Back to blog
awsaws-organizationssecurityterraformplatform-engineering

Separate production and development in AWS without overbuilding

What to separate into AWS accounts, when to use Organizations or Control Tower, and how to plan the transition with clear permissions, dependencies and tests.

Separate production and development in AWS without overbuilding

If a development deployment can modify production using the same credentials, the problem is not what the environments are called: it is where each identity's authority ends. A shared account requires that separation to be expressed through permissions and resources that coexist within the same scope.

AWS recommends separating workloads using accounts, because an account provides a strong boundary for security, access and billing. For a small team, that recommendation does not require designing today the entire organisation it might need years from now.

The reasonable decision is to separate production and non-production with a foundation the team can maintain. But creating accounts does not move applications, data or dependencies. The work is complete when you can demonstrate what is isolated, what remains shared, and how each workload is operated and recovered.

What an account separates and what remains connected

A VPC, a tag or a Terraform directory helps organise resources. They do not create another account boundary. They can be part of the design, but they do not replace the decision about who owns the resources and which identities can modify them.

Nor is putting production in another account enough if development retains a role that allows it to administer production. The same applies if a test application still writes to the production database: the infrastructure is separated, but the data flow is not.

The design must identify three connections:

  • Cross-account access: who can assume roles and which resource policies allow access from outside the account.
  • Operational dependencies: networking, DNS, artifact repositories, pipelines and shared services.
  • Data dependencies: reads, writes, copies and recovery across environments.

Sharing a component does not automatically invalidate the design. It requires a reason, an owner and a verifiable permission. The account boundary reduces certain risks; it does not eliminate the impact of a privileged identity or a shared dependency.

Start with a small structure, not one account for everything

The AWS multi-account strategy guide proposes starting with a subset and expanding it as needed. It also recommends organising around controls and operational needs, rather than copying the organisation chart into a deep hierarchy of organisational units, or OUs.

As a conceptual structure, distinguish management, production and non-production. Management serves to govern the organisation; production contains the workloads that sustain the service; non-production allows development and testing under different constraints. Logging and security need explicit locations, access and owners, even if their specific distribution requires a separate design.

This is not a universal account count or the mandatory Control Tower template. If testing and development have incompatible needs, or two products have different owners and risks, further separation may be justified. Each new boundary also adds access, configuration and maintenance work.

AWS advises keeping product workloads out of the management account and limiting who can access it. If production already runs there, record that risk and plan the transition. Do not turn a valid recommendation into an instruction to move immediately without understanding its dependencies.

Organizations and Control Tower are not equivalent alternatives

Organizations organises accounts and enables common governance. Control Tower adds orchestration of a landing zone—the configuration and controls foundation of the multi-account environment—on top of Organizations. The integration documentation explicitly describes that relationship and its preventive and detective controls.

The purchasing decision is therefore not between two interchangeable products. It is between maintaining your own foundation on Organizations and adopting Control Tower's automation and conventions.

Your own foundation may fