SAP-C03 · CHEAT SHEET
AWS Certified Solutions Architect - Professional
Key concepts grouped by official exam weight.
Download as PDF (free, no sign-up)1. Architecture Design and Implementation Built for the Cloud
25–25%Design architectures built for the cloud
- SQS standard queues and Lambda event source mappings deliver at least once, so consumers must be idempotent. FIFO queues order messages per message group ID and deduplicate with a deduplication ID or content-based deduplication; SNS FIFO topics fan out to SQS queues, not directly to Lambda. In Kinesis, the partition key decides the shard and the per-key order.
- A stateless application keeps session and user data outside the instance or task, in ElastiCache, DynamoDB, EFS or S3, so servers can be replaced and scaled without losing anything. DynamoDB global tables accept writes in every Region and TTL removes expired items; in an ElastiCache global datastore only the primary cluster accepts writes. S3 presigned URLs let clients upload files without going through the server.
- A typical serverless pipeline ingests into Kinesis Data Streams, delivers to S3 with Amazon Data Firehose (which converts JSON to Parquet with a Glue Data Catalog schema and partitions dynamically) and queries with Athena. Glue job bookmarks prevent reprocessing data; in Athena, columnar formats, partitioning and partition projection reduce scanned data and maintenance.
- Step Functions Standard workflows are exactly-once, support .sync and .waitForTaskToken and fit long, auditable processes; Express workflows are for short, high-volume events, at-least-once (asynchronous) or at-most-once (synchronous), and the type cannot change after creation. Lambda durable functions write the workflow in code with checkpointed steps and waits that incur no compute charges.
- To receive events from other accounts, the target bus needs a resource-based policy that allows specific accounts or the whole organization by its ID, and sender rules target the bus ARN with an IAM role. EventBridge does not forward events received from another account on to a third account. Global endpoints fail over to a second Region with a Route 53 health check and recover automatically when replication is enabled.
Integrate AI and ML services into architectures
- Amazon Bedrock cross-Region inference uses Geographic profiles (US, EU, APAC), suited to data residency, or Global profiles; if SCPs block Regions, they must allow every destination Region in the profile. Tagged application inference profiles enable chargeback and usage metrics per team, and an interface VPC endpoint for bedrock-runtime keeps traffic off the internet.
- Amazon Bedrock AgentCore splits an agent into parts: Runtime hosts agents from any framework in isolated microVM sessions, Gateway turns OpenAPI APIs, Lambda functions and services into MCP tools with inbound and outbound authentication, Identity handles OAuth 2.0 on behalf of users with a token vault, and Memory keeps short- and long-term context. Knowledge Bases provides RAG with metadata filters.
- SageMaker AI offers real-time inference (persistent, low-latency endpoint), serverless inference (intermittent traffic, scales to zero, no GPUs or VPC configuration), asynchronous inference (queued requests, large payloads in S3, long processing, scale to zero and SNS notifications) and batch transform (offline datasets without an endpoint). Multi-model endpoints share a fleet across many models, and SageMaker AI Runtime supports PrivateLink.
- Amazon Bedrock Knowledge Bases implements RAG: it ingests documents, splits them into chunks (fixed, hierarchical, semantic or none), embeds them and stores them in a vector store. Syncing is incremental. At retrieval, results can be filtered with metadata from .metadata.json files, and hybrid search (vector plus raw text) is available for RDS, OpenSearch Serverless and MongoDB vector stores.
Design hybrid cloud integration patterns
- For maximum resiliency, Direct Connect uses connections on separate devices in more than one location; high resiliency is one connection per location. On a transit gateway, static routes beat propagated routes and Direct Connect gateway routes beat VPN routes, so a BGP VPN is an automatic backup. ECMP across VPN tunnels needs dynamic routing and the same ASN, and does not work between different attachment types.
- DataSync moves files and objects between on-premises storage (NFS, SMB, HDFS, object) and AWS services with scheduled, incremental tasks; with a VPC service endpoint the traffic avoids the internet. S3 File Gateway presents S3 as an NFS or SMB share with a local cache, and Transfer Family provides managed SFTP, FTPS, FTP and AS2 into S3 or EFS with post-upload workflows.
