AWS Well-Architected Framework
Six pillars, design principles, and practical AWS services and patterns for building and reviewing workloads
AWS Well-Architected Framework ┌────────────────┬──────────┬─────────────┬─────────────┬──────────┬───────────────┐ │ Operational │ Security │ Reliability │ Performance │ Cost │ Sustainability│ │ Excellence │ │ │ Efficiency │ Optimiz. │ │ ├────────────────┴──────────┴─────────────┴─────────────┴──────────┴───────────────┤ │ Each pillar = design principles + best practices + questions (per workload) │ ├──────────────────────────────────────────────────────────────────────────────────┤ │ Tooling: Well-Architected Tool · Lenses · Trusted Advisor · Config · Security Hub│ └──────────────────────────────────────────────────────────────────────────────────┘
Overview Intro

The Well-Architected Framework is AWS's set of design principles, best practices, and review questions for evaluating architectures. It is not a certification or a checklist to pass: it is a structured way to find risks in a workload, make trade-offs explicit, and prioritise improvements.

The six pillars

PillarCore question
Operational ExcellenceCan we run, observe, and continuously improve this workload?
SecurityAre data, systems, and assets protected, and can we detect and respond to events?
ReliabilityDoes it recover from failure and meet demand without disruption?
Performance EfficiencyAre we using the right resources efficiently as demand and technology change?
Cost OptimizationAre we delivering business value at the lowest price point?
SustainabilityAre we minimising the environmental impact of running the workload?
Sustainability was added in December 2021, making the original five pillars into six. The framework is revised periodically, so check the current whitepaper for the exact question wording.

General design principles

  • Stop guessing capacity needs: scale on measured demand
  • Test systems at production scale, then tear them down
  • Automate to make architectural experimentation cheap
  • Allow evolutionary architectures: design for change
  • Drive architectures using data (metrics, load tests, game days)
  • Improve through game days: simulate events and learn

Shared responsibility

AWS is responsible for security of the cloud (hardware, facilities, managed-service internals). You are responsible for security in the cloud (IAM, data, network config, guest OS and application). The split moves toward AWS as you move from EC2 to containers to serverless.
Operational Excellence Pillar 1

The ability to support development, run workloads effectively, gain insight into their operation, and continuously improve supporting processes and procedures.

Design principles

  • Perform operations as code (IaC for infrastructure, runbooks, and procedures)
  • Make frequent, small, reversible changes
  • Refine operations procedures frequently
  • Anticipate failure (pre-mortems, game days)
  • Learn from all operational failures and share what you learn

Best-practice areas

AreaWhat it covers
OrganizationShared ownership, clear priorities, team roles, risk tolerance
PrepareTelemetry design, IaC, CI/CD, readiness checks, runbooks
OperateWorkload health, KPIs, alerting, event response, dashboards
EvolvePost-incident analysis, feedback loops, continuous improvement

Key services

ServiceRole
CloudFormation / CDK / TerraformInfrastructure as code, repeatable environments
CodePipeline / CodeDeployCI/CD with blue/green, canary, and automatic rollback
CloudWatchMetrics, logs, alarms, dashboards, composite alarms
X-Ray / CloudWatch Application SignalsDistributed tracing and service-level objectives
Systems ManagerRunbook automation, patching, Parameter Store, Session Manager
EventBridgeEvent-driven remediation and automation

Example: alarm on user-facing symptoms, not just causes

Shell
# Alarm on ALB 5xx rate, notify an SNS topic
aws cloudwatch put-metric-alarm \
  --alarm-name checkout-alb-5xx \
  --namespace AWS/ApplicationELB \
  --metric-name HTTPCode_Target_5XX_Count \
  --dimensions Name=LoadBalancer,Value=app/checkout/50dc6c495c0c9188 \
  --statistic Sum --period 60 --evaluation-periods 3 \
  --threshold 20 --comparison-operator GreaterThanThreshold \
  --treat-missing-data notBreaching \
  --alarm-actions arn:aws:sns:us-east-1:111122223333:oncall

Operating areas in detail

AreaTopicsSummary
OrganizationOrganization priorities · Operating model · Organizational cultureTeams share an understanding of the workload and business goals. Evaluate internal and external customer needs, governance and compliance obligations, and keep a risk registry. Every workload, process and procedure has a named owner; define team responsibilities and agreements, and provide mechanisms to request changes and exceptions. Leadership sponsors best practices; encourage experimentation, learning time, and diverse perspectives. Review priorities regularly. Control Tower and Organizations support centralised governance.
PrepareImplement observability · Design for operations · Mitigate deployment risks · Operational readiness and change managementDesign workloads to emit metrics, logs, events and traces, tied to KPIs that matter to the business. Treat the whole stack as code (CloudFormation), use consistent tagging, and run pre-mortems. Prefer frequent, small, reversible changes with fast feedback and a rollback plan. Use operational readiness reviews with checklists, runbooks (routine tasks) and playbooks (issue resolution) before go-live.
OperateUtilizing workload observability · Understanding operational health · Responding to eventsSet baselines and alert thresholds, and watch for deviations. Define expected outcomes and metrics for both the workload and the operations activities around it (deployments, incident response). Every alert has an owner and a process; prioritise by customer and business impact, define escalation paths, and automate scripted responses to well-understood events. Dashboards are tailored to audience: customer, business, developers, operations.
EvolveLearn, share, and improveDedicate time to continuous small improvements. Run post-incident analysis on every customer-impacting event, identify contributing factors and preventive actions, and share lessons across teams. Log analytics on S3 with Glue and Athena can reveal trends. Provide safe sandbox, dev, test and production environments with increasing operational control.

Best practices by topic

TopicBest practices
Organization priorities (OPS01)BP01 Evaluate external customer needs · BP02 Evaluate internal customer needs · BP03 Evaluate governance requirements · BP04 Evaluate compliance requirements · BP05 Evaluate threat landscape · BP06 Evaluate trade-offs while managing benefits and risks
Operating modelNo numbered best practices. Describes centralised, decentralised and federated organisation models and the team-level path towards "you build it, you run it" with two-pizza teams that own a workload end to end.
Organizational culture (OPS03)BP01 Provide executive sponsorship · BP02 Empower team members to take action when outcomes are at risk · BP03 Encourage escalation · BP04 Timely, clear and actionable communications · BP05 Encourage experimentation · BP06 Encourage team members to maintain and grow their skill sets · BP07 Resource teams appropriately
Implement observability (OPS04)BP01 Identify key performance indicators · BP02 Implement application telemetry · BP03 Implement user experience telemetry · BP04 Implement dependency telemetry · BP05 Implement distributed tracing
Design for operations (OPS05)BP01 Use version control · BP02 Test and validate changes · BP03 Use configuration management systems · BP04 Use build and deployment management systems · BP05 Perform patch management · BP06 Share design standards · BP07 Implement practices to improve code quality · BP08 Use multiple environments · BP09 Make frequent, small, reversible changes · BP10 Fully automate integration and deployment
Mitigate deployment risks (OPS06)BP01 Plan for unsuccessful changes · BP02 Test deployments · BP03 Employ safe deployment strategies · BP04 Automate testing and rollback
Operational readiness and change management (OPS07)BP01 Ensure personnel capability · BP02 Ensure a consistent review of operational readiness · BP03 Use runbooks to perform procedures · BP04 Use playbooks to investigate issues · BP05 Make informed decisions to deploy systems and changes · BP06 Create support plans for production workloads
Utilizing workload observability (OPS08)BP01 Analyse workload metrics · BP02 Analyse workload logs · BP03 Analyse workload traces · BP04 Create actionable alerts · BP05 Create dashboards
Understanding operational health (OPS09)BP01 Measure operations goals and KPIs with metrics · BP02 Communicate status and trends to ensure visibility into operations · BP03 Review operations metrics and prioritise improvement
Responding to events (OPS10)BP01 Use a process for event, incident and problem management · BP02 Have a process per alert · BP03 Prioritise operational events based on business impact · BP04 Define escalation paths · BP05 Define a customer communication plan for service-impacting events · BP06 Communicate status through dashboards · BP07 Automate responses to events
Learn, share and improve (OPS11)BP01 Have a process for continuous improvement · BP02 Perform post-incident analysis · BP03 Implement feedback loops · BP04 Perform knowledge management · BP05 Define drivers for improvement · BP06 Validate insights · BP07 Perform operations metrics reviews · BP08 Document and share lessons learned · BP09 Allocate time to make improvements
When responding to events, replace failed components with known-good versions instead of repairing them in place, and analyse the failed resource out of band. Every alert should have an owner and a process.
Prefer deployments you can undo in minutes. Canary or blue/green with automatic rollback on alarm turns a bad release into a non-event.
Security Pillar 2

The ability to protect data, systems, and assets, taking advantage of cloud technologies to improve your security posture.

Design principles

  • Implement a strong identity foundation (least privilege, separation of duties, no long-lived credentials)
  • Maintain traceability: monitor, alert, and audit actions and changes
  • Apply security at all layers (edge, VPC, subnet, load balancer, instance, OS, application)
  • Automate security best practices
  • Protect data in transit and at rest
  • Keep people away from data
  • Prepare for security events with incident response runbooks and simulations

Best-practice areas

AreaKey services and practices
Security foundationsMulti-account with AWS Organizations, Control Tower, SCPs, separate accounts per environment
Identity and access managementIAM Identity Center (SSO), IAM roles, permission boundaries, IAM Access Analyzer, MFA
DetectionCloudTrail (org trail), GuardDuty, Security Hub, Config, Inspector, Detective
Infrastructure protectionVPC design, security groups, WAF, Shield, Network Firewall, Systems Manager Session Manager
Data protectionKMS, Secrets Manager, ACM, S3 Block Public Access, Macie, TLS everywhere
Incident responsePlaybooks, isolation automation, forensic accounts, snapshots, regular exercises
Application securityThreat modelling, dependency scanning, SAST/DAST in the pipeline, signed artifacts

Example: least-privilege policy with conditions

JSON
{
  "Version": "2012-10-17",
  "Statement": [{
    "Effect": "Allow",
    "Action": ["s3:GetObject"],
    "Resource": "arn:aws:s3:::my-app-data/reports/*",
    "Condition": {
      "Bool": { "aws:SecureTransport": "true" },
      "StringEquals": { "aws:PrincipalOrgID": "o-exampleorgid" }
    }
  }]
}

Baseline quick wins

Shell
# Block public access on every bucket in the account
aws s3control put-public-access-block --account-id 111122223333 \
  --public-access-block-configuration \
  BlockPublicAcls=true,IgnorePublicAcls=true,BlockPublicPolicy=true,RestrictPublicBuckets=true

# Require IMDSv2 on an instance
aws ec2 modify-instance-metadata-options --instance-id i-0abc123 \
  --http-tokens required --http-endpoint enabled

# Enable GuardDuty
aws guardduty create-detector --enable

Identity and access management (SEC02, SEC03)

Get the right people and machines the right access under the right conditions. Two identity types: human (workforce, third parties, application users) and machine (EC2, Lambda, tools, external systems).

IDBest practice
SEC02-BP01Use strong sign-in mechanisms
SEC02-BP02Use temporary credentials
SEC02-BP03Store and use secrets securely
SEC02-BP04Rely on a centralised identity provider
SEC02-BP05Audit and rotate credentials periodically
SEC02-BP06Employ user groups and attributes
SEC03-BP01Define access requirements
SEC03-BP02Grant least-privilege access
SEC03-BP03Establish an emergency access process
SEC03-BP04Reduce permissions continuously
SEC03-BP05Define permission guardrails for your organisation
SEC03-BP06Manage access based on lifecycle
SEC03-BP07Analyse public and cross-account access
SEC03-BP08Share resources securely within your organisation
SEC03-BP09Share resources securely with a third party

Policy types

TypeWhat it does
Identity-based (managed / inline)Attached to users, groups or roles. Managed policies (AWS-managed or customer-managed) are preferred; create customer-managed policies for least privilege. Inline policies are tied one-to-one to an identity and deleted with it.
Resource-basedAttached to a resource, such as an S3 bucket policy; can grant to principals in the same or another account.
Permissions boundarySets the maximum permissions an identity can have, so developers can create roles without escalating their own privilege.
ABACGrants permissions by matching tags on principals and resources, so new resources inherit access without policy edits.
Organizations SCPMaximum permissions for accounts in an OU. Limits but never grants.
Session policyLimits the permissions of an assumed-role or federated session. Limits but never grants.

Detection (SEC04)

Detection has two halves: spotting unexpected configuration changes (in templates and pipelines before deployment, in deployed environments, and in running workloads outside a known deployment or scaling event) and spotting unexpected behaviour (GuardDuty for malicious activity, plus alerts on mutating API calls the workload shouldn't make and calls that change security posture). Review detection mechanisms regularly against internal and external policy.

