Cloud Computing
AWS Well-Architected: The Six Pillars and How to Review a Real Workload
Intermediate9 min readUpdated
The Well-Architected Framework turns vague architecture opinions into a checklist of questions with defensible answers. This guide explains each of the six pillars, the concrete decisions it drives on a real system, and the trade-offs you accept when you favour one pillar over another. It closes with a practical review procedure you can run in a day.
What you will learn
- Name the six Well-Architected pillars and state the question each one answers
- Map a pillar's design principles onto concrete architectural decisions in a workload
- Recognise where pillars conflict and argue a trade-off explicitly rather than silently
- Run a Well-Architected review, record findings as risks and turn them into a remediation backlog
- Decide when a lens or a custom lens adds value over the base framework
Every architecture review eventually reaches the same impasse: two engineers disagree, both have plausible arguments, and the decision is made by whoever is more senior or more persistent. The Well-Architected Framework exists to replace that with something reproducible. It is a set of questions, grouped into six pillars, that a team answers about a specific workload, together with the design principles and improvement suggestions behind each question. The output is not a score but a list of risks the team has consciously accepted or agreed to remove.
This guide assumes you already know the services covered in AWS core services. It walks the six pillars, shows what each one changes in a real deployment, spends time on the places where the pillars pull in opposite directions — which is where the framework earns its keep — and finishes with a procedure for running your own review.
What the framework is, and what it is not
The framework is a question bank plus a facilitation method. Each pillar contains roughly ten questions of the form "How do you …?", each with a set of best-practice answers, and each best practice you cannot claim becomes a risk item rated high, medium or none. AWS publishes it as documentation, encodes it in the free Well-Architected Tool in the console, and trains partners to facilitate reviews.
What it is not: a certification, a compliance standard, or a prescription. Nothing in it says "use Amazon Aurora". It says that you should understand your recovery objectives, that you should test recovery, and that your data store should meet them. It is also not a one-off exercise. A workload passes a review, ships six months of features, and is a different workload; the framework is designed for repeated application at milestones — before launch, after a significant redesign, following a major incident.
The unit of review matters more than beginners expect. A workload is a set of components that together deliver business value: a payments API and its database and its queue, not "our AWS account" and not "one Lambda function". Reviewing an entire account produces answers so general that no action follows.
The six pillars
The framework began with five pillars; sustainability was added in 2021 and is now a peer of the others.
| Pillar | Central question | Fails as |
|---|---|---|
| Operational excellence | Can we run, observe and evolve this safely? | Long deployments, unexplained incidents, tribal knowledge |
| Security | Are identity, data and network boundaries defensible? | Standing credentials, unencrypted data, flat networks |
| Reliability | Does it survive the failures we expect, and recover from the ones we do not? | Single points of failure, untested backups, retry storms |
| Performance efficiency | Are we using the right resources, sized correctly, for the demand? | Oversized instances, wrong data store, no load testing |
| Cost optimisation | Are we spending in proportion to the value delivered? | Idle capacity, no ownership of spend, no purchase strategy |
| Sustainability | Are we minimising the energy and hardware the workload consumes? | Perpetual over-provisioning, data kept forever, chatty clients |
Operational excellence
This pillar treats operations as a product with its own backlog. Its design principles push towards small, reversible changes, operations defined as code, and learning from every failure. In practice it drives decisions such as: deployments go through a pipeline rather than a console, runbooks live in version control next to the service they describe, every service emits structured logs and business-level metrics rather than only CPU, and alarms fire on symptoms users notice rather than on internal thresholds. The review question that catches most teams is "how do you know your workload is healthy?" — answering "CloudWatch is green" is not the same as knowing that checkout succeeds.
A useful concrete test: can a new engineer deploy a one-line change to production, unaided, on their second day, and roll it back in under five minutes? If not, the operational-excellence answers are aspirational. The disciplines here overlap heavily with continuous delivery pipelines.
Security
The security pillar is organised around identity, detection, infrastructure protection, data protection and incident response. Its most load-bearing principle is to use short-lived credentials everywhere: workloads assume IAM roles rather than carrying access keys, humans federate through an identity provider rather than holding IAM users, and secrets come from a managed secret store with rotation. Encryption at rest and in transit is expected by default rather than justified case by case, because on AWS both are close to free to switch on.
The pillar also asks how you would investigate. That means CloudTrail enabled in every region and delivered to an account nobody else can write to, VPC flow logs available when needed, and a written path from "an alert fired" to "we contained it". Teams that can describe their preventative controls in detail and go quiet when asked about detection have a high-risk item whether or not they admit it.
Reliability
Reliability starts from measurable objectives. Without a stated recovery time objective and recovery point objective, "the system should be highly available" is unfalsifiable. The pillar then asks about foundations (are you near a service quota you cannot raise in an incident?), change management (does a deployment cause an outage?), and failure management (do you test recovery, or only backups?).
The design principles are the familiar ones — recover automatically, test recovery procedures, scale horizontally, stop guessing capacity, manage change through automation — and the framework is unusually direct that a backup you have never restored is not a backup. The mechanics live in high availability and disaster recovery.
Performance efficiency
This pillar is about selection and measurement: choosing the right compute, storage, database and network approach, then re-evaluating as new options appear. It is the pillar most likely to expose a decision made three years ago and never revisited — a relational database used as a queue, a fleet of general-purpose instances running a memory-bound workload, an object store queried row by row.
Its practices are democratised advanced technology (use the managed service rather than operating your own), going global in minutes when latency matters, using serverless to remove idle capacity, experimenting often, and considering mechanical sympathy — matching the technology to how the data is actually accessed. The evidence expected is benchmarks and load tests, not intuition.
Cost optimisation
Cost optimisation asks whether the workload's spend is understood, attributed and proportionate. Its practices are financial management as a shared responsibility, cost-aware architecture, matching supply to demand, measuring efficiency in business terms (cost per order, not cost per instance), and reviewing purchase commitments. A workload that cannot answer "what does one customer cost us per month?" has an immediate finding. The full discipline is covered in cloud cost management.
Sustainability
The newest pillar asks about the environmental impact of the resources you consume, and it is more actionable than it first sounds because most of its advice reduces cost as well. Its practices include placing workloads in regions with lower carbon intensity where latency and sovereignty allow, maximising utilisation so fewer physical machines are needed, adopting newer and more efficient instance generations and managed services, reducing the volume and lifetime of stored data, and reducing the work done by clients — smaller payloads, better caching, fewer polling requests.
Where sustainability genuinely diverges from cost is in things like retaining fewer copies of data across regions, or accepting slightly higher latency to run in a cleaner region. Those are real trade-offs and the pillar exists so they get discussed.
Where the pillars conflict
The framework is honest that you cannot maximise everything, and the interesting part of a review is naming the trade-offs.
- Reliability against cost. A warm standby in a second region roughly doubles infrastructure spend for a system that is idle most of the time. The question is not "can we afford it?" but "what is an hour of downtime worth?"
- Security against operational excellence. Break-glass access, strict change approval and network isolation all slow delivery. The resolution is usually automation — policy as code, automated evidence collection — rather than choosing a side.
- Performance against cost and sustainability. Provisioned capacity gives predictable latency; scaling to zero saves money and energy but introduces cold starts. Serverless workloads often accept a p99 penalty in exchange.
- Operational excellence against reliability. Shipping frequently reduces the size of each change but increases the number of change events, which are the leading cause of incidents. Progressive delivery — canaries, feature flags, automated rollback — is how you take both.
Record these as explicit decisions. A one-page architecture decision record saying "we accept a four-hour RTO because a warm standby costs more than the modelled loss" is worth more than an unwritten assumption that the system is multi-region.
Running a review on a real workload
Define the workload and its context
Write down the components, the business owner, the criticality tier, the regions used, and the current RTO and RPO. Anything outside that boundary is out of scope for this review.
Create the workload in the Well-Architected Tool
The tool stores the workload, the answers and the risk list, and lets you re-run the review later against the same baseline. Milestones freeze a point in time so you can show progress.