- To modernize legacy systems without changing them, API Gateway REST APIs transform bodies with VTL mapping templates on non-proxy integrations (HTTP APIs transform only parameters), and passthroughBehavior decides whether unmapped content types are rejected. Amazon MQ is managed ActiveMQ or RabbitMQ using JMS, AMQP or MQTT without rewriting messaging code, and Lambda can consume and transform its messages.
- Route 53 VPC Resolver (formerly Route 53 Resolver) handles hybrid DNS with inbound endpoints (queries from on premises) and outbound endpoints with conditional forwarding rules that can be shared across accounts; forwarding to the VPC+2 address from on premises is not supported. Split-horizon uses same-named public and private zones; a missing private record returns NXDOMAIN, and a Resolver rule overrides a private zone of the same name.
- Outposts brings AWS-managed capacity to customer sites as an extension of a Region, with a service link to the Region and a local gateway to the on-premises network. Local Zones place resources near large population centers. ECS Anywhere registers your servers as external instances of an ECS cluster, EKS Hybrid Nodes joins them to EKS over a reliable connection, and EKS Anywhere is customer-managed and can run disconnected.
Design network architectures
- VPC peering links two VPCs (also across accounts and Regions) with non-overlapping CIDRs, is not transitive, cannot use the other VPC's VPN, Direct Connect, NAT or internet gateway, and needs routes on both sides. AWS PrivateLink exposes a service behind a Network Load Balancer to consumer VPCs through interface endpoints, one-way and with overlapping CIDRs allowed. Transit Gateway adds transitive routing at scale.
- Amazon VPC IPAM organizes IP space into scopes, hierarchical pools per Region and environment and VPC allocations, monitors overlap and compliance, and with AWS Organizations shares pools through AWS RAM. An SCP denying ec2:CreateVpc and ec2:AssociateVpcCidrBlock without ec2:Ipv4IpamPoolId enforces it. Public subnets route to the internet gateway, private subnets egress through NAT, and subnet sharing centralizes the network.
- Centralized inspection through Transit Gateway needs appliance mode on the inspection VPC attachment so traffic is symmetric. Gateway Load Balancer transparently inserts third-party appliances for all IP traffic. AWS WAF protects resources such as ALB, CloudFront and API Gateway at layer 7. Security groups are stateful and allow-only, network ACLs are stateless and support deny rules, and Flow Logs only record traffic.
- An IPv4 VPC can move to dual stack by associating IPv6 CIDR blocks with the VPC and subnets; IPv4 cannot be disabled and there is no migration from IPv4-only to IPv6-only subnets. Outbound-only IPv6 uses an egress-only internet gateway. IPv6-only workloads reach IPv4 services with DNS64 on the subnet and NAT64 on a NAT gateway. A dualstack ALB publishes an AAAA record.
- Route 53 offers failover (active-passive), latency, geolocation (continent, country or US state, with a default record for unmapped locations), geoproximity with bias, weighted, multivalue and IP-based routing; Evaluate Target Health on load balancer alias records steers traffic away from unhealthy resources. Global Accelerator provides static anycast IPs, TCP and UDP support, and regional failover that does not depend on DNS caching.
Design data, storage, and database architectures
- A data lake on Amazon S3 is cataloged in the AWS Glue Data Catalog and governed with AWS Lake Formation: column, row and cell permissions, LF-Tags to scale, and cross-account sharing through AWS RAM without copies. Redshift Spectrum queries S3 from Redshift through external tables that do not support updates. Apache Iceberg tables add record-level updates and deletes, schema evolution and snapshot isolation.
- Amazon MSK runs managed Apache Kafka that works with existing applications, and Redshift streaming ingestion reads from MSK or Kinesis Data Streams directly into materialized views. Amazon AppFlow moves SaaS data such as Salesforce into S3 or Redshift without code, on a schedule or on events, including private flows. On EMR, On-Demand primary and core nodes with Spot task nodes protect HDFS while cutting processing cost.
- Database choice follows the access pattern: DynamoDB for keyed access with millisecond latency at any scale and on-demand mode without capacity planning (no JOIN), DAX for microsecond eventually consistent reads, Aurora and RDS for relational OLTP with transactions and read-only replicas, and Redshift for analytics; an Aurora zero-ETL integration brings transactional data into Redshift in near real time.
- Aurora Global Database replicates at the storage layer to other Regions; managed failover recovers from a Regional outage and adds the old Region back as a secondary, and write forwarding lets a secondary accept writes. DynamoDB global tables are active-active: MREC replicates asynchronously and MRSC synchronously with an RPO of zero. An RDS Multi-AZ DB cluster gives two readable standbys within one Region.