Best practiceRisk if absentSummary
SEC04-BP01 Configure service and application loggingHighRetain security event logs for audit, investigation and compliance. Baseline sources: CloudTrail (an org trail with S3, or CloudTrail Lake with up to seven years of retention and SQL queries), VPC Flow Logs and Route 53 Resolver query logs. Add ELB, WAF, Config, GuardDuty findings, EKS audit, and OS/application logs. Store in S3 (cheap, lifecycle, Athena) or CloudWatch Logs (Logs Insights); typically 3–12 months queryable and up to 7 years retained. Avoid keeping logs forever or deleting too soon, open access to logs, and logging everything "just in case".
SEC04-BP02 Capture logs, findings and metrics in standardised locationsMediumAggregate security data into a separate Log Archive account (prevents tampering from compromised workloads, gives one integration point and one retention model). Run analytics from a separate Security Tooling account with cross-account read. Amazon Security Lake normalises data to OCSF in Parquet on S3, with no ETL; Control Tower creates both accounts. Capture data from opted-in Regions even if unused, and respect data-sovereignty rules.
SEC04-BP03 Correlate and enrich security alertsLowAlerts record who did what to which resource; map identifiers across sources so alerts can be correlated. Context sets true severity (a low-severity GuardDuty DNS alert plus hundreds of instances launched by the same identity becomes high). Sources include GuardDuty, Security Hub, Macie, Inspector, Config, IAM Access Analyzer, Network Access Analyzer and billing alarms. Amazon Detective builds relationship graphs; or build your own with Lambda, Athena and EventBridge. The AWS Threat Technique Catalog extends MITRE ATT&CK.
SEC04-BP04 Initiate remediation for non-compliant resourcesMediumDefine standard remediations in code (Lambda, Systems Manager Automation) and start them manually or automatically via Config remediation actions or Security Hub custom actions through EventBridge. Use least-privilege roles, test before automating, keep human oversight, and log actions and outcomes (CloudWatch Logs, Security Hub finding notes). Automated Security Response on AWS provides pre-built remediations.
Shell
# Organization-wide, multi-Region CloudTrail trail with log file validation
aws cloudtrail create-trail --name org-trail \
  --s3-bucket-name my-log-archive-bucket \
  --is-organization-trail --is-multi-region-trail \
  --enable-log-file-validation
aws cloudtrail start-logging --name org-trail

# VPC Flow Logs to S3 for a VPC
aws ec2 create-flow-logs --resource-type VPC --resource-ids vpc-0abc123 \
  --traffic-type ALL --log-destination-type s3 \
  --log-destination arn:aws:s3:::my-log-archive-bucket/flowlogs/

Infrastructure protection (SEC05, SEC06)

Defence in depth: trust boundaries (network and account), system hardening and patching, OS authentication, and policy-enforcement points such as WAF and API gateways. Take a Zero Trust approach where components do not trust each other. Know Regions, AZs (independent power, cooling and physical security, with encrypted inter-AZ traffic), Local Zones and Outposts.

IDBest practiceNotes
SEC05-BP01Create network layersProtecting networks: layered VPC design backed by inspection and automation. For serverless or edge-only workloads see the Serverless Applications Lens.
SEC05-BP02Control traffic flow within your network layers
SEC05-BP03Implement inspection-based protection
SEC05-BP04Automate network protection
SEC06-BP01Perform vulnerability managementProtecting compute (EC2, containers, Lambda, databases, IoT): defence in depth, vulnerability management, smaller attack surface, automated configuration, and acting at a distance.
SEC06-BP02Provision compute from hardened images
SEC06-BP03Reduce manual management and interactive access
SEC06-BP04Validate software integrity
SEC06-BP05Automate compute protection
For the network layers themselves (subnets, route tables, security groups, NACLs, endpoints) see the AWS VPC Guide.

Data protection (SEC07, SEC08, SEC09)

Classify data first, then protect it at rest and in transit. Encryption makes data unreadable without a key; tokenisation substitutes a meaningless token (not derived from the data, so a hash doesn't count) and can shrink compliance scope; masking redacts part of a value. Audit key use: KMS key usage appears in CloudTrail and can be queried with CloudWatch Logs Insights.

AreaBest practices
Data classification (SEC07)BP01 Understand your data classification scheme · BP02 Apply data protection controls based on sensitivity · BP03 Automate identification and classification · BP04 Define scalable data lifecycle management
Data at rest (SEC08)BP01 Implement secure key management · BP02 Enforce encryption at rest · BP03 Automate data-at-rest protection · BP04 Enforce access control
Data in transit (SEC09)BP01 Implement secure key and certificate management · BP02 Enforce encryption in transit · BP03 Authenticate network communications
AWS PrivateLink gives private connectivity between VPCs, on-premises networks and AWS or third-party services without an internet gateway or NAT, even across accounts with overlapping CIDRs. Traffic stays on the Amazon backbone.

Incident response (SEC10)

Even with strong preventive and detective controls, prepare to respond. Preparation (tools and access in place beforehand, rehearsed through game days) determines how well you can isolate, contain, investigate, and restore a known-good state.

AspectWhat it involves
PreparationPeople (stakeholders, training), process (documented architectures, plans, playbooks), technology (access, logging, alerting, investigation tooling). All three are required.
OperationsNIST phases: detect (find a potential event), analyse (is it an incident, and what is the scope), contain (limit the scope), eradicate (remove unauthorised artefacts and fix the cause), recover (restore a known-safe state and monitor). Phases can loop; after containment you may return to analysis.
Post-incident activityIterate on incidents and simulations to improve detection, response and time to safe operations, with clear ownership of improvements.
IDBest practice
SEC10-BP01Identify key personnel and external resources
SEC10-BP02Develop incident management plans
SEC10-BP03Prepare forensic capabilities
SEC10-BP04Develop and test security incident response playbooks
SEC10-BP05Pre-provision access
SEC10-BP06Pre-deploy tools
SEC10-BP07Run simulations
SEC10-BP08Establish a framework for learning from incidents

Cloud response design goals

  • Establish response objectives with stakeholders and legal counsel (contain, recover, preserve forensics, learn)
  • Respond in the cloud, where the event and data live
  • Know what you have and need: copy logs, snapshots and evidence into a central response account with tags and retention
  • Use redeployment: fix misconfigurations by redeploying; for compromise, verify the root cause is mitigated
  • Automate common triage and response; use humans for unique, complex or sensitive incidents
  • Choose solutions that scale with your cloud usage
  • Learn and improve; simulations are a safe way to find gaps

Application security (SEC11)

Design, build and test security properties as a normal part of the SDLC. Defects are cheaper to fix earlier, and threat modeling in design is the cheapest control of all. Four themes: organisation and culture, security of the pipeline, security in the pipeline, and dependency management.

IDBest practice
SEC11-BP01Train for application security
SEC11-BP02Automate testing throughout the development and release lifecycle
SEC11-BP03Perform regular penetration testing
SEC11-BP04Conduct code reviews
SEC11-BP05Centralise services for packages and dependencies
SEC11-BP06Deploy software programmatically
SEC11-BP07Regularly assess security properties of the pipelines
SEC11-BP08Build a program that embeds security ownership in workload teams

Operating workloads securely (SEC01)

Operating securely spans the whole lifecycle: design, build, run, and improve. The framework's answer is governance (consistent decisions that don't rely on individual judgment), automation (the same response every time), and shared components (common accounts, identity, logging, and base images) instead of each team building its own security. The best practices below follow SEC01-BP01 (separate workloads using accounts) and BP02 (secure account root user and properties).

