SaaS isolation New in C03
Model
How
Trade-off
Silo
Each tenant gets its own resources (account, VPC, database, or stack)
Strongest isolation and easy per-tenant cost; most expensive and hardest to operate at scale
Pool
Tenants share resources; tenant ID in every row and request
Cheapest and most efficient; noisy neighbors and isolation must be enforced in code and policy
Bridge (hybrid)
Shared app tier, siloed data tier (or premium tenants siloed)
Balance; common in practice
Tools for isolation
Cognito tenant context: a custom attribute (custom:tenant_id) in the JWT that every service reads.
Dynamic IAM with ABAC: assume a role with a tenant session tag; policies allow only resources tagged with the same tenant (DynamoDB dynamodb:LeadingKeys, S3 prefixes).
Row-level security (RLS): PostgreSQL policies so queries only see the tenant's rows.
Tenant-specific encryption: one KMS key per tenant (crypto-isolation; deleting the key "shreds" that tenant's data).
Per-tenant throttling with API Gateway usage plans to stop noisy neighbors.
"Regulated enterprise tenants demand dedicated infrastructure; small tenants must be cheap" → bridge/tiered model: silo the premium tier, pool the rest.
1.2Integrating generative AI and agents New in C03
You're the architect, not the data scientist: know which AWS building block to wire in and how to secure, scale and share it across accounts. Training and fine-tuning are out of scope.
Building block
What it does
Pick it when…
Amazon Bedrock
Serverless API (InvokeModel, Converse) to foundation models from Amazon and other providers
Add GenAI to an app without managing model servers
Bedrock Knowledge Bases
Managed RAG: ingestion, chunking, embeddings, vector store, retrieval with citations
Answer from company documents; data changes often
Bedrock Agents
Managed agent with action groups (Lambda / OpenAPI) and Knowledge Bases
Simple managed agent inside Bedrock
Bedrock AgentCore
Run any agent framework securely at scale: Runtime, Memory, Gateway (APIs/Lambda → MCP tools), Identity, Policy (Cedar), Observability, built-in Browser and Code Interpreter
Production agents, any framework or model, tool access with enterprise identity
SageMaker AI endpoints
Host your own or open-weight models
Custom model, specific instance/GPU control, model not on Bedrock
Multi-account GenAI patterns
Central AI platform account exposes models through API Gateway + Lambda (or a gateway on ECS) so workload accounts get one governed endpoint with logging, quotas, guardrails and chargeback.
Or let each workload account call Bedrock directly, governed by SCPs (allow only approved model IDs/Regions), mandatory Guardrails (IAM condition bedrock:GuardrailIdentifier), and org-wide model invocation logging.
Application inference profiles carry tags per team or app for cost tracking.
Cross-Region inference profiles (geographic such as US or EU, or global) spread requests across Regions for throughput and resilience, with no extra routing charge; geographic profiles keep processing inside that geography.
Private access: Bedrock interface VPC endpoints (PrivateLink) keep traffic off the internet.
RAG architecture, in order
Ingest documents from S3 (or connectors such as SharePoint, Confluence, web crawler); sync on a schedule or event.
Chunk them (fixed-size, hierarchical, semantic, or none).
Create embeddings with an embeddings model.
Store vectors (OpenSearch Serverless, Aurora PostgreSQL with pgvector, Neptune Analytics, Amazon S3 Vectors, or partner stores).
At query time retrieve the top chunks (optionally hybrid search, metadata filters, reranking).
Send question + chunks to the model; return answer with citations.
Answers must reflect documents updated daily → RAG with Knowledge Bases, not fine-tuning. Already on Aurora PostgreSQL → pgvector with the fewest new services. Different users may see only their documents → metadata filtering at retrieval on a user/department attribute.
ML inference options (SageMaker AI)
Type
Pick it when…
Real-time endpoint
Steady traffic, low latency
Serverless inference
Spiky or idle traffic; scales to zero (cold starts)
Asynchronous inference
Large payloads (up to 1 GB), long processing; queue-based; can scale to zero
Batch transform
Score a whole dataset offline
Multi-model endpoints / inference components
Many models sharing one endpoint's instances to cut cost
Bedrock = rent a model by the word. SageMaker endpoint = run your own model's server. AgentCore = the secure office your agents work in.
1.3Hybrid connectivity with redundancy
Option
What it is
Strengths
Limits
Site-to-Site VPN
IPsec over the internet, 2 tunnels per connection, BGP or static
Up in minutes, cheap, encrypted
~1.25 Gbps per tunnel; internet variability. Use ECMP over Transit Gateway to aggregate; Accelerated VPN uses Global Accelerator
Direct Connect (DX)
Private fiber link to AWS: dedicated 1/10/100/400 Gbps, or hosted from a partner (50 Mbps–25 Gbps)
Consistent latency, high bandwidth, lower egress price
Weeks to provision; not encrypted by default (use MACsec on dedicated links or VPN over DX)
DX + VPN backup
VPN as failover for DX
Cheap redundancy
Backup bandwidth is lower
Direct Connect building blocks
Virtual interfaces (VIFs):private VIF (to one VPC's VGW or a DX gateway), transit VIF (to a DX gateway attached to Transit Gateways), public VIF (AWS public endpoints like S3 over DX).
Direct Connect gateway: global object; one DX connection reaches VPCs/TGWs in many Regions and accounts.
SiteLink: send traffic between your own sites over the AWS backbone.
Resiliency models (DX Resiliency Toolkit): Maximum resiliency = separate connections on separate devices in two or more locations (99.99% SLA); High resiliency = one connection at each of two locations (99.9%); Development and test = two connections on separate devices at one location (no SLA).
DX is a private highway; VPN is a tunnel through public roads. Redundancy = two highways from two different on-ramps.
"Encrypted, 10 Gbps, private" → DX with MACsec (or IPsec VPN over a DX public/transit VIF). "Need connectivity next week while DX is ordered" → VPN now, DX later, keep VPN as backup. Prefer BGP with AS_PATH prepending or local preference to make one path primary.
1.3Hybrid data sync, legacy integration and DNS
Service
Job
Pick it when…
AWS DataSync
Online, scheduled, incremental copy (NFS, SMB, HDFS, object storage, other clouds) ↔ S3, EFS, FSx; verifies data
Move or sync large file sets over the network
Storage Gateway – S3 File Gateway
NFS/SMB share on-premises backed by S3, local cache
On-prem apps keep using a file share; data lands in S3
Storage Gateway – Volume Gateway
iSCSI block volumes backed by S3 with EBS snapshots (cached or stored)
Block storage backup/DR to AWS
Storage Gateway – Tape Gateway
Virtual tape library to S3 Glacier
Replace physical tapes with no backup-software change
AWS Transfer Family
Managed SFTP, FTPS, FTP, AS2 endpoints into S3/EFS
Partners already send files by SFTP or AS2
Modernizing legacy integration
Put a facade in front of the old system: API Gateway + Lambda to transform payloads (XML/SOAP ↔ JSON), or Amazon MQ to keep JMS/AMQP producers unchanged while new consumers come online. EventBridge Pipes can filter and enrich between sources and targets.
Hybrid DNS with Route 53 Resolver
Direction
Use
Inbound endpoint
On-premises DNS servers forward queries into AWS to resolve private hosted zone names
Outbound endpoint + forwarding rules
VPC resources resolve on-premises names (corp.example.com) by forwarding to on-prem DNS. Share rules with other accounts through AWS RAM.
Private hosted zones
Associate with many VPCs, including in other accounts (authorize, then associate)
Split-horizon DNS
Same name, public zone answers internet users, private zone answers inside the VPC
INbound = queries coming IN to AWS. OUTbound = queries going OUT to your data center.
Hybrid compute New in C03
Option
Pick it when…
AWS Outposts (rack or server)
AWS-managed infrastructure in your data center: low latency to on-prem systems, local data processing, data residency
AWS Local Zones
AWS-run infrastructure in a metro area for single-digit-millisecond latency to users (media, gaming, real-time)
Amazon EKS Anywhere
Run Kubernetes on your own hardware with EKS tooling (you manage the hardware and clusters)
Amazon ECS Anywhere
Register on-prem servers as ECS capacity, managed from the ECS control plane
Try it
Which hybrid option?
Think of your answer, then tap a card to check it.
1.4Multi-VPC and multi-account connectivity
Option
Transitive?
Overlapping CIDRs?
Scale
Pick it when…
VPC peering
No
No
1:1, full mesh gets messy past ~10 VPCs
Few VPCs, highest bandwidth, no hourly charge
Transit Gateway (TGW)
Yes (hub and spoke)
No
Thousands of attachments; route tables for segmentation; inter-Region peering; VPN and DX attachments
Many VPCs + on-prem; central inspection; isolated route domains
AWS PrivateLink
N/A (service-level)
Yes
One service exposed to many consumers; consumer sees an ENI in its own VPC
Expose one app/API to other accounts or SaaS customers privately, one direction
Amazon VPC Lattice
N/A (service-level)
Yes
Service network across VPCs and accounts with IAM auth policies
Many microservices across accounts talking to each other at layer 7
AWS Cloud WAN
Yes
No
Global network from a central policy; segments; multi-Region automatically
Global multi-Region WAN with central policy, many sites and Regions
Shared VPC (RAM subnet sharing)
—
—
Many accounts deploy into one VPC's subnets
Central network team owns the VPC; app teams own their resources
Try it
Pick the connectivity option
Peering = a private hallway between two rooms. TGW = a central train station. PrivateLink = a service window in the wall: you can buy coffee but can't walk into the kitchen.
VPC design, IPAM and segmentation
Amazon VPC IP Address Manager (IPAM): plan, allocate and track CIDRs across accounts and Regions from pools (shared via RAM) so VPCs never overlap; audits usage and compliance.
Subnets per AZ in tiers (public for load balancers and NAT, private for apps, isolated for data). Reserve big ranges: a VPC can have extra secondary CIDRs.
Segment with separate VPCs/accounts and TGW route tables (prod can't reach dev), not just security groups.
1.4Network security and inspection
Control
Layer / scope
Key traits
Security groups
ENI / instance
Stateful, allow rules only, can reference other SGs
Network ACLs
Subnet
Stateless (allow return traffic explicitly), allow and deny, rules evaluated in number order
Block DNS lookups of malicious domains (exfiltration)
VPC Flow Logs
Visibility
Metadata of accepted/rejected flows to CloudWatch Logs, S3 or Firehose (not packet contents)
Centralized inspection pattern
Spoke VPCs→Transit Gateway→Inspection VPC (Network Firewall or GWLB endpoints)→Egress VPC (NAT) / other spokes / on-prem
Turn on TGW appliance mode on the inspection VPC attachment so both directions of a flow hit the same AZ's firewall (keeps stateful inspection symmetric). Firewall Manager can push Network Firewall, WAF and security group policies across the organization.
Block outbound traffic to all domains except an approved list → Network Firewall domain allow list (stateful) or DNS Firewall. A security group can't filter by domain name, and NACLs can't either.
1.4IPv6 and global traffic routing
IPv6 and dual-stack migration New in C03
Add an Amazon-provided (or BYOIP / IPAM-allocated) IPv6 /56 to the VPC, a /64 to each subnet; run dual-stack first, then IPv6-only subnets for new workloads.
All IPv6 addresses are globally unique, so there's no NAT. For private outbound-only IPv6 use an egress-only internet gateway.
Update security groups, NACLs and route tables for ::/0; ALB/NLB/CloudFront support dual-stack.
Motivation: IPv4 exhaustion inside large organizations, and public IPv4 addresses carry an hourly charge.
Route 53 routing policies
Policy
Use
Simple
One record, no health checks
Failover
Active-passive DR with health checks
Weighted
Split traffic by percentage (canary, blue/green)
Latency
Send users to the Region with lowest latency
Geolocation
By the user's country/continent (content rights, compliance)
Geoproximity
By distance, with a bias to grow or shrink a Region's share
Multivalue answer
Up to 8 healthy records, simple client-side load spreading
IP-based
By the client's IP range (route a known ISP to a specific endpoint)
Global Accelerator vs CloudFront
AWS Global Accelerator
Amazon CloudFront
What
2 static anycast IPs; traffic enters the AWS backbone at the nearest edge
CDN that caches content at the edge
Protocols
TCP and UDP (any app)
HTTP/HTTPS (and WebSockets)
Failover
Seconds, no DNS caching problem
Origin failover groups
Pick when
Gaming/VoIP/IoT, fixed IPs for allow-lists, fast multi-Region failover
Static and dynamic web content, cost of egress, edge security
CloudFront keeps copies. Global Accelerator finds the fastest lane and never changes its phone number.
1.5Data lakes, lakehouses and analytics pipelines
Lakehouse on AWS New in C03
Amazon S3 stores data in open formats (Parquet) and open table formats like Apache Iceberg (ACID transactions, row-level updates/deletes for GDPR, schema evolution, time travel). S3 Tables are S3 buckets with built-in Iceberg management.
AWS Glue Data Catalog holds table definitions; Glue crawlers discover schemas; Glue ETL jobs (Spark) transform.
AWS Lake Formation governs access: fine-grained table, column, row and cell permissions, LF-Tags (tag-based access control), and cross-account sharing.
Query engines read the same data: Athena (serverless SQL), Redshift Spectrum (warehouse joins with lake data), EMR (Spark, Hive, Presto).
Real-time vs batch pipelines
Need
Services
Streaming ingestion
Kinesis Data Streams, Amazon MSK (Kafka), Firehose for delivery
Stream processing
Lambda, Amazon Managed Service for Apache Flink, Spark streaming on EMR/Glue
Batch processing
Glue (serverless Spark), EMR (on EC2, EKS or Serverless) for big custom Spark/Hadoop
Warehouse
Amazon Redshift (provisioned RA3 or Serverless), zero-ETL integrations from Aurora/RDS/DynamoDB, data sharing between clusters/accounts
SaaS ingestion
Amazon AppFlow (Salesforce, SAP, ServiceNow, Slack… → S3/Redshift), no code
Search and log analytics
Amazon OpenSearch Service
Dashboards
Amazon QuickSight
Cataloging and cross-account sharing
Need
Answer
Share catalog tables with another account, with column-level control
Lake Formation cross-account grants (uses RAM under the hood)
Share live warehouse data, no copies
Redshift data sharing
Two companies analyze combined data without either seeing the other's raw records
AWS Clean Rooms (analysis rules, aggregation thresholds, differential privacy) New in C03
"Analysts in many accounts, one governed lake, column-level permissions, least operational overhead" → central Glue Data Catalog + Lake Formation with LF-Tags, not S3 bucket policies per prefix.
"Read-heavy reporting is slowing the primary" → read replica / Aurora replica (reader endpoint). "Survive an AZ failure" → Multi-AZ. "Survive a Region failure, RPO about a second" → Aurora Global Database. "Writes in several Regions at once" → DynamoDB global tables. Multi-AZ is for availability; replicas are for scale.
1.5Storage architectures and caching
Service
Type
Pick it when…
Amazon S3
Object
Unlimited objects, data lakes, static sites, backups; 11 nines durability
Many Linux instances/containers/Lambdas share files
FSx for Windows File Server
SMB, Active Directory integrated
Windows file shares, DFS, user home folders
FSx for Lustre
High-performance parallel file system, links to S3
HPC, ML training, media rendering
FSx for NetApp ONTAP
NFS + SMB + iSCSI, snapshots, SnapMirror
Lift NetApp workloads; multi-protocol
FSx for OpenZFS
NFS, ZFS features
Lift ZFS/Linux file workloads with low latency
Try it
Pick the storage
Caching strategies
Cache
Where
Use
CloudFront
Edge
Static and cacheable dynamic content near users
API Gateway cache
API (REST)
Repeated identical API responses
ElastiCache
App tier
Any database or computed results; sessions
DynamoDB Accelerator (DAX)
In front of DynamoDB only
Microsecond reads with no code rewrite beyond the DAX client
Lazy loading (cache-aside): read cache, on miss read DB and fill cache; data can be stale, set a TTL. Write-through: update cache on every write; always fresh, more writes.
DAX is a speed boost bolted to DynamoDB. ElastiCache is a general-purpose speed boost you wire into your code yourself.
Single sign-on for people into many AWS accounts and apps. Connect an external IdP (Okta, Entra ID) by SAML 2.0, sync users and groups automatically by SCIM, grant access with permission sets (become roles in each account). Supports ABAC from IdP attributes.
SAML 2.0
XML-based standard for browser SSO between an IdP and a service provider
OpenID Connect (OIDC)
Identity layer on OAuth 2.0 using JWTs; used by Cognito, web/mobile logins and CI systems (GitHub Actions → IAM role, no stored keys)
SCIM
System for Cross-domain Identity Management: automatic user/group provisioning and deprovisioning
IAM SAML / OIDC identity providers
Federation directly into one account's roles (older pattern; Identity Center is preferred for many accounts)
AWS Directory Service
Managed Microsoft AD (real AD in AWS, trusts with on-prem AD), AD Connector (proxy to on-prem AD, no directory in AWS), Simple AD (basic Samba)
"Employees use the corporate IdP; access to 200 accounts; leavers lose access automatically" → IAM Identity Center + SAML 2.0 + SCIM. "Apps in AWS need to join the on-prem domain with no directory replication" → AD Connector; "needs a domain in AWS that survives a link outage" → Managed Microsoft AD with a trust.
2.1Application authentication and workload identities New in C03
Need
Answer
Sign up / sign in end users of a web or mobile app; social and SAML/OIDC federation; JWTs
Cognito user pools (authentication; issues ID, access and refresh tokens)
Give app users temporary AWS credentials to call S3/DynamoDB directly
Cognito identity pools (trade a token for scoped IAM role credentials; supports guest access)
Fine-grained in-app permissions ("can this user edit this document?")
Amazon Verified Permissions (Cedar policy language, central authorization decisions)
Machine-to-machine API calls
OAuth 2.0 client credentials grant (Cognito resource server + scopes)
User login in an SPA or mobile app
OAuth 2.0 authorization code grant with PKCE
Pods on EKS need AWS permissions
EKS Pod Identity (simplest: associate role with service account; no OIDC provider per cluster) or IRSA (IAM Roles for Service Accounts)
Servers on-premises or in another cloud need AWS credentials without long-term keys
IAM Roles Anywhere: X.509 certificate from your CA (or AWS Private CA) registered as a trust anchor → temporary credentials
ECS tasks / Lambda / EC2
Task role / execution role / instance profile (never store access keys)
User pools = who are you (login). Identity pools = what AWS keys do you get. Verified Permissions = what may you do inside the app.
2.1Least privilege and IAM policy evaluation
For a request inside one account, AWS checks every policy type. An explicit deny anywhere wins. Otherwise every guardrail layer that applies (SCP, RCP, permissions boundary, session policy) must allow it, and an identity-based or resource-based policy must grant it.
A role in account A calls s3:GetObject on a bucket. Set each policy layer, then read the verdict.
SCP on the caller's account
RCP on the bucket's account
Identity policy on the role
Bucket policy (resource-based)
Permissions boundary on the role
Session policy (passed at AssumeRole)
Least-privilege tools
Tool
Use
IAM roles + STS temporary credentials
Cross-account access: trust policy in the target account, sts:AssumeRole from the source
External ID
Third party assumes a role in your account; stops the confused deputy problem
Permissions boundary
Maximum permissions for a role/user; lets developers create roles safely (delegated admin)
Session policy
Narrow a session further at AssumeRole time
ABAC
Allow when aws:PrincipalTag/project equals aws:ResourceTag/project; scales without editing policies for each new project
Session tags / sts:SourceIdentity
Carry user attributes and identity through role chains for ABAC and auditing
SCPs, RCPs, boundaries and session policies are fences: they never grant, only limit. Identity and resource policies are keys. You need a key and must stay inside every fence.
2.1Data perimeters and external access
A data perimeter means: only trusted identities access trusted resources from expected networks.
Perimeter
Typical condition keys
Enforced with
Identity: only my org's principals reach my resources
aws:PrincipalOrgID, aws:PrincipalIsAWSService
RCPs, resource policies
Resource: my principals only reach my org's resources
aws:ResourceOrgID
SCPs, VPC endpoint policies
Network: access only from my VPCs or corporate IPs
VPC endpoint policies: limit what can be done through an endpoint (for example only buckets in my org).
IAM Access Analyzer:external access findings (resources shared outside your zone of trust), internal access findings, unused access (unused roles, keys, permissions), policy validation, custom policy checks in pipelines, and policy generation from CloudTrail.
AWS Verified Access: zero-trust access to internal web apps (and TCP resources) without a VPN; checks identity (Identity Center, OIDC) and device posture on every request.
"Stop anyone outside the organization reading any S3 bucket, even if a developer writes a bad bucket policy, across all accounts" → an RCP denying s3:* when aws:PrincipalOrgID isn't yours (with an exception for AWS service principals). An SCP can't do this: SCPs only limit your own principals.
2.2Encryption at rest with AWS KMS
Key type
Who controls it
Notes
AWS owned keys
AWS, shared across accounts
Default for some services; invisible to you; no audit
AWS managed keys (aws/s3…)
AWS in your account
Auto-rotated yearly; you can't change the key policy or share cross-account
Customer managed keys
You
Your key policy and grants, cross-account sharing, enable/disable, scheduled deletion, configurable automatic rotation (90–2,560 days) and on-demand rotation
Envelope encryption: KMS encrypts a data key; the data key encrypts the data. That's how services encrypt huge data with a key that never leaves KMS.
Multi-Region keys: same key material and ID in several Regions; decrypt in another Region without re-encrypting (DR, global tables, client-side encryption).
Key policy is the primary control; without a statement allowing the account, IAM policies alone can't use the key.
Bring your own: imported key material (you can delete it instantly), CloudHSM key store (keys in your CloudHSM cluster), external key store (XKS) (keys stay in your HSM outside AWS).
Cross-account: grant the other account in the key policy and give its principals IAM permission. Snapshots/AMIs encrypted with AWS managed keys can't be shared; re-encrypt with a customer managed key.
KMS keys are the master key in a bank vault; data keys are the keys to individual safe-deposit boxes, handed out and locked away again by the vault.
2.2Encryption in transit, HSMs and post-quantum crypto
Certificates
AWS Certificate Manager (ACM): free public TLS certificates with auto-renewal for ALB, NLB, CloudFront, API Gateway. CloudFront certificates must be in us-east-1. ACM also offers paid exportable public certificates for use on your own servers.
AWS Private CA: your own private certificate authority for internal TLS, mutual TLS (mTLS), IoT devices, and IAM Roles Anywhere trust anchors; short-lived certificate mode is cheaper.
Enforce modern TLS (1.2/1.3) with load balancer and CloudFront security policies; require HTTPS in bucket policies with aws:SecureTransport.
AWS CloudHSM
Single-tenant FIPS 140-3 Level 3 HSMs in your VPC that only you control (AWS can't access keys). Run a cluster across AZs. Pick it when a regulation demands dedicated HSMs or sole key custody, or when apps need standard HSM APIs (PKCS#11, JCE, Microsoft CNG/KSP), SSL/TLS offload, or Oracle/SQL Server TDE. Otherwise KMS is simpler, integrated and also uses FIPS 140-3 validated HSMs.
Post-quantum cryptography New in C03
Threat
Algorithm
On AWS
"Harvest now, decrypt later": attackers record encrypted traffic today and decrypt it once quantum computers arrive
ML-KEM (FIPS 203) key encapsulation, used in hybrid TLS key exchange (classic ECDH + ML-KEM)
Hybrid post-quantum TLS on AWS KMS, ACM and Secrets Manager endpoints; enable in the AWS SDK / CRT client
Forged signatures on long-lived artifacts (firmware, roots of trust, documents)
A strategy: inventory where you use public-key crypto; prioritize data that must stay secret for many years; adopt hybrid PQ TLS for those connections first; use ML-DSA for new long-lived signing keys; design for crypto agility (swap algorithms without re-architecting). Symmetric AES-256 (what KMS uses for data) is already considered quantum-resistant.
KEM = Keeps secrets in transit (key exchange). DSA = Digital Signature: proves who signed.
2.3Multi-account structure and landing zones
AWS Organizations: accounts grouped into organizational units (OUs); consolidated billing; policies (SCPs, RCPs, tag policies, backup policies, AI opt-out, declarative policies); delegated administrator accounts for security services.
AWS Control Tower: sets up a landing zone (management, Log Archive and Audit accounts, Identity Center, central CloudTrail and Config) and applies controls: preventive (SCPs/RCPs), detective (Config rules), proactive (CloudFormation Hooks). Account Factory vends new compliant accounts; Account Factory for Terraform (AFT) and customizations (CfCT) automate baselines.
Data perimeter: block outside-org access to S3, KMS, SQS, Secrets Manager, STS role assumption
Affects management account?
No
No
Grants permissions?
Never
Never
SCP = rules for your people. RCP = rules for your stuff.
Cross-account resource sharing
AWS Resource Access Manager (RAM) shares resources instead of copying them: VPC subnets (shared VPC), Transit Gateways, Route 53 Resolver rules, IPAM pools, prefix lists, Network Firewall policies, License Manager configs, Aurora clusters for cloning, VPC Lattice service networks, and more. Share with the whole org or an OU with no invitation needed.
2.3Centralized logging, compliance monitoring and patching
Need
Answer
All API activity from every account in one place
CloudTrail organization trail → S3 in the Log Archive account (or CloudTrail Lake)
Configuration history and compliance across accounts
Amazon GuardDuty with a delegated admin and auto-enable for new accounts
One normalized security data lake
Amazon Security Lake (OCSF format in S3; CloudTrail, VPC Flow Logs, Route 53, Security Hub findings; subscribers like SIEMs)
Security posture checks and prioritized findings
Security Hub CSPM (best-practice controls and standards: AWS Foundational Security Best Practices, CIS, PCI DSS, NIST) + Security Hub (correlates findings into exposures, attack paths, response workflows)
Log GenAI prompts and responses
Bedrock model invocation logging to S3 / CloudWatch Logs in each account, aggregated centrally; AgentCore Observability for agents
Find vulnerable software
Amazon Inspector: continuous CVE scanning of EC2, ECR images and Lambda; network reachability; SBOM export; org-wide via delegated admin
Patch fleets on schedule
Systems Manager Patch Manager (patch baselines, maintenance windows; patch policies across the org via Quick Setup)
Auto-fix non-compliance
Config rule → automatic remediation with an SSM Automation document
"New accounts must automatically get GuardDuty, Config and CloudTrail with no manual steps" → Control Tower landing zone / Organizations integrations with delegated administrator + auto-enable. "Detect, then auto-fix an open security group" → Config rule + SSM Automation remediation (or Security Hub automation rule → EventBridge → Lambda).
2.4Meeting compliance, sovereignty and audit requirements
Data residency and sovereignty New in C03
Deploy only in approved Regions; enforce with an SCP denying actions when aws:RequestedRegion isn't allowed (Control Tower has a built-in Region deny control). Exempt global services (IAM, CloudFront, Route 53…).
Block transfers out: deny S3 replication to other Regions, deny KMS key sharing, restrict snapshot copy and AMI sharing.
For GenAI, use geographic cross-Region inference profiles (EU, US…) so prompts stay in the geography; avoid global profiles where residency applies.
Encryption doesn't move data; residency is about where data is stored and processed.
Audit trails, including AI services
CloudTrail management events (default) plus data events (S3 object reads, Lambda invokes, DynamoDB items) when needed; log file integrity validation (digest files) proves logs weren't changed.
Store logs in a separate Log Archive account, encrypted with KMS, with S3 Object Lock (compliance mode) for write-once-read-many retention.
CloudTrail records every Bedrock API call; model invocation logging records the actual prompts and outputs (it's off by default).
Continuously collect evidence from your accounts for your audit
AWS Audit Manager (not on the C03 list, but still the matching service), Config conformance packs, Security Hub CSPM standards
Assess risk before building
Threat modeling (for example STRIDE: spoofing, tampering, repudiation, information disclosure, denial of service, elevation of privilege), Well-Architected reviews, mitigations mapped to controls New in C03
2.5Security and responsible AI controls New in C03
Amazon Bedrock Guardrails
Policy
Blocks or checks
Content filters
Hate, insults, sexual, violence, misconduct (text and images), and prompt attacks (jailbreaks, injection)
Denied topics
Subjects you define ("no investment advice")
Word filters
Specific words, profanity
Sensitive information filters
Block or mask PII; custom regex (account numbers)
Contextual grounding checks
Answers not supported by the source or not relevant (hallucinations in RAG)
Automated Reasoning checks
Verify answers against formal rules (policies) using logic
Guardrails work across models, and the ApplyGuardrail API can check text for any model, even ones outside Bedrock. Enforce them centrally: an IAM policy with the bedrock:GuardrailIdentifier condition key denies any inference call that doesn't use the approved guardrail.
Access controls for agents
AgentCore Identity: inbound authentication of users/apps calling the agent (Cognito, Okta, Entra ID…), and outbound credentials (OAuth tokens, API keys in a secure token vault) so agents act on behalf of a user with that user's consent and scope.
Policy in AgentCore: Cedar rules for which tools/actions an agent may call, enforced at the Gateway.
Least privilege for the agent's role: scope its IAM role to only the tools and data it needs; private VPC connectivity; log everything.
Human oversight workflows
AI proposes action→Step Functions task with .waitForTaskToken→SNS / email / ticket to reviewer→Approve or reject (SendTaskSuccess / Failure)→Execute or roll back
Use for high-impact or low-confidence actions (refunds above a limit, production changes). Lambda durable function callbacks and Bedrock Agents' user confirmation / return of control are alternatives.
"Every team's GenAI app must block PII leakage and prompt injection, enforced centrally" → one approved Guardrail + IAM/SCP deny when bedrock:GuardrailIdentifier doesn't match + invocation logging. "Agent must call Salesforce as the signed-in user" → AgentCore Identity (OAuth on behalf of user).
Layer it like a cake: Savings Plans for the steady base, On-Demand for the bumps, Spot for anything that can be interrupted.
3.1Cost-optimized architectures and rightsizing
Graviton (Arm) instances: up to ~40% better price-performance; also for Lambda, Fargate, RDS, ElastiCache. Usually a recompile or a multi-arch container image.
Step Functions: Standard bills per state transition; Express bills per request and duration. High-volume short workflows → Express is far cheaper.
Lambda: bills per ms × memory; right-size memory (Lambda Power Tuning), use Arm, avoid idle waits (durable functions waits don't bill compute).
Serverless vs always-on: low or spiky utilization → Lambda/Fargate/Aurora Serverless; steady high utilization → EC2/containers with Savings Plans.
AWS Compute Optimizer: ML rightsizing for EC2, Auto Scaling groups, EBS, Lambda, ECS on Fargate, RDS/Aurora, and commercial licenses. Turn on enhanced infrastructure metrics for 3 months of lookback.
Scaling for cost: scheduled scaling (shut dev down at night), target tracking, predictive scaling; Instance Scheduler for stop/start.
Optimizing AI/ML costs New in C03
Lever
How it saves
Bedrock service tiers
Standard (pay per token), Priority (fastest, premium price), Flex (cheaper, for non-urgent work), Reserved (term commitment for predictable throughput)
Batch inference
Large non-urgent jobs from S3 at a lower price than on-demand
Prompt caching
Repeated long prompt prefixes (system prompt, documents) billed at a reduced rate and served faster
Smaller or distilled model; intelligent prompt routing
Send easy requests to a cheaper model
Response caching
Cache answers to frequent questions (ElastiCache / semantic cache) so you don't call the model at all
Cap output tokens; trim context
Fewer tokens billed
SageMaker: serverless or async (scale to zero), multi-model endpoints, Savings Plans, Inferentia/Graviton instances
Stop paying for idle GPUs
Bill too high for a nightly document-summary job → Bedrock batch inference (or Flex tier). Same 20-page policy sent with every chat → prompt caching.
3.2Data transfer and network costs
Traffic
Cost rule of thumb
Internet → AWS (inbound)
Free
AWS → internet
Charged per GB (the big one); CloudFront usually cheaper and origin → CloudFront from AWS origins is free
Same AZ, private IP
Free
Across AZs in a Region
Charged in each direction (includes cross-AZ VPC peering)
Across Regions
Charged per GB (replication, peering, TGW peering)
NAT gateway
Hourly + per-GB processing charge, on top of transfer
Gateway VPC endpoints (S3, DynamoDB)
Free
Interface endpoints (PrivateLink)
Hourly per AZ + per-GB processed (cheaper per GB than NAT)
Transit Gateway
Per attachment-hour + per-GB processed
VPC peering
No hourly or processing fee; only normal cross-AZ / cross-Region transfer
Direct Connect
Port-hours + lower data-transfer-out rate than internet
Try it
Cut the network bill
Think of your answer, then tap a card to check it.
Content delivery
CloudFront: cache static and cacheable dynamic content; long TTLs with versioned file names; cache policies that include only needed headers/cookies/query strings (higher hit ratio).
Origin Shield: an extra central caching layer that collapses requests to the origin (good for multi-Region or expensive origins).
Origin failover: origin groups switch to a secondary origin on errors.
Price classes limit edge locations to cheaper regions.
Cost-aware replication
Replicate only what DR needs: S3 Replication with filters (prefix/tag), lifecycle in the destination (replicate into cheaper storage classes), S3 Replication Time Control only when you need the 15-minute SLA; DynamoDB global tables charge for replicated writes in every Region, so add Regions deliberately.
Data is free to check in, cheap to stay in one room, and costs every time it changes rooms (AZ), buildings (Region) or leaves (internet).
3.3Cost allocation, visibility and alerts
Need
Answer
Attribute spend to business units, projects, environments
Cost allocation tags (activate user-defined tags in Billing; AWS-generated tags like aws:createdBy)
Standardize tag keys and values
Tag policies (Organizations)
Block untagged resources at creation
SCP denying create actions when aws:RequestTag/CostCenter is missing
Detect untagged resources after creation
AWS Config required-tags rule (+ remediation)
Group costs by rules (team = these accounts + these tags)
AWS Cost Categories
Split shared EKS/ECS cluster costs by pod or task
Split cost allocation data in the Cost and Usage Report
Cost and Usage Report (CUR 2.0 via Data Exports) → S3 → Athena / QuickSight
Explore and forecast spend, RI/SP recommendations
AWS Cost Explorer
Alert (or act) before spend passes a threshold
AWS Budgets (cost, usage, RI/SP coverage & utilization); budget actions can apply an SCP/IAM policy or stop EC2/RDS
Catch unusual spikes automatically
AWS Cost Anomaly Detection (ML monitors per service, account, tag or cost category; SNS/email alerts)
Showback = show each team its bill. Chargeback = actually send them the bill.
Budgets is a threshold you set; Anomaly Detection finds spikes you didn't predict. "Stop a sandbox account spending past $500" → Budgets with a budget action that applies a deny SCP.
3.4Storage cost: S3 classes and lifecycle
Class
Retrieval
Min. duration
AZs
Pick it when…
S3 Standard
ms
—
≥3
Frequently accessed
S3 Intelligent-Tiering
ms (optional archive tiers slower)
—
≥3
Unknown or changing access; no retrieval fees; small monitoring fee per object (objects <128 KB aren't tiered)
S3 Standard-IA
ms, per-GB retrieval fee
30 days
≥3
Infrequent but must be instant (monthly)
S3 One Zone-IA
ms
30 days
1
Re-creatable infrequent data (secondary copies, thumbnails)
S3 Glacier Instant Retrieval
ms
90 days
≥3
Rarely accessed (quarterly) but needed instantly (medical images)
S3 Glacier Flexible Retrieval
minutes to 12 hours (free bulk)
90 days
≥3
Archives, backups
S3 Glacier Deep Archive
12–48 hours
180 days
≥3
Cheapest; compliance archives kept 7–10 years
Try it
Pick the S3 storage class
Lifecycle rules transition objects by age (for example Standard → Standard-IA at 30 days → Glacier Flexible at 90 → Deep Archive at 365) and expire old versions and incomplete multipart uploads. S3 Storage Lens shows where to optimize. EFS lifecycle moves cold files to EFS Infrequent Access and Archive; EFS One Zone is cheaper for non-critical data.
Glacier names by speed: Instant (now), Flexible (later today), Deep (the day after tomorrow).
3.4Block storage, database, backup and observability costs
EBS volume types
Type
Performance
Cost note
gp3
3,000 IOPS and 125 MiB/s baseline at any size; provision more IOPS/throughput independently
About 20% cheaper per GB than gp2; usually the answer
gp2
3 IOPS per GiB (burst to 3,000 on small volumes); performance tied to size
Teams over-provision size just to get IOPS → migrate to gp3 (no downtime)
io2 Block Express
Up to 256,000 IOPS, sub-ms latency, 99.999% durability, Multi-Attach
Only for critical high-IOPS databases
st1 / sc1
Throughput-optimized / cold HDD
Big sequential (logs, data warehouse) / rarely accessed
Delete unattached volumes and old snapshots; EBS Snapshots Archive is up to 75% cheaper for snapshots kept 90+ days.
Database cost levers
Lever
Use when
DynamoDB on-demand
Unpredictable traffic, new tables, spiky
DynamoDB provisioned + auto scaling (+ reserved capacity)
Predictable traffic; cheaper at steady load
DynamoDB Standard-IA table class
Storage dominates cost, data rarely read
RDS reserved DB instances / Database Savings Plans
Steady databases
Aurora Serverless v2
Variable or intermittent load; scales in fine-grained ACUs (can scale to 0 with auto-pause)
Aurora I/O-Optimized
I/O charges are more than ~25% of Aurora spend; no per-I/O charges, higher instance/storage price
Aurora Standard
Low-to-moderate I/O
Backup retention
AWS Backup plans: schedules, retention, lifecycle to cold storage (must stay 90+ days), cross-Region and cross-account copies.
Amazon Data Lifecycle Manager (DLM): automate EBS snapshot and AMI creation, retention, cross-Region copy and archiving, for free. New in C03
Keep only what the RPO and regulations require; tier long-term backups to cold/archive.
Cost-effective observability New in C03
Set CloudWatch Logs retention on every log group (default is never expire).
Use the Infrequent Access log class for logs you rarely query (cheaper ingestion).
Filter at the source (log level, subscription filter patterns) and sample debug logs; export old logs to S3 Glacier.
Aggregate metrics: Embedded Metric Format, fewer high-cardinality dimensions, standard (not high-resolution) metrics unless needed; metric filters instead of custom metric calls.
Avoid duplicate pipelines sending the same logs to several tools.
VMware workloads to Amazon Elastic VMware Service (EVS)
Repurchase ("drop and shop")
Switch to SaaS
On-prem CRM → Salesforce
Replatform ("lift, tinker and shift")
Small changes to use managed services
Self-managed MySQL → RDS; app server → Elastic Beanstalk or ECS
Refactor / re-architect
Rebuild cloud-native
Monolith → microservices on Lambda / containers
Try it
Which R?
Think of your answer, then tap a card to check it.
Effort and benefit both rise as you go Rehost → Replatform → Refactor: move it, tweak it, rebuild it.
4.1Discovery, wave planning and rehosting
Discovery and portfolio assessment: collect server inventory, utilization and network dependencies, then group tightly coupled apps into the same migration wave. AWS Transform does agentic discovery and wave planning; Migration Evaluator builds the business case. (Application Discovery Service and Migration Hub are closed to new customers.)
AWS Transform MGN (previously AWS Application Migration Service): installs an agent, does continuous block-level replication to a low-cost staging area, launches test instances, then cuts over in minutes. Post-launch actions can install agents and modernize. Can be driven by the AWS Transform agent (discovery → landing zone → network → rehost).
AWS Transform (agentic): also modernizes VMware, Windows/.NET, mainframe (COBOL to Java) and code upgrades.
Land in a prepared multi-account landing zone (Control Tower) with networking in place before wave 1.
"Move hundreds of servers quickly with minimal downtime and no code changes" → Transform MGN. Don't pick hand-built AMIs, VM Import scripts, or the old SMS (Server Migration Service, retired).
4.1Database and data migration
Tool
Job
AWS DMS
Migrate databases with full load + change data capture (CDC) for near-zero downtime; homogeneous (MySQL→MySQL) or heterogeneous (Oracle→PostgreSQL); DMS Serverless; also continuous replication to targets like S3, Kinesis, Redshift
DMS Schema Conversion / AWS SCT
Convert schema, stored procedures and code between engines (heterogeneous); SCT also has data extraction agents for warehouses
Native tools
Backups/restore to RDS (for example SQL Server native backup to S3), Percona XtraBackup to Aurora MySQL, logical replication
AWS DataSync
File and object data over the network, scheduled and incremental, with verification
AWS Data Transfer Terminal
Physical AWS facility: bring your own storage devices and upload at high speed (Enterprise Support customers) New in C03
Snowball Edge
Shippable device (closed to new customers since Nov 2025; may still appear in old questions)
Direct Connect
Sustained high-throughput network transfer
Try it
Network or physical? Estimate transfer time
Rule of thumb: 1 Gbps moves about 10 TB a day at full speed. If the math says weeks and the deadline says days, go physical or add bandwidth.
4.1Replatforming, refactoring and legacy modernization
Replatform targets: RDS/Aurora (no DB admin), ECS/Fargate (containerize without Kubernetes overhead), Elastic Beanstalk (managed app platform), Lambda (event-driven pieces), Amazon MQ (keep broker protocol), FSx (keep file protocol).
Strangler fig pattern: put a router (API Gateway, ALB rules, CloudFront) in front of the monolith; move one feature at a time to a new service; route that path to the new service; eventually retire the monolith. Low risk, incremental.
Decomposition: split by business capability / domain (orders, payments); give each service its own data store; communicate through events. Add an anti-corruption layer so new services aren't polluted by the legacy model.
Serverless adoption: move scheduled jobs to EventBridge Scheduler + Lambda, file processing to S3 events + Lambda, workflows to Step Functions.
Legacy (mainframe, .NET Framework, old databases): AWS Transform for mainframe and .NET (agentic code analysis and conversion), replatform COBOL runtimes, move DB2/Oracle to Aurora with DMS + schema conversion.
The strangler fig vine grows around the old tree until the tree can be removed. You never cut it down in one go.
4.2Disaster recovery strategies
RPO (recovery point objective): how much data you can lose, measured back in time. RTO (recovery time objective): how long you can be down.
Strategy
What runs in the DR Region
RPO
RTO
Cost
Backup and restore
Nothing but backups (AWS Backup copies, snapshots, S3 replication); rebuild with IaC
Hours
Hours (up to 24)
$
Pilot light
Data replicated live (DB replica); app servers defined but off / not deployed
Minutes
Tens of minutes to hours
$$
Warm standby
Full stack running at reduced capacity; scale up on failover
Seconds to minutes
Minutes
$$$
Multi-site active/active (or hot standby)
Full capacity serving traffic in both Regions
Near zero
Near zero
$$$$
Try it
Pick the cheapest DR strategy that meets the targets
A pilot light in a furnace gives no heat but lights the burners fast. Warm standby is the furnace on low.
Services for DR
Service
Role
AWS Elastic Disaster Recovery (DRS)
Continuous block-level replication of servers (on-prem or another Region) to a cheap staging area; RPO in seconds, RTO in minutes; point-in-time recovery (ransomware); failback
AWS Resilience Hub
Define RTO/RPO policy per app, assess the architecture, get recommendations, SOPs, alarms and FIS tests; track resilience drift
Fast failover between Regional endpoints behind static IPs
Amazon Application Recovery Controller (ARC)
Region switch (orchestrated multi-Region failover plans), routing controls (highly available on/off switches for DNS failover), zonal shift / zonal autoshift (move traffic away from an impaired AZ), readiness checks
Replicating data across Regions
Data
Replication
Aurora
Global Database (sub-second lag typical, managed failover)
RDS
Cross-Region read replica (promote) or automated backup replication
DynamoDB
Global tables (MRSC for zero RPO)
S3
Cross-Region Replication; Replication Time Control (15-minute SLA); Multi-Region Access Points with failover controls
EFS / FSx
EFS replication; FSx backups or ONTAP SnapMirror
EBS / anything
AWS Backup cross-Region copy; DLM snapshot copy
Containers
ECR replication + IaC/GitOps to recreate clusters
Fail over using data plane actions (health checks, ARC routing controls, pre-scaled capacity), not control-plane actions that might be impaired during an event, like creating resources or editing DNS records in the console. This is called static stability.
Multi-cluster containers for HA/DR New in C03
Run identical EKS or ECS clusters in two Regions (or two clusters per Region for blast-radius isolation), deployed from the same IaC/GitOps (for example Argo CD/Flux) source.
Replicate images (ECR replication) and data stores; route with Route 53 / Global Accelerator / ARC.
Keep state outside clusters (Aurora Global, DynamoDB global tables, S3) so a cluster can be replaced; back up Kubernetes resources and volumes with AWS Backup for EKS or Velero.
Blue/green clusters also make Kubernetes version upgrades safe.
4.3Backup governance and ransomware protection
AWS Backup with Organizations:backup policies applied to OUs; a delegated administrator; cross-account monitoring; resources selected by tags.
Cross-account and cross-Region copies into a separate, locked-down backup account protect against a compromised production account.
Logically air-gapped vaults keep backups isolated and shareable for recovery.
Control
Protects against
S3 Object Lock – governance mode
Deletion/overwrite, except by users with a special permission
S3 Object Lock – compliance mode
Nobody, not even root, can delete before retention ends
Legal hold
Deletion indefinitely until removed
EBS snapshot lock
Snapshot deletion for a set period (governance or compliance mode)
AWS Backup Vault Lock
Deletion of recovery points or changes to retention (compliance mode is immutable after a grace period)
S3 Versioning + MFA Delete
Accidental overwrite/delete
Governance mode = a lock with a master key. Compliance mode = a time-lock safe: nobody opens it early.
4.3Resilient design, failover tests and chaos engineering
Design for failure
Redundancy: at least 2–3 AZs for every tier; Multi-AZ databases; NAT gateway per AZ.
Fault tolerance: load balancers with health checks, queues to absorb spikes, timeouts and retries with backoff.
Blast-radius reduction: cell-based architecture (independent copies of the stack, each serving a subset of customers), shuffle sharding, separate accounts.
Static stability: pre-provision enough capacity in each AZ to survive losing one, so recovery needs no new launches.
AWS Fault Injection Service (FIS) New in C03
Define the steady state (normal metrics: p99 latency, error rate, orders/minute).
Write a hypothesis: "if one AZ loses network, orders/minute stay within 5%".
Build an experiment template: actions (stop instances, inject CPU stress, throttle APIs, AZ power interruption scenario, fail over RDS), targets (by tag), and stop conditions (CloudWatch alarms that halt the experiment).
Run in pre-prod, then production with a small blast radius; measure recovery time against RTO.
Fix weaknesses and repeat, ideally automated in pipelines.
Test failover too: flip Route 53 / ARC to the DR Region on a schedule and measure actual RTO and RPO.
"Prove the app survives an AZ outage, safely, with automatic abort if customers are affected" → FIS experiment with a CloudWatch alarm stop condition. Resilience Hub can recommend and run these tests.
Never hard-code secrets in templates. Use dynamic references ({{resolve:secretsmanager:MySecret:SecretString:password}}), let RDS manage the master password in Secrets Manager (ManageMasterUserPassword), and pass secrets to ECS/Lambda by ARN.
5.2CI/CD and DevSecOps pipelines
Source (CodeCommit / GitHub via CodeConnections)→Build & test (CodeBuild)→Scan (Inspector, ECR, Guard)→Sign (AWS Signer)→Deploy to test account→Manual approval→Deploy to prod accounts/Regions (CodeDeploy / CloudFormation)
Managed builds and tests (buildspec.yml); runs scanners
AWS CodeDeploy
Deploys to EC2/on-prem, Lambda and ECS with in-place or blue/green, canary or linear traffic shifting, and automatic rollback on CloudWatch alarms
AWS CodeCommit
Managed private Git repositories
Security in the pipeline New in C03
Amazon Inspector CI/CD scanning (SBOM generator) in CodeBuild fails builds with critical CVEs; ECR enhanced scanning rescans images continuously after push.
Static checks: Guard/cdk-nag on IaC, Access Analyzer custom policy checks on IAM changes, secrets scanning.
Sign artifacts (AWS Signer for container images and Lambda code signing) and verify before deploy, so only trusted builds run.
Multi-account, multi-Region pipelines New in C03
Pipeline lives in a central tooling/deployment account.
Artifacts in S3 encrypted with a customer managed KMS key whose key policy allows the target accounts (AWS managed keys can't be shared).
Each target account has a cross-account deployment role that the pipeline assumes (role delegation); CloudFormation runs with a least-privilege execution role there.
Artifact promotion: build once, promote the same signed artifact dev → test → prod; never rebuild per environment.
Multi-Region: add actions per Region (CodePipeline cross-Region actions with an artifact bucket per Region) or use StackSets.
Evaluating AI applications New in C03
Amazon Bedrock Evaluations compares models and prompts on your own dataset with automatic metrics, LLM-as-a-judge (correctness, helpfulness, harmfulness…) or human reviewers, and evaluates RAG (retrieval relevance, faithfulness to sources). Run it as a pipeline gate before switching a model or prompt version.
5.2Deployment strategies with rollback
Strategy
How
Downtime
Rollback speed
Extra cost
All at once (in-place)
Update every instance together
Yes
Slow (redeploy old)
None
Rolling
Update in batches
Reduced capacity
Slow
None
Rolling with additional batch
Add a batch first, then roll
No
Slow
Small
Immutable
New instances in a new ASG, then swap
No
Fast (terminate new)
Temporary double
Blue/green
Full new environment; shift traffic (DNS, ALB, CodeDeploy)
No
Instant (switch back)
Double during cutover
Canary
Small % first (for example 10% for 5 minutes), then the rest
No
Fast, automatic on alarm
Low
Linear
Equal increments every N minutes
No
Fast, automatic on alarm
Low
Feature flags
Deploy dark, enable with AWS AppConfig
No
Instant toggle
None
Try it
Pick the deployment strategy
ECS: CodeDeploy blue/green, ECS-native blue/green deployments, and the deployment circuit breaker (auto-rollback when tasks fail to become healthy). Lambda: aliases with weighted traffic + CodeDeploy canary/linear + CloudWatch alarm rollback.
Vendor-neutral instrumentation for metrics and traces
AWS Distro for OpenTelemetry (ADOT): send once, to CloudWatch, X-Ray, Amazon Managed Prometheus or partners
APM: service health, latency, errors, SLOs out of the box
CloudWatch Application Signals
Container metrics and logs (ECS, EKS)
CloudWatch Container Insights (enhanced observability for EKS/ECS) New in C03
Watch many accounts from one
CloudWatch cross-account observability (monitoring account linked to source accounts through Organizations)
Scripted checks from outside, 24/7
CloudWatch Synthetics canaries (synthetic transactions: log in, add to cart) New in C03
Real browser experience (page load, JS errors) of actual users
CloudWatch RUM (real user monitoring) New in C03
Synthetics = a mystery shopper visiting your store every 5 minutes. RUM = cameras watching real shoppers.
Multi-account log aggregation
Log groups in every account→Subscription filter (account-level or per group)→Central destination (Kinesis Data Streams / Firehose) in the logging account→S3 (Glacier tiers via lifecycle) / OpenSearch / SIEM
Alternatives: CloudWatch cross-account observability for live viewing; organization-wide CloudTrail and Security Lake for security logs.
Alerting
CloudWatch alarm → SNS (email, SMS, chat, Lambda) and/or EventBridge (route to automation) and/or Systems Manager OpsCenter (OpsItems for operators). Use composite alarms to cut noise (page only when latency AND errors are high).
AgentCore Observability (OpenTelemetry traces into CloudWatch)
Model performance and drift (data quality, model quality, bias drift, feature attribution drift)
SageMaker Model Monitor on endpoints
Endpoint latency, GPU/CPU utilization
SageMaker endpoint CloudWatch metrics
Answer quality over time
Scheduled Bedrock Evaluations on sampled traffic; user feedback
Service-level objectives
SLI (indicator): what you measure, such as % of requests under 300 ms.
SLO (objective): the target, such as 99.9% over 28 days.
Error budget: 100% − SLO; how much badness you can spend before freezing risky releases.
CloudWatch Application Signals creates SLOs on latency/availability and alarms on burn rate.
Try it
Error budget calculator
5.4Operational automation and safe remediation
Health, capacity and responses
CloudWatch alarm / AWS Health event / Config change→EventBridge rule→SSM Automation runbook / Lambda / Step Functions→Notify via SNS; record OpsItem
Service Quotas: view and request limit increases, alarm on usage approaching quotas, quota request templates applied automatically to new accounts in the org.
Systems Manager: Automation (runbooks), Run Command, State Manager (keep config), Patch Manager, Maintenance Windows, Change Calendar (block changes during freezes), Session Manager (no SSH keys or bastions), Fleet Manager, OpsCenter.
Controlled remediation with safety constraints New in C03
Pattern
Plain meaning
On AWS
Circuit breaker
Stop calling a failing dependency for a while; fail fast instead of piling up
Keep a metric at a target (CPU 50%, requests per target) — the default choice
Step scaling
Add more capacity the further the alarm is breached
Scheduled
Known patterns (business hours, month-end)
Predictive scaling
ML forecast from history (needs at least 24 hours of data; better with 2 weeks) launches capacity before daily/weekly peaks; combine with target tracking
Warm pools
Pre-initialized stopped instances for slow-booting apps
Automated failure recovery
Auto Scaling health checks: EC2 status checks by default; turn on ELB health checks so instances failing the app check are replaced.
ECS replaces tasks that stop or fail container/ALB health checks; EKS restarts pods on liveness probe failure.
Lambda: async invokes retry twice, then go to an on-failure destination or DLQ; stream sources can bisect batches and skip poison records.
EC2 auto recovery moves an instance to new hardware on a system status check failure.
Route 53 health checks remove failed endpoints from DNS.
Runbooks and incident response
Automate routine work with SSM Automation runbooks (AWS-provided like AWS-RestartEC2Instance, or your own): patching through Patch Manager in maintenance windows, remediation triggered by Config or alarms, scaling or failover steps. Start them from EventBridge, alarms or a chat channel; record every run for audit. (Systems Manager Incident Manager is closed to new customers.)
Questions usually hinge on choosing between close neighbors. Each table lists the members, what each does, and the clue that points to it. Click any name to open its card.
Connecting networks
Service
Job
Clue in the question
VPC peering
Private link between two VPCs
"two VPCs", "lowest cost", no transitive need
Transit Gateway
Regional hub for VPCs, VPN, DX
"hundreds of VPCs", "transitive", "central inspection", "isolated route domains"
Cloud WAN
Global WAN from one policy
"global network", "many Regions and branches", "segments by policy"
PrivateLink
Expose one service privately
"SaaS provider", "overlapping CIDRs", "one-way access to a service"
VPC Lattice
Service-to-service app network with IAM auth
"microservices across accounts", "no network plumbing", "auth policies"
"lift and shift", "minimal downtime", "continuous replication"
Migrate databases
DMS + SCT
"CDC", "heterogeneous", "minimal downtime"
Move files online
DataSync
"NFS/SMB to S3/EFS/FSx", "scheduled", "verify"
Ongoing hybrid file access
Storage Gateway
"on-prem apps keep a file share/iSCSI/tape"
Partner file exchange
Transfer Family
"SFTP", "AS2"
Physical bulk upload
Data Transfer Terminal
"bring drives", "network too slow"
Server-level DR
Elastic Disaster Recovery
"RPO seconds", "failback", "on-prem DR to AWS"
Observability and operations
Job
Service
Metrics, logs, alarms, dashboards
CloudWatch
Traces
X-Ray · ADOT
Who did what
CloudTrail
What changed
Config
Fix things automatically
Systems Manager · EventBridge
Test resilience
FIS · Resilience Hub
Limits
Service Quotas
Rightsizing
Compute Optimizer
Amazon Bedrock and AgentCore parts
Feature
Job
Clue
Bedrock model APIs (Converse, InvokeModel)
Call foundation models
"no infrastructure", "many model providers"
Knowledge Bases
Managed RAG
"answer from our documents", "citations"
Guardrails
Content, PII, prompt attacks, grounding
"block topics", "mask PII", "stop jailbreaks"
Agents
Managed agent with action groups
"agent calls Lambda APIs"
Evaluations
Compare models, prompts, RAG
"is the cheaper model good enough"
Batch inference / service tiers / prompt caching
Cost
"overnight", "non-urgent", "same long prompt"
Cross-Region inference profiles
Throughput and resilience across Regions
"throttling", "keep data in the EU"
Model invocation logging
Audit prompts and responses
"record what users asked the AI"
AgentCore Runtime
Serverless, isolated agent hosting
"run agents from any framework"
AgentCore Memory
Short- and long-term memory
"remember across sessions"
AgentCore Gateway
APIs / Lambda → MCP tools
"give agents our APIs as tools"
AgentCore Identity
Inbound auth and outbound credentials
"act on behalf of the user", "OAuth"
Policy in AgentCore
Cedar rules on tool use
"limit which tools an agent can call"
AgentCore Observability
Traces and metrics
"debug agent steps", "agent latency"
Reference
Services A–Z
Every service and major feature you're likely to see, including the ones used as wrong answers. Each card says what it is, when to pick it, and what it gets confused with. Tags follow the SAP-C03 lists: In scope on the in-scope list (or a feature of a listed service), Out of scope, Not listed on neither list.
Try it
Service flashcards
Read the situation, name the service in your head, then reveal.
Night before
Quick review: "If you see this, pick that"
Each row maps a common exam clue to the answer it usually points to.
Architecture and networking
Security and governance
Cost
Resilience and migration
Operations and AI
Acronyms A–Z
Every acronym used in this guide. Type any of them in the search box to jump here.