- Choose storage by protocol and access pattern: FSx for Lustre is a parallel POSIX file system linked to S3 with automatic import and export; FSx for NetApp ONTAP serves the same data over NFS and SMB; EFS (NFS only, no Windows) uses Elastic throughput for unpredictable spikes and lifecycle policies to Infrequent Access; EBS Multi-Attach works only for io1/io2 in one Availability Zone.
2. Security, Compliance, and Governance
25–25%Design identity security architectures
- IAM Identity Center connects an external IdP with SAML 2.0 for sign-in and SCIM for provisioning users and groups; the SAML NameID must match the attribute mapped to Username, and synchronized attributes can be passed as session tags for ABAC. For external workloads such as GitHub Actions, OIDC federation exchanges a token for temporary role credentials.
- Cognito user pools authenticate users and issue OAuth 2.0 tokens; the client credentials grant with resource server scopes covers machine-to-machine access. Identity pools exchange those tokens for temporary AWS credentials and choose the role by rules or token claims. EKS Pod Identity maps roles to service accounts with session tags, but it does not support pods on Fargate, which use IRSA.
- Cross-account access requires both the role trust policy and the principal's identity policy to allow it. A unique external ID per customer prevents the confused deputy problem in third-party access. Permissions boundaries set a role's maximum permissions and let teams delegate role creation with the iam:PermissionsBoundary condition, and session policies restrict a single session without granting permissions.
- A data perimeter combines RCPs with aws:PrincipalOrgID to protect your resources from outside principals, SCPs and VPC endpoint policies with aws:ResourceOrgID to limit which resources can be reached, and exceptions for AWS services. IAM Access Analyzer finds resources shared outside the zone of trust in each Region and checks policies before deployment. Verified Access evaluates every request without a VPN.
- IAM Roles Anywhere gives temporary credentials to workloads outside AWS that authenticate with X.509 certificates from a CA registered as a trust anchor (AWS Private CA or an external CA). The role trusts rolesanywhere.amazonaws.com and can filter on certificate attributes such as aws:PrincipalTag/x509Subject/CN; imported CRLs must be re-imported, and the credential helper supports credential_process, update, serve and PKCS#11.
Design solutions to implement encryption and key management strategies
- Customer managed keys let you control the key policy and turn on automatic rotation with your own period, only on symmetric keys with KMS-generated material; keys with imported material rotate only on demand. AWS managed keys rotate yearly and cannot be shared. Multi-Region keys share key material and rotation is enabled on the primary; S3 Bucket Keys cut the number of KMS calls.
- ACM renews the certificates it issues automatically when they use DNS validation and are associated with a service; imported certificates do not renew, and CloudFront needs its certificate in us-east-1. AWS Private CA is shared across accounts with AWS RAM. Load balancers apply TLS 1.3 security policies and mutual TLS with a trust store, and an NLB TCP listener passes TLS through without decrypting it.
- AWS CloudHSM provides single-tenant, FIPS-validated HSMs under the customer's control. A KMS CloudHSM key store lets AWS services use keys generated in the cluster, which needs HSMs in two Availability Zones, but it supports neither automatic rotation nor multi-Region keys, so keys are rotated manually. CloudHSM also offloads TLS for web servers and its backups can be copied to another Region for DR.
- AWS KMS offers post-quantum asymmetric keys: ML-DSA (SIGN_VERIFY usage) for signatures and ML-KEM (ENCAPSULATE_DECAPSULATE usage) to establish shared secrets; the private key never leaves KMS unencrypted. These keys do not support imported material or custom key stores and are rotated manually with aliases. Calls to KMS are protected with hybrid post-quantum TLS, which you can verify in the CloudTrail tlsDetails field.
Design multi-account governance and compliance solutions
- AWS Control Tower builds a landing zone on AWS Organizations and applies controls per OU: preventive (SCPs and RCPs), detective (AWS Config rules) and proactive (CloudFormation hooks). Account Factory provisions standard accounts, and Account Factory Customization adds Service Catalog blueprints without pipelines to maintain. Recommended OUs include Security, Infrastructure, Workloads, Sandbox, Policy Staging and Suspended.