Best practiceRisk if absentSummary
BP03 Identify and validate control objectivesHighDerive control objectives from compliance requirements and your threat model (SOC 2, PCI-DSS, ISO 27001, NIST SP 800-53, HITRUST). Use AWS Artifact to see which controls AWS covers and which are yours. Enforce with preventive controls (SCPs, resource and trust policies) and detect with AWS Config rules, Security Hub CSPM standards (FSBP, CIS AWS Foundations) and Config conformance packs. Start Config rules in report-only mode, then move to enforcement. Report effectiveness through automation.
BP04 Stay up to date with threats and recommendationsHighFollow MITRE ATT&CK, CVE, OWASP Top 10 and AWS Security Bulletins. Prefer services that update their own threat intelligence: GuardDuty, Inspector (CVE database), and WAF / Shield Advanced managed rule groups. Pair with automated fleet management and patching.
BP05 Reduce security management scopeMediumUse managed services (RDS, ECS/EKS, serverless) to shift controls to AWS and reduce patching, backup and provisioning work. Include security upkeep in TCO comparisons against self-hosting on EC2. You still own update timing and impact (RDS engine versions, Lambda runtimes); AWS Health helps track these lifecycle events.
BP06 Automate deployment of standard security controlsMediumDefine controls as IaC (CloudFormation), keep them in version control, test in CI/CD (CloudFormation Guard) and detect drift. Separate pipeline accounts and separate duties for authors and deployers. Offer approved templates as self-service via Service Catalog, Control Tower and Account Factory. Back up control configurations and scripts.
BP07 Identify threats and prioritise mitigations with a threat modelHighThreat model early and keep it as a living document; revisit on major changes, new services and landscape shifts. Use Shostack's four questions and STRIDE. Focus on custom code, where most vulnerabilities originate. Threat Composer helps produce human- and machine-readable models.
BP08 Evaluate and implement new security services and features regularlyLowSubscribe to What's New, the AWS Security Blog and the Daily Feature Updates SNS topic, plus GuardDuty and Security Hub CSPM announcements. Attend re:Inforce / re:Invent sessions and use your account team or TAM.

Threat modeling: Shostack's four questions

QuestionPractice
What are we working on?Data flow diagram, written assumptions, explicit scope
What can go wrong?Brainstorm with STRIDE: Spoofing, Tampering, Repudiation, Information disclosure, Denial of service, Elevation of privilege
What are we going to do about it?Choose mitigations you control under the shared responsibility model; cover IAM, data protection, infrastructure, logging and monitoring, and application code (input validation, session handling, bounds checks)
Did we do a good job?Review, practise, and improve model quality and speed over time

Example: preventive guardrail (SCP) and a CI check (CloudFormation Guard)

JSON
// SCP: member accounts cannot switch off GuardDuty
{
  "Version": "2012-10-17",
  "Statement": [{
    "Sid": "ProtectGuardDuty",
    "Effect": "Deny",
    "Action": ["guardduty:DeleteDetector", "guardduty:UpdateDetector",
               "guardduty:DisassociateFromMasterAccount"],
    "Resource": "*"
  }]
}
Guard
# rules.guard: fail the pipeline if a bucket has no encryption configured
rule s3_buckets_encrypted {
  Resources.*[ Type == 'AWS::S3::Bucket' ] {
    Properties.BucketEncryption exists
  }
}

# Run in CI
cfn-guard validate --data template.yaml --rules rules.guard
Avoid IAM users with long-lived access keys. Use IAM roles, IAM Identity Center, and short-lived STS credentials. Never put credentials in code, AMIs, or environment files committed to git.
Reliability Pillar 3

The ability of a workload to perform its intended function correctly and consistently, including operating and testing it through its total lifecycle, and recovering from failure.

Design principles

  • Automatically recover from failure (detect via KPIs, trigger remediation)
  • Test recovery procedures
  • Scale horizontally to increase aggregate availability
  • Stop guessing capacity
  • Manage change through automation

Best-practice areas

AreaKey practices
FoundationsService quotas, network topology, multi-AZ subnets, adequate IP space
Workload architectureService-oriented design, loose coupling, idempotent APIs, avoid hard dependencies
Change managementAuto Scaling, deploy with canary or blue/green, feature flags
Failure managementBackups, multi-AZ and multi-Region DR, chaos testing, health checks, graceful degradation

Resilience patterns

Timeouts + retries

Bounded timeouts on every call
Exponential backoff with jitter
Retry budgets to avoid retry storms
Idempotency keys for writes

Circuit breaker + bulkhead

Stop calling a failing dependency
Isolate resource pools per dependency
Fail fast, serve degraded responses

Load shedding

Reject excess work early
Prioritise critical requests
Keep latency bounded under overload

Static stability

Pre-provision capacity across AZs
Data plane works without control plane
Avoid needing to launch resources during an outage

Availability maths

AvailabilityDowntime / yearDowntime / month
99%3.65 days7.3 hours
99.9%8.76 hours43.8 minutes
99.95%4.38 hours21.9 minutes
99.99%52.6 minutes4.38 minutes
99.999%5.26 minutes26.3 seconds
Serial dependencies multiply: two components at 99.9% each give about 99.8% overall. Redundant parallel components improve availability: two independent 99.9% components in parallel give about 99.9999%, assuming truly independent failure.

Disaster recovery strategies

StrategyRTO / RPODescriptionRelative cost
Backup & restoreHours / hoursBackups copied to another Region, rebuilt from IaC on demand$
Pilot lightTens of minutes / minutesData replicated, core infrastructure provisioned but scaled to zero$$
Warm standbyMinutes / secondsScaled-down but fully functional copy running in DR Region$$$
Multi-site active/activeNear zero / near zeroFull capacity in multiple Regions serving traffic$$$$

Key services

ServiceRole
Auto Scaling, ELB/ALB/NLBSelf-healing capacity across AZs
Route 53 (health checks, failover)DNS-based failover and routing policies
RDS Multi-AZ / Aurora Global DatabaseSynchronous in-Region replication; cross-Region replication
DynamoDB global tables, S3 CRRMulti-Region data replication
AWS BackupCentralised, policy-based backups with cross-Region/account copies
SQS / SNS / EventBridgeDecoupling and buffering between components
AWS Resilience Hub, Fault Injection ServiceAssess resilience and run chaos experiments
Elastic Disaster RecoveryContinuous replication for server-based workloads

Example: Auto Scaling group spread across AZs with ELB health checks

Shell
aws autoscaling create-auto-scaling-group \
  --auto-scaling-group-name web-asg \
  --launch-template LaunchTemplateName=web,Version='$Latest' \
  --min-size 3 --max-size 12 \
  --vpc-zone-identifier "subnet-aaa,subnet-bbb,subnet-ccc" \
  --target-group-arns arn:aws:elasticloadbalancing:...:targetgroup/web/abc \
  --health-check-type ELB --health-check-grace-period 120

Reliability areas in detail

AreaTopicsSummary
FoundationsManage service quotas and constraints · Plan your network topologyRequirements that reach beyond a single workload. AWS supplies the capacity; you must track service quotas and plan IP space and connectivity before architecting the workload.
Workload architectureDesign your workload service architecture · Design interactions to prevent failures · Design interactions to mitigate or withstand failuresUpfront software and infrastructure decisions drive reliability across all pillars. Choose service boundaries deliberately, then design the distributed-system interactions to avoid failures (for example idempotency) and to survive them (timeouts, retries, graceful degradation).
Change managementMonitor workload resources · Design the workload to adapt to changes in demand · Implement changePlan for external change (demand spikes) and internal change (deployments, security patches). Monitor resources, scale automatically with demand, and deploy through automation.
Failure managementBack up data · Use fault isolation · Design to withstand component failures · Test reliability · Plan for disaster recoveryEverything eventually fails, so build resiliency in. Back up data, isolate faults (AZs, cells, bulkheads), remove single points of failure, test recovery paths (including with fault injection), and choose a DR strategy that matches RTO and RPO. The people building and running the workload must know the business reliability goals.

Best practices by topic

TopicBest practices
Manage service quotas and constraints (REL01)BP01 Be aware of service quotas and constraints · BP02 Manage service quotas across accounts and Regions · BP03 Accommodate fixed quotas and constraints through architecture · BP04 Monitor and manage quotas · BP05 Automate quota management · BP06 Keep a sufficient gap between current quotas and maximum usage to accommodate failover
Plan your network topology (REL02)BP01 Use highly available connectivity for public endpoints · BP02 Provision redundant connectivity between cloud and on-premises private networks · BP03 Ensure IP subnet allocation accounts for expansion and availability · BP04 Prefer hub-and-spoke over many-to-many mesh · BP05 Enforce non-overlapping private IP ranges in connected address spaces
Design your workload service architecture (REL03)BP01 Choose how to segment your workload (monolith, SOA, microservices) · BP02 Build services focused on specific business domains and functionality · BP03 Provide service contracts per API
Prevent failures in distributed systems (REL04)BP01 Identify the kind of distributed system you depend on · BP02 Implement loosely coupled dependencies · BP03 Do constant work · BP04 Make mutating operations idempotent
Mitigate or withstand failures in distributed systems (REL05)BP01 Implement graceful degradation (hard dependencies become soft) · BP02 Throttle requests · BP03 Control and limit retry calls · BP04 Fail fast and limit queues · BP05 Set client timeouts · BP06 Make systems stateless where possible · BP07 Implement emergency levers
Monitor workload resources (REL06)BP01 Monitor all components (generation) · BP02 Define and calculate metrics (aggregation) · BP03 Send notifications (real-time processing and alarming) · BP04 Automate responses · BP05 Analyse logs · BP06 Regularly review monitoring scope and metrics · BP07 Monitor end-to-end tracing of requests
Adapt to changes in demand (REL07)BP01 Use automation when obtaining or scaling resources · BP02 Obtain resources upon detection of impairment · BP03 Obtain resources upon detection that more are needed · BP04 Load test your workload
Implement change (REL08)BP01 Use runbooks for standard activities such as deployment · BP02 Integrate functional testing into deployment · BP03 Integrate resiliency testing into deployment · BP04 Deploy using immutable infrastructure · BP05 Deploy changes with automation
Back up data (REL09)BP01 Identify and back up all data that needs it, or reproduce it from sources · BP02 Secure and encrypt backups · BP03 Perform backups automatically · BP04 Periodically recover data to verify backup integrity and process
Fault isolation (REL10)BP01 Deploy the workload to multiple locations · BP02 Automate recovery for components constrained to a single location · BP03 Use bulkhead architectures to limit scope of impact
Withstand component failures (REL11)BP01 Monitor all components to detect failures · BP02 Fail over to healthy resources · BP03 Automate healing on all layers · BP04 Rely on the data plane, not the control plane, during recovery · BP05 Use static stability to prevent bimodal behaviour · BP06 Send notifications when events affect availability · BP07 Architect to meet availability targets and uptime SLAs
Test reliability (REL12)BP01 Use playbooks to investigate failures · BP02 Perform post-incident analysis · BP03 Test scalability and performance requirements · BP04 Test resiliency using chaos engineering · BP05 Conduct game days regularly
Plan for disaster recovery (REL13)BP01 Define recovery objectives for downtime and data loss · BP02 Use defined recovery strategies to meet them · BP03 Test the DR implementation · BP04 Manage configuration drift at the DR site or Region · BP05 Automate recovery
Monitoring: alerting should be decoupled from the systems it watches, or an outage can take your alerting down with it. Focus on tail latency (99.9th and 99.99th percentile), not just averages, and beware of "silent" failures that only customers notice. Availability and DR share practices (monitoring, multi-location deployment, failover) but availability concerns components, while DR concerns discrete copies of the whole workload and time to recovery.
Managed services already give strong baseline durability, for example EBS volumes replicate within an AZ and S3 stores objects across at least three AZs. They don't remove your responsibility for application-level resilience.
A DR plan you have not exercised is a hypothesis. Run restores and failovers regularly and measure actual RTO and RPO against targets.
Performance Efficiency Pillar 4

The ability to use computing resources efficiently to meet system requirements, and to maintain that efficiency as demand changes and technologies evolve.

Design principles

  • Democratise advanced technologies by consuming them as managed services
  • Go global in minutes (multi-Region, edge locations)
  • Use serverless architectures to remove undifferentiated operations
  • Experiment more often (easy A/B comparison of instance types and designs)
  • Consider mechanical sympathy: choose the technology that fits the access pattern

Selecting resources

DomainOptions and when to choose them
ComputeEC2 (full control, Graviton for price/performance), containers on ECS/EKS/Fargate, Lambda (event-driven, spiky load)
StorageEBS (block, low latency), EFS/FSx (shared file), S3 (object, durable), instance store (ephemeral, very fast)
DatabaseAurora/RDS (relational), DynamoDB (key-value at scale), ElastiCache/MemoryDB (in-memory), OpenSearch (search), Redshift (analytics), Neptune (graph)
NetworkingCloudFront for edge caching, Global Accelerator, placement groups, ENA/EFA, VPC endpoints, Direct Connect

Review and monitor

Measure before tuning

Load test with realistic traffic
Track p50/p95/p99 latency, not averages
Define SLOs and alert on burn rate

Right-size continuously

Compute Optimizer recommendations
Revisit instance family each generation
Benchmark Graviton for compatible workloads

Cache at every layer

CloudFront at the edge
API Gateway cache
ElastiCache / DAX in front of the database

Trade-offs

Consistency vs latency
Compression vs CPU
Batching vs responsiveness
Choose deliberately and document

Best practices by focus area

Focus areaBest practices
Architecture selection (PERF01)BP01 Learn about available cloud services and features · BP02 Use guidance from your cloud provider or a partner for patterns and best practices · BP03 Factor cost into architectural decisions · BP04 Evaluate how trade-offs affect customers and architecture efficiency · BP05 Use policies and reference architectures · BP06 Use benchmarking to drive architectural decisions · BP07 Use a data-driven approach for architectural choices
Compute and hardware (PERF02)BP01 Select the best compute options · BP02 Understand compute configuration and features · BP03 Collect compute-related metrics · BP04 Configure and right-size compute · BP05 Scale compute dynamically · BP06 Use optimized hardware-based accelerators
Data management (PERF03)BP01 Use a purpose-built data store · BP02 Evaluate configuration options for the data store · BP03 Collect and record data store performance metrics · BP04 Improve query performance · BP05 Implement data access patterns that use caching
Networking and content delivery (PERF04)BP01 Understand how networking affects performance · BP02 Evaluate networking features · BP03 Choose dedicated connectivity or VPN appropriately · BP04 Use load balancing to distribute traffic · BP05 Choose network protocols to improve performance · BP06 Choose workload location based on network requirements · BP07 Optimize network configuration based on metrics
Process and culture (PERF05)BP01 Establish KPIs for workload health and performance · BP02 Use monitoring to find where performance matters most · BP03 Define a process to improve performance · BP04 Load test your workload · BP05 Automate remediation of performance issues · BP06 Keep workload and services up to date · BP07 Review metrics at regular intervals

Process and culture principles

  • Infrastructure as code and a deployment pipeline: repeatable, low-cost, source-controlled environments
  • Well-defined metrics: technical and business KPIs, interpreted with maximum or p99 rather than averages
  • Automated performance tests: start new environments with test data, run benchmarks and load tests after quick tests pass, and tie results to the build
  • Load generation: idempotent, uncoupled scripts that mimic production user journeys (Spot Instances keep it cheap)
  • Visibility and visualisation: per-build metrics, with performance overlaid on architecture diagrams to spot hot spots and wait states
  • Regular review and continual optimisation: poor performance usually traces back to a missing or broken review process
Watch for hot partitions in DynamoDB and noisy-neighbour effects on shared resources. Averages hide tail latency, so always look at percentiles.
Cost Optimization Pillar 5

The ability to run systems that deliver business value at the lowest price point, while understanding where money is being spent.

Design principles

  • Implement cloud financial management (FinOps)
  • Adopt a consumption model: pay only for what you use
  • Measure overall efficiency (business output per unit cost)
  • Stop spending money on undifferentiated heavy lifting
  • Analyse and attribute expenditure to teams and workloads

Best-practice areas

AreaKey practices
Practice Cloud Financial ManagementOwnership, cost-aware culture, budgets and forecasts
Expenditure and usage awarenessCost allocation tags, Cost Explorer, CUR, Budgets, Anomaly Detection
Cost-effective resourcesRight-sizing, pricing models, managed services, data transfer awareness
Manage demand and supplyAuto Scaling, queues for buffering, scheduling non-prod environments off
Optimise over timeReview new services and instance generations, decommission unused resources

Pricing models

ModelBest forNotes
On-DemandSpiky, unknown, or short-lived loadNo commitment, highest unit price
Compute Savings PlansSteady baseline across EC2, Fargate, Lambda1 or 3 year commitment, flexible across families and Regions
EC2 Instance Savings Plans / RIsStable, known instance family and RegionDeeper discount, less flexibility
SpotFault-tolerant, stateless, batch, CIUp to ~90% off; can be interrupted with a 2-minute notice
S3 storage classesData with differing access patternsIntelligent-Tiering or lifecycle rules to move colder data

Example: budget with alert, and tag-based cost visibility

Shell
# Monthly cost budget, alert at 80% of actual spend
aws budgets create-budget --account-id 111122223333 \
  --budget '{"BudgetName":"monthly","BudgetLimit":{"Amount":"5000","Unit":"USD"},
             "TimeUnit":"MONTHLY","BudgetType":"COST"}' \
  --notifications-with-subscribers '[{
    "Notification":{"NotificationType":"ACTUAL","ComparisonOperator":"GREATER_THAN",
                    "Threshold":80,"ThresholdType":"PERCENTAGE"},
    "Subscribers":[{"SubscriptionType":"EMAIL","Address":"finops@example.com"}]}]'

# Cost by tag over a period
aws ce get-cost-and-usage --time-period Start=2026-10-01,End=2026-11-01 \
  --granularity MONTHLY --metrics UnblendedCost \
  --group-by Type=TAG,Key=team

Cloud Financial Management: see, save, plan, run

TenetQuestionServices
SeeHow do we measure, monitor and assign accountability for spend?Organizations, Control Tower, cost allocation tags, tag policies, Resource Groups, Cost Categories, Cost Explorer, Cost and Usage Report
SaveWhich levers (pricing models, resource choices) are we using?Savings Plans / RIs, EC2 Auto Scaling, Spot Instances
PlanHow do we forecast and budget for variable usage?Cost Explorer, Cost and Usage Report, Budgets
RunWhich guardrails and processes keep spend within budget?Billing console, IAM, SCPs, Service Catalog, Cost Anomaly Detection, Budgets

Cloud financial management practices (COST01)

IDPractice
COST01-BP01Establish ownership of cost optimization
COST01-BP02Establish a partnership between finance and technology
COST01-BP03Establish cloud budgets and forecasts
COST01-BP04Implement cost awareness in your organisational processes
COST01-BP05Report and notify on cost optimization
COST01-BP06Monitor cost proactively
COST01-BP07Keep up to date with new service releases
COST01-BP08Create a cost-aware culture
COST01-BP09Quantify business value from cost optimization

The other cost areas

AreaTopicsSummary
Expenditure and usage awarenessGovernance · Monitor cost and usage · Decommission resourcesAttribute costs to workloads, organisation units and product owners to drive efficient behaviour and reduce waste. Gather data, analyse, and report; change in usage drives change in cost.
Cost-effective resourcesEvaluate cost when selecting services · Select the correct resource type, size and number · Select the best pricing model · Plan for data transferMatch services, resource sizes and configurations to the workload, then pick the pricing model that fits its usage pattern.
Manage demand and supply (COST09)BP01 Analyse workload demand · BP02 Implement a buffer or throttle to manage demand · BP03 Supply resources dynamicallySupply resources to match demand instead of over-provisioning, or smooth demand with throttles, buffers and queues. Know the acceptable and maximum delay before shaping demand, and balance just-in-time supply against failure, availability and provisioning time.
Optimise over timeDefine a review process and analyse your workload regularly · Automating operationsReview architecture as AWS releases new services and features, decommission what you no longer need, and automate time-consuming operations.

Best practices by topic (COST02 to COST11)

TopicBest practices
Governance (COST02)BP01 Develop policies based on organisation requirements · BP02 Implement goals and targets · BP03 Implement an account structure · BP04 Implement groups and roles · BP05 Implement cost controls · BP06 Track project lifecycle
Monitor cost and usage (COST03)BP01 Configure detailed information sources · BP02 Add organisation information to cost and usage · BP03 Identify cost attribution categories · BP04 Establish organisation metrics · BP05 Configure billing and cost management tools · BP06 Allocate costs based on workload metrics
Decommission resources (COST04)BP01 Track resources over their lifetime · BP02 Implement a decommissioning process · BP03 Decommission resources · BP04 Decommission resources automatically · BP05 Enforce data retention policies
Evaluate cost when selecting services (COST05)BP01 Identify organisation requirements for cost · BP02 Analyse all components of the workload · BP03 Perform a thorough analysis of each component · BP04 Select software with cost-effective licensing · BP05 Select components to optimise cost in line with organisation priorities · BP06 Perform cost analysis for different usage over time
Resource type, size and number (COST06)BP01 Perform cost modeling · BP02 Select resource type, size and number based on data · BP03 Select them automatically based on metrics · BP04 Consider using shared resources
Best pricing model (COST07)BP01 Perform pricing model analysis · BP02 Choose Regions based on cost · BP03 Select third-party agreements with cost-efficient terms · BP04 Implement pricing models for all components of the workload · BP05 Perform pricing model analysis at the management account level
Plan for data transfer (COST08)BP01 Perform data transfer modeling · BP02 Select components to optimise data transfer cost · BP03 Implement services to reduce data transfer costs
Manage demand and supply (COST09)BP01 Analyse workload demand · BP02 Implement a buffer or throttle · BP03 Supply resources dynamically
Review process (COST10)BP01 Develop a workload review process · BP02 Review and analyse the workload regularly
Automating operations (COST11)BP01 Perform automation for operations

COST02-BP02 Implement goals and targets

Risk if not established: High. A goal is the direction you want to go; a target is how far, and by when. Make targets SMART: specific, measurable, assignable, realistic and timely.

TypeExample
Usage vs cost goalPlatform usage grows a lot with only a small, non-linear rise in cost. Target: usage +20% with cost up by less than 5%.
Efficiency goalEvery workload gets more efficient over time. Target: cost per business metric down 5% every 6 to 12 months.
Waste goalMinimise over-provisioning. Targets: waste no more than 10% of compute cost for tier-1 production workloads and 5% for tier-2.

Example KPIs

CategoryKPIDescription
ComputeEC2 usage coverageEC2 cost or hours covered by Savings Plans, RIs and Spot, versus total
ComputeCompute SP/RI utilisationUtilised SP or RI hours versus total available
ComputeEC2 cost per hour / vCPU costEC2 cost divided by instances running in that hour; cost per vCPU across all instances
ComputeLatest instance generationShare of instances on Graviton or other modern generations
DatabaseRDS coverage and utilisationRDS cost or hours covered by RIs; utilised RI hours versus available
DatabaseRDS cost per hour; latest generationRDS cost divided by instances running in that hour; share on Graviton or modern types
StorageStorage utilisationOptimised storage cost (Glacier, deep archive, Infrequent Access) divided by total storage cost
TaggingUntagged resourcesCost Explorer: exclude credits, discounts, taxes, refunds and marketplace, show only untagged resources, divide by monthly cost

Target values differ by organisation. Other common KPIs are percent of spend on-demand and adoption of gp3 EBS volumes.

Implementation steps

  1. Define expected usage: talk to application owners, marketing and the business about demand, seasonality and campaigns.
  2. Define resourcing and cost: work out what resources that usage needs and predict how cost changes as usage changes.
  3. Define business goals: combine usage and cost changes with expected technology changes. Keep goals simple and high-level, covering usage, cost and the relationship between them.
  4. Define targets: give each goal a measurable target and a deadline.
  • Start with basic KPIs and refine them as the business evolves.
  • Infrastructure metrics (latency, uptime) don't show business impact. Pair them with business metrics such as revenue or customer satisfaction.
  • Aim for near real-time KPI visibility. AWS recommends the KPI dashboard in Cloud Intelligence Dashboards (CID), built on the Cost and Usage Report (CUR), with custom goals and progress tracking. Whatever tool you use, all cloud financial management stakeholders should adopt it.

Pricing model details

  • Compute Savings Plans: up to about 66% off and the most flexible (any family, size, OS, tenancy, Region, and EC2, Fargate or Lambda). Instance Savings Plans: up to about 72%, less flexible. Payment options: no, partial or all upfront.
  • Savings Plan practice: buy small commitments continually, purchase in an account with no usage (such as the management account) so they apply to the highest discounts first, and don't chase a coverage or utilisation percentage. Watch the potential savings in Cost Explorer recommendations and buy until they fall below your required discount.
  • Reserved Instances: still apply to RDS, OpenSearch, ElastiCache, Redshift and DynamoDB. For EC2, Savings Plans give the same discount with more flexibility. Judge RIs by net savings in Cost Explorer.
  • Spot: up to about 90% off, interruptible with a 2-minute notice (on average interrupted less than 5% of the time). Be flexible across instance types and AZs, design for statelessness, and combine with On-Demand and Savings Plans. EC2 Fleet blends On-Demand and Spot automatically.
  • Third-party pricing: prefer terms that scale with the outcomes they deliver, not with your whole bill. Review pricing every two weeks to a month.
Common hidden costs: cross-AZ and cross-Region data transfer, NAT Gateway processing charges, idle load balancers, unattached EBS volumes and old snapshots, over-retained CloudWatch Logs, and non-production environments left running overnight.
Sustainability Pillar 6

The ability to continually improve sustainability impacts by reducing energy consumption and increasing efficiency across all components of a workload, maximising the benefit from the provisioned resources and minimising the total resources required.