Answer the questions as a group, without editing reality
Work through one pillar at a time. For each best practice, the honest answers are "yes, and here is where it is implemented" or "no". "Partly" means no. Capture the evidence link — the pipeline, the dashboard, the policy — because that is what makes the answer auditable later.
Convert risks into a backlog
Each high or medium risk becomes a ticket with an owner and a target milestone. Group them: five findings often share one root cause, such as the absence of infrastructure as code.
Schedule the next review
Six to twelve months, or immediately after a major redesign or a severity-one incident.
The tool is scriptable, which matters if you review dozens of workloads. Creating a workload from the CLI looks like this:
aws wellarchitected create-workload \
--workload-name "payments-api-prod" \
--description "Card authorisation API and its data stores" \
--environment PRODUCTION \
--aws-regions eu-west-1 eu-central-1 \
--lenses wellarchitected serverless \
--review-owner "[email protected]"
aws wellarchitected list-lens-review-improvements \
--workload-id "$WORKLOAD_ID" --lens-alias wellarchitected \
--query 'ImprovementSummaries[?RiskFacingRisk!=`NONE`].[PillarId,QuestionTitle,Risk]' \
--output table
A pattern worth adopting: pair the review with a small amount of automated evidence. AWS Config rules, Security Hub standards and Trusted Advisor checks can answer a subset of the questions mechanically — is encryption enabled, are backups configured, are there public buckets — leaving the workshop time for the questions that need judgement.
Common mistakes
Reviewing the architecture you intended to build. The answers must describe production as it is today, including the manual step everyone forgets and the instance somebody launched by hand. Pull the evidence from the account, not the diagram.
Treating the risk count as a score to minimise. Teams quickly learn which answers make the number go down. If nobody ever records a high risk, the review has become theatre.
Skipping the pillars that feel like someone else's job. Developers skip cost, platform teams skip performance, everyone skips sustainability. Those are exactly the pillars where an outside question finds something.
Doing the review and never scheduling the follow-up. The framework's value is in the delta between reviews. A single review that produces a document nobody reopens is an expensive afternoon.
Key takeaways
- The Well-Architected Framework is a structured question set applied to one workload at a time, producing a list of accepted or remediated risks rather than a score.
- The six pillars are operational excellence, security, reliability, performance efficiency, cost optimisation and sustainability; each drives concrete, testable decisions.
- The framework's real contribution is forcing trade-offs into the open — reliability against cost, security against delivery speed, latency against efficiency.
- A review needs the people who build, operate and fund the workload, evidence from the live account, and an outside facilitator to keep the answers honest.
- Findings only matter if they become owned backlog items with a follow-up review scheduled; lenses extend the base questions for serverless, SaaS, analytics or your own standards.
Sources and further reading
Frequently asked questions
- Is the Well-Architected Framework only useful on AWS?
- The questions are provider-neutral even though the answers reference AWS services. Ask whether you can deploy without downtime, whether failures are isolated, whether identities are short-lived, and the discussion is identical on any platform. What you lose off AWS is the tooling: the Well-Architected Tool, the service-specific improvement plans and the partner review programme. Teams on other clouds usually copy the question set into their own checklist.
- How long does a review take and who should be in the room?
- A focused review of one workload against all six pillars takes roughly four to six hours, often split across two sessions. You need someone who knows the architecture, someone who operates it on call, someone who owns the roadmap, and a facilitator who has not built the system and can therefore ask naive questions. Reviews conducted alone by the architect who designed the system reliably produce optimistic answers.
- Do I have to fix every high risk the review finds?
- No, and pretending otherwise is why reviews get abandoned. A high-risk item means the workload would suffer material harm if the scenario occurred; the business may still accept that risk for an internal tool with ten users. What matters is that the risk is written down, owned by a named team, and revisited on a schedule, rather than discovered during the incident it predicted.
- What is a lens?
- A lens is an additional question set for a specific technology or industry — serverless, SaaS, machine learning, financial services, data analytics — layered on top of the six core pillars. Lenses ask what the base framework cannot: cold-start behaviour for serverless, tenant isolation for SaaS, model drift for machine learning. You can also define a custom lens encoding your own organisation's standards and review workloads against it.