- An organization trail captures activity from every account, and a Config aggregator gives a read-only view of configuration and compliance. GuardDuty is delegated per Region. Security Lake normalizes logs to OCSF in buckets you own. Bedrock model invocation logging stores prompts and responses only in destinations in the same account and Region; CloudWatch Logs centralization copies them to a central account.
- AWS RAM shares resources such as subnets, transit gateways, Route 53 Resolver rules, IPAM pools and prefix lists with accounts, OUs or the whole organization. With sharing with Organizations enabled, no invitations are needed. In a shared VPC the owner controls routes, ACLs and NAT, and participants their own resources and security groups. Shared transit gateway attachments are accepted manually unless auto accept is on.
- SCPs cap the maximum permissions of principals in member accounts, and RCPs cap those of resources in those accounts, even against external identities. Neither grants permissions, neither affects the management account or service-linked roles, and an action must be allowed at every level above the account. Combined, for example with aws:ResourceOrgID and aws:PrincipalOrgID, they form a data perimeter.
- AWS Config organization rules and conformance packs are deployed per Region to every account, including new ones, and members cannot modify them; remediation uses Systems Manager Automation documents. Security Hub CSPM checks standards such as PCI DSS or CIS through Config rules; its central configuration manages accounts and Regions, and automation rules plus EventBridge automate the response.
Design architectures to meet compliance requirements
- Customer managed KMS keys let you control the key policy and lifecycle, and their use is logged in CloudTrail; AWS managed keys cannot be changed or shared. RDS cannot encrypt an existing instance in place, so you use a blue/green deployment or an encrypted snapshot copy. Lake Formation applies row, column and cell-level permissions and audits data access.
- For data sovereignty, AWS KMS offers imported key material, which can be deleted immediately and imported again, and external key stores, where keys are used in your own key manager outside AWS in exchange for owning availability. AWS Artifact provides AWS compliance reports and lets the management account accept agreements such as the BAA for the whole organization.
- CloudTrail log file integrity validation delivers signed digest files that prove whether a log changed, and S3 Object Lock in compliance mode stops anyone, even the root user, from deleting it. In Bedrock, InvokeModel is a management event, but InvokeAgent and Retrieve are data events that need advanced event selectors. CloudWatch Logs data protection policies mask sensitive data in invocation logs.
- For data residency, Control Tower offers a Region deny control per OU (CT.MULTISERVICE.PV.1) and another for the whole landing zone. Security Lake supports rollup Regions per geography, and S3 Cross-Region Replication copies new objects while Batch Replication copies existing ones. In Bedrock, geographic inference profiles keep processing inside a geography such as the EU.
- AWS Artifact provides AWS's own compliance reports (SOC, PCI, ISO), watermarked and meant to be shared with auditors through a secure channel, and lets the management account accept agreements such as the BAA for the whole organization (all features required), covering current and future accounts. Evidence about your own configuration comes from Security Hub CSPM and AWS Config.
Design security and responsible AI controls
- Amazon Bedrock Guardrails combines content filters, denied topics, word filters, sensitive information filters (block or mask, with regex), contextual grounding checks and Automated Reasoning. ApplyGuardrail evaluates any text without invoking a model, including models outside Bedrock. Bedrock policies in AWS Organizations enforce a numeric guardrail version on accounts and OUs without code changes.
- AgentCore Identity manages agent identities and credentials. Inbound, Runtime and Gateway accept IAM SigV4 or a JWT authorizer that validates audiences, clients and scopes from any OAuth 2.0 provider. Outbound, credential providers keep OAuth tokens (2LO, or 3LO with user consent) and API keys in a KMS-encrypted token vault, outside the agent code.
- Human approvals are designed with Step Functions Standard workflows and .waitForTaskToken tasks, which wait without compute until the approval system calls SendTaskSuccess or SendTaskFailure from the same account. The wait has no default timeout: TimeoutSeconds plus a Catch on States.Timeout allow escalation. Express workflows do not support callbacks. Bedrock agents can request user confirmation.
3. Cost-Optimized Architecture Design
15–15%Design architectures based on AWS pricing models
- Compute Savings Plans apply across families, Regions, Fargate and Lambda; EC2 Instance Savings Plans to one family in one Region; Database Savings Plans across database engines and services. Neither Savings Plans nor regional RIs reserve capacity: zonal RIs or On-Demand Capacity Reservations do, the latter with no term for immediate use. Spot suits interruption-tolerant workloads.
- Step Functions Standard workflows are billed per state transition and suit long-running, exactly-once processes; Express workflows are billed by executions, duration and memory and suit short, idempotent, high-volume work. The type cannot change after creation, and a Standard workflow can start nested Express ones. Lambda on arm64 (Graviton) improves price performance when every dependency has an arm64 build.
- Compute Optimizer gives rightsizing recommendations for EC2, Auto Scaling groups, EBS, Lambda, ECS on Fargate and databases, organization-wide when the management account opts in; EC2 memory needs the CloudWatch agent. Scheduled scaling combines with target tracking. Fargate Spot warns with a SIGTERM two minutes before stopping a task; a base on FARGATE guarantees regular tasks.
- Amazon Bedrock offers Reserved, Priority, Standard and Flex tiers (a discount for latency-tolerant work), chosen with service_tier. Batch inference processes S3 files asynchronously at a lower price on supported models. Prompt caching cuts the cost of long repeated contexts when the static prefix comes first. SageMaker AI asynchronous and serverless inference can scale to zero.
Optimize data transfer and network costs
- CloudFront Origin Shield adds a caching layer that consolidates requests from all caches to the origin; enable it in the Region with the lowest latency to the origin, and it adds charges. The cache policy defines the cache key (fewer values, more hits) and the origin request policy forwards data without fragmenting it. Origin groups fail over only GET, HEAD and OPTIONS requests.
- Gateway endpoints for S3 and DynamoDB have no additional charge and avoid NAT gateways, which bill per hour and per GB processed; interface endpoints bill per endpoint and per GB. Transfer between Availability Zones is charged in both directions, so keep the resources that exchange the most traffic in the same zone and avoid crossing Regions unnecessarily.
- S3 replication filters by prefix or tag, can write replicas into a colder storage class, and does not copy existing objects: S3 Batch Replication does that. S3 RTC adds a 15-minute objective that pays off only for strict RPOs. DynamoDB global tables are active-active but bill replicated writes in every Region; for RPOs of hours, AWS Backup with cross-Region copy is enough.
- Transit Gateway bills per attachment hour and per traffic processed; VPC peering has no processing charge and is the cheapest option for a few VPCs with non-overlapping CIDRs. PrivateLink bills the endpoint owner per hour and per GB, supports overlapping CIDRs and consumer-initiated connections only. Centralizing interface endpoints behind Transit Gateway needs your own private hosted zone.
- Amazon Bedrock cross-Region inference uses inference profiles: geographic profiles (EU, US, APAC) keep processing inside that geography and global profiles route to any commercial Region at a lower token price. There is no additional routing cost, pricing follows the source Region and data stays on the AWS network. Region-deny SCPs must allow the destination Regions or use bedrock:InferenceProfileArn.
Design cost allocation and visibility solutions
- Cost allocation tags (user-defined or AWS-generated) appear in Cost Explorer and reports only after they are activated from the management account, and they are not retroactive unless a backfill is requested. AWS Organizations account tags apply to all usage in the account, including credits and refunds, and combine with resource tags to split costs by business unit, project and environment.
- AWS Data Exports delivers the fixed-schema CUR 2.0 to Amazon S3 with resource IDs, hourly granularity and SQL row and column filters; with the Athena integration it arrives as Parquet with a template that creates the Glue catalog. Cost Explorer analyzes costs without code, and the management account enables recent hourly and resource-level data in its preferences.
- AWS Organizations tag policies standardize capitalization and values and, with enforced_for, block noncompliant tagging, but they do not require missing keys; for that use SCPs with aws:RequestTag and Null, or required tag keys with the AWS::TagPolicies::TaggingComplianceValidator hook in CloudFormation. The AWS Config required-tags rule detects untagged resources but does not prevent their creation.
- Cost categories group accounts, tags and services into custom values, and their split charge rules (proportional, fixed or even) allocate shared costs, with results shown only on their details page. Split cost allocation data brings per-task ECS and per-pod EKS costs into CUR. For Bedrock, tags on the IAM principal or on an application inference profile allocate costs and tokens.
- AWS Budgets alerts on actual or forecasted thresholds and can run automatic or approved actions, such as applying an IAM policy or, from the management account, an SCP to another account. AWS Cost Anomaly Detection uses machine learning to detect unusual spend; its AWS managed monitors include new accounts or values automatically, and alerts combine absolute and percentage thresholds.
Optimize storage and database costs
- S3 Intelligent-Tiering moves objects between tiers by access, with no retrieval fees, and its optional Archive tiers are asynchronous; it does not monitor objects under 128 KB. S3 Lifecycle moves objects to cheaper classes, but each transition is billed and archived objects carry extra metadata, so aggregate small objects first. In EFS, lifecycle policies apply to the whole file system.
- With gp2, baseline IOPS grow with size and rely on burst credits; gp3 costs less per GiB and lets you provision IOPS and throughput without adding capacity. io2 Block Express gives sub-millisecond latency and the highest durability for critical databases. Elastic Volumes changes type, IOPS or size (increase only) without detaching, and Compute Optimizer recommends type, size and IOPS from usage.
- DynamoDB on-demand charges per request and scales by itself; provisioned mode suits steady, predictable workloads. Aurora I/O-Optimized removes I/O charges when I/O is a large part of the bill, and Aurora Serverless v2 scales with load and can pause at 0 ACUs. RDS reserved DB instances are tied to engine, type and Region, while Database Savings Plans keep applying across engine, family and Region changes.
- AWS Backup sets retention and can move backups of supported resources (EFS, DynamoDB with advanced features, EBS through archival) to cold storage, where they must stay at least 90 days; AWS Organizations backup policies apply plans through inheritance. Amazon Data Lifecycle Manager automates EBS snapshots and AMIs and archives snapshots only on monthly or yearly schedules.
- Lambda log-level filtering (with the JSON format) avoids ingesting logs that are not needed. CloudWatch Logs keeps data indefinitely unless a retention is set, and the Essentials class (formerly Infrequent Access) lowers ingestion cost but has no metric filters, subscriptions, Live Tail or EMF, and the class cannot be changed later. For metrics, use EMF and avoid high-cardinality dimensions.
4. Resilience, Migration, and Business Continuity
15–15%Design migration strategies by using AWS frameworks
- AWS Transform replaces Application Discovery Service and Migration Hub, which are closed to new customers. Its discovery tool inventories VMware vCenter, Hyper-V and imported servers and, with OS access, collects databases and network connections; it also accepts RVTools, modelizeIT or MPA data. Planning groups applications, generates move groups from technical and business dependencies and builds waves.
- AWS Transform MGN (formerly AWS Application Migration Service) continuously replicates physical, virtual or cloud servers at block level and launches them on EC2 with cutovers of minutes. Its replication, launch and post-launch templates let you send data over private IP (VPN or Direct Connect), keep the private IP, right-size the instance type and automate actions such as installing the SSM Agent.
- AWS DMS migrates databases with full load plus CDC to keep the target in sync and reduce downtime; DMS Serverless removes sizing and patching of replication instances, and DMS Schema Conversion assesses and converts schemas between engines, but not data. AWS DataSync copies NFS, SMB, HDFS or object data to S3, EFS or FSx with verification, scheduling, incremental transfers and private VPC endpoints.
- Replatforming (lift, tinker and shift) moves an application with some optimization, such as moving to a managed or serverless service, or from VMs to containers without code changes. Amazon RDS manages backups, patching and recovery, with Multi-AZ for failover; AWS Fargate runs ECS containers with no servers to manage, and Lambda with S3 events and EventBridge Scheduler replaces short jobs on servers.
- The strangler fig pattern migrates a monolith to microservices incrementally: a proxy layer (for example, Amazon API Gateway) routes each path to the monolith or to the new service. An anti-corruption layer translates internal calls to new interfaces without changing callers, and the transactional outbox avoids events from rolled-back transactions.
Design disaster recovery architectures
- The four DR strategies go from lowest to highest cost: backup and restore, pilot light (replicated data always on, servers switched off), warm standby (scaled-down copy running) and multi-site active/active. Continuous replication does not protect against data corruption, so point-in-time recovery and copies in another account are also needed.
- AWS Elastic Disaster Recovery continuously replicates servers at block level to a low-cost staging area and launches recovery instances in minutes, from the latest state or from an earlier snapshot. By default it replicates over the internet; the private IP option uses VPN or Direct Connect. Between Regions, failback uses reversed replication.
- AWS Resilience Hub defines tiered resiliency policies with RTO and RPO per disruption type, assesses each application against its policy and recommends alarms, SOPs and FIS experiments. SOPs are delivered as a CloudFormation template with SSM Automation documents, and daily assessment with drift notification alerts through SNS.
- For failover between Regions, Route 53 health checks give automatic DNS failover, Global Accelerator provides static anycast IP addresses and avoids DNS caching, and ARC routing controls are data plane switches for manual failover with safety rules. ARC Region switch orchestrates multi-account recovery and measures the actual recovery time.
- DR validation is automated by running AWS Resilience Hub assessments with StartAppAssessment after each published version and blocking the pipeline when the result is PolicyBreached. Recommended FIS experiments are deployed as a CloudFormation template and run in CI/CD. For planned regional rotation, an Aurora Global Database switchover has an RPO of 0.
Design solutions to meet business continuity and resiliency requirements
- AWS Organizations backup policies are inherited from the root and OUs and merged into one immutable effective plan per account; they merge only when they share the plan name. Cross-account copies need customer managed keys for resources not fully managed by AWS Backup. A delegated administrator manages policies and monitors jobs without the management account.
- To measure recovery, AWS FIS runs actions such as aws:rds:failover-db-cluster and produces PDF reports with CloudWatch dashboards as evidence. Route 53 can base health checks on CloudWatch alarms when resources have only private IP addresses. ARC zonal autoshift requires practice runs with an outcome alarm.
- A resilient architecture heals itself: Auto Scaling replaces instances that fail ELB health checks when they are turned on, and static stability provisions capacity for the loss of a zone before the failure, without depending on the control plane. Limited retries with exponential backoff and jitter at one layer and circuit breakers prevent cascading failures.
- S3 Object Lock in compliance mode and EBS snapshot lock in compliance mode prevent anyone, even the root user, from deleting data before the retention expires; governance mode can be bypassed with permissions. An AWS Backup logically air-gapped vault comes with Vault Lock compliance mode and can be shared through AWS RAM with a recovery account.
- Chaos engineering with AWS FIS starts from a steady-state hypothesis that becomes CloudWatch alarms used as stop conditions. Templates define actions, targets and a role; the scenario library offers predefined events such as AZ Availability: Power Interruption, and multi-account experiments are orchestrated from a central account using AZ IDs.
5. Operational Excellence and Automation
20–20%Design infrastructure as code
- CloudFormation is declarative, and AWS CDK lets you define infrastructure in programming languages that synthesize to CloudFormation; each target environment is bootstrapped with cdk bootstrap. StackSets with service-managed permissions deploys to OUs and to new accounts automatically. The IaC generator brings in existing resources, and drift detection reveals changes made outside CloudFormation.
- Policy as code combines layers: CDK aspects validate at synthesis time, CloudFormation Hooks (and Control Tower proactive controls) inspect resources before provisioning in WARN or FAIL mode, SCPs block every creation path, and AWS Config rules with SSM Automation remediation correct later drift.
- CloudFormation dynamic references (secretsmanager and ssm-secure) keep secrets out of templates: the value is resolved only when the resource is created or updated, ssm-secure cannot read parameters from other accounts, and a secret in another account is referenced by its full ARN with a customer managed KMS key. For RDS and Aurora, ManageMasterUserPassword lets RDS create and rotate the secret.
- StackSets deploy one template across many accounts and Regions. With service-managed permissions, CloudFormation creates the roles, supports automatic deployment to new accounts in an OU and can be run from a delegated administrator, but it never deploys to the management account. Region concurrency and per-Region failure tolerance control the blast radius of an operation.
Design CI/CD and DevSecOps pipelines
- CodePipeline V2 pipelines support triggers filtered on branches, file paths and Git tags, QUEUED and PARALLEL execution modes besides SUPERSEDED, and stage conditions (Entry, On Failure, On Success) with rules such as CloudWatchAlarm and fail, skip or rollback results. CodeBuild publishes test reports, and CodeDeploy deploys to EC2 and to registered on-premises servers.
- ECR basic scanning covers operating system vulnerabilities; enhanced scanning uses Amazon Inspector to also scan programming language packages, continuously, and emits EventBridge events. In CodePipeline, the InspectorScan action fails the pipeline when severity thresholds are exceeded and produces an SBOM. AWS Signer signs images and Lambda packages, and the Enforce policy blocks unsigned code.
- A cross-account pipeline uses a customer managed KMS key in its Region, bucket and key policies for the target accounts, and roles in those accounts that the service role assumes. An artifact cannot move between two accounts if neither is the pipeline account. Cross-Region actions need an artifact bucket and key in each Region, owned by the pipeline account.
- Amazon Bedrock Evaluations offers automatic evaluations with computed metrics, judge-based evaluations (built-in metrics such as Correctness or Harmfulness plus custom metrics), evaluations with your own human workers, and RAG evaluations that are retrieve-only or retrieve-and-generate. Models outside Bedrock can be evaluated by supplying their responses in the dataset.
- In ECS, rolling updates are protected by the deployment circuit breaker and by CloudWatch alarms that roll back to the last completed deployment; blue/green, canary and linear strategies shift traffic with bake times and keep the previous revision for rollback. CodeDeploy offers canary and linear configurations for Lambda and ECS, but with an NLB it supports only all-at-once.
Design observability and monitoring solutions
- AWS Distro for OpenTelemetry lets you instrument once and send correlated traces and metrics to X-Ray, CloudWatch and Managed Service for Prometheus, with a sidecar collector on ECS and a managed layer on Lambda. X-Ray recommends OpenTelemetry over its SDK. Transaction Search ingests every span as logs, and the embedded metric format extracts metrics from log events.
- Container Insights collects metrics and logs from ECS and EKS, including Fargate; ECS enhanced observability goes down to container level and is turned on with an account setting for new clusters and update-cluster-settings for existing ones. On EKS, the CloudWatch Observability add-on installs the agent and Fluent Bit and enables Container Insights and Application Signals. The agent can also scrape Prometheus metrics.
- Cross-account subscriptions send events to a CloudWatch Logs destination that wraps a Kinesis or Firehose stream; the destination must be in the same Region as the log group, and its policy can allow the whole organization with aws:PrincipalOrgID. Centralization rules copy logs from the whole organization into a central account. To archive years of logs, S3 Lifecycle moves them to S3 Glacier Deep Archive.
- CloudWatch alarms can notify SNS, run EC2 and Auto Scaling actions or create Systems Manager OpsItems. Composite alarms combine alarm states with AND, OR and NOT to reduce noise and support suppressor alarms during deployments. Anomaly detection follows a metric's seasonality, and the missing data treatment (breaching, notBreaching, ignore, missing) avoids false positives.
- SageMaker AI endpoints publish ModelLatency and OverheadLatency to separate model latency from service overhead. Amazon Bedrock publishes latency, throttling and token metrics per ModelId in AWS/Bedrock, and invocation logging stores requests and responses in destinations in the same account and Region. AgentCore Observability sends agent sessions, latency, tokens, errors and traces to CloudWatch through ADOT.
Design operational automation and resilient systems
- Service Quotas request templates automatically request increases for every new account in the organization, and Service Quotas creates CloudWatch alarms near the limit. Predictive scaling forecasts daily and weekly patterns and pre-launches instances, but it never scales in. AWS Health events reach EventBridge, which can start Systems Manager Automation runbooks as the operational response.
- Automated remediation needs brakes: in Systems Manager Automation, concurrency caps how many resources are touched at once and the error threshold stops sending to new targets (with 0, after the first failure). The ECS circuit breaker rolls back to the last COMPLETED deployment, and retries should use exponential backoff with jitter, a maximum number of attempts and a single layer.
- Target tracking holds a metric that changes in proportion to capacity (for queues, backlog per instance with metric math). Predictive scaling forecasts daily or weekly patterns, starts in forecast only mode, can pre-launch instances and never scales in, so it is paired with a dynamic policy. Scheduled actions change minimum and maximum capacity at set times.
- Auto Scaling ignores Elastic Load Balancing health checks by default: turn them on for the group and set a grace period so bootstrapping instances are not terminated. ECS replaces tasks whose essential container fails the health check in the task definition (not one only in the Dockerfile). For Kinesis streams, Lambda can cap retries, bisect batches and send discarded records to an on-failure destination.
- Quick Setup patch policies apply to chosen OUs and Regions with separate scan and install schedules. An EventBridge rule can start an Automation runbook, and aws:approve pauses it until designated principals approve, though it does not support multi-account runs. Those need a delegated administrator, a central administration role and an execution role in every target account.