Design principles

  • Understand your impact: measure against a business metric (e.g. emissions per transaction)
  • Establish sustainability goals and model return on investment
  • Maximise utilisation: right-size and consolidate
  • Anticipate and adopt more efficient hardware and software offerings
  • Use managed services, which share infrastructure at high utilisation
  • Reduce the downstream impact of your cloud workloads (e.g. smaller payloads, fewer device upgrades)

Practical levers

LeverExamples
Region selectionPrefer Regions with lower carbon intensity where latency and data residency allow
Alignment to demandAuto Scaling, scale to zero, schedule batch jobs for flexible times
Software and architectureEfficient code, async processing, avoid polling, reduce data movement
DataLifecycle policies, compression, delete what you no longer need, tiered storage
Hardware and servicesGraviton (better performance per watt), managed and serverless services, newer instance generations
Development processEphemeral test environments, efficient CI, device-farm usage on demand

Best practices by area

AreaBest practicesFocus
Region selection (SUS01)BP01 Choose Region based on both business requirements and sustainability goalsRegion choice affects performance, cost and carbon footprint together.
Alignment to demand (SUS02)BP01 Scale infrastructure dynamically · BP02 Align SLAs with sustainability goals · BP03 Stop creating and maintaining unused assets · BP04 Optimise geographic placement based on networking requirements · BP05 Optimise team member resources for activities performed · BP06 Implement buffering or throttling to flatten the demand curveMatch resources to what users actually consume, remove what is unused, and place resources close to users.
Software and architecture (SUS03)BP01 Optimise for asynchronous and scheduled jobs · BP02 Remove or refactor low- or no-use components · BP03 Optimise code that consumes the most time or resources · BP04 Optimise impact on devices and equipment · BP05 Use patterns that best support data access and storageKeep utilisation high and smooth load; retire idle components; reduce the need for customer device upgrades.
Data management (SUS04)BP01 Implement a data classification policy · BP02 Use technologies that support access and storage patterns · BP03 Use lifecycle policies for datasets · BP04 Use elasticity and automation to expand block storage or file systems · BP05 Remove unneeded or redundant data · BP06 Use shared file systems or storage for common data · BP07 Minimise data movement across networks · BP08 Back up data only when it is difficult to recreateProvision less storage, move cold data to more efficient tiers, and delete what you don't need.
Hardware and services (SUS05)BP01 Use the minimum hardware to meet your needs · BP02 Use instance types with the least impact · BP03 Use managed services · BP04 Optimise use of hardware-based compute acceleratorsProvision less hardware and choose the most efficient hardware and services for the workload.
Process and culture (SUS06)BP01 Communicate and cascade your sustainability goals · BP02 Adopt methods that can rapidly introduce sustainability improvements · BP03 Keep your workload up to date · BP04 Increase utilisation of build environments · BP05 Use managed device farms for testingReduce impact through development, test and deployment practices.
Sustainability and cost usually point the same direction: fewer idle resources means less spend and less energy. The AWS Customer Carbon Footprint Tool reports emissions estimates at account level.
Trade-offs Between Pillars Concept

Pillars pull against each other. The framework does not tell you which wins; it asks you to make the decision consciously, record it, and revisit it as the workload and business change.

DecisionGainsCosts
Multi-Region active/activeReliability, latencyCost, operational complexity, data consistency challenges
Extra security layers (inspection, encryption)SecurityLatency, cost, operational overhead
Aggressive Spot usageCost, sustainabilityInterruption handling, reliability engineering effort
CachingPerformance, costStaleness, invalidation complexity
Over-provisioned headroomReliability, performanceCost, sustainability
Record trade-offs as Architecture Decision Records (ADRs) and attach them to the workload in the Well-Architected Tool, so future reviewers understand why a risk was accepted.
Running a Review Process

A Well-Architected Review is a conversation, not an audit. Done with the team that owns the workload, it surfaces risks and produces an improvement plan. The AWS Well-Architected Tool (in the console, free to use) walks through the questions and tracks results.

Review flow

Process
1. Define the workload   scope, owners, environments, regions, accounts
2. Choose lenses         Well-Architected Framework + any specialist lenses
3. Answer questions      per pillar; choose the best-practice choices that apply
4. Identify risks        each answer maps to None / Medium / High Risk Issues
5. Prioritise            by business impact and effort; create improvement items
6. Remediate             track in your normal backlog; save a milestone
7. Repeat                on major changes, and at least yearly

Well-Architected Tool via CLI

Shell
# Create a workload
aws wellarchitected create-workload \
  --workload-name checkout-service \
  --description "Customer checkout API" \
  --environment PRODUCTION \
  --lenses wellarchitected \
  --review-owner team-checkout@example.com \
  --aws-regions us-east-1 \
  --pillar-priorities security reliability costOptimization \
  --architectural-design "https://wiki.example.com/checkout"

# List workloads
aws wellarchitected list-workloads

# See answers and risk counts for a workload
aws wellarchitected list-answers --workload-id <id> \
  --lens-alias wellarchitected --pillar-id security

# Save a milestone (point-in-time snapshot)
aws wellarchitected create-milestone --workload-id <id> \
  --milestone-name "2026-Q4-review"

Good practice

  • Review per workload, not per company, and include engineers who build and run it
  • Don't aim for zero risks: aim for risks that are known, owned, and accepted or scheduled
  • Review early in design and again before major launches, not only after incidents
  • Feed findings into the same backlog as feature work
Lenses Extensions

Lenses extend the core framework with guidance for a specific technology domain or industry. They apply the same six pillars with more specific questions and best practices.

LensFocus
Serverless ApplicationsLambda, API Gateway, Step Functions, event-driven design
SaaSTenant isolation, onboarding, tiering, per-tenant cost attribution
Data AnalyticsPipelines, lakes, warehouses, governance
Machine LearningTraining, deployment, MLOps, responsible AI
Generative AIFoundation models, prompts, guardrails, evaluation
Container BuildECS/EKS workload practices, image supply chain
Financial ServicesRegulated workloads, resilience and compliance
Custom lensesYour organisation's own standards, authored in JSON and shared across accounts
The set of available lenses changes over time. Check the Well-Architected Tool or the AWS lens catalog for the current list.
Cheat Sheet Reference

Operational Excellence

IaC for everything
Small, reversible changes
Telemetry: metrics, logs, traces
Runbooks and post-incident reviews

Security

Least privilege, roles not keys
Multi-account + SCPs
CloudTrail, GuardDuty, Security Hub
Encrypt in transit and at rest (KMS)

Reliability

Multi-AZ by default
Timeouts, retries with jitter, breakers
Test restores and failovers
Know your RTO / RPO and quotas

Performance Efficiency

Pick resources by access pattern
Measure p99, load test
Cache, CDN, Graviton, serverless
Re-evaluate every few quarters

Cost Optimization

Tag and attribute spend
Right-size, Savings Plans, Spot
Watch data transfer and NAT
Turn off what is not in use

Sustainability

Maximise utilisation
Managed and serverless services
Lifecycle and delete old data
Efficient hardware (Graviton)

Useful services by purpose

NeedService
Automated best-practice checksTrusted Advisor, Security Hub, Config conformance packs
Architecture risk trackingWell-Architected Tool
Resilience assessmentResilience Hub, Fault Injection Service
Right-sizingCompute Optimizer
Spend visibilityCost Explorer, Budgets, Cost Anomaly Detection, CUR
Account governanceOrganizations, Control Tower, SCPs

Found this guide useful? Buy me a coffee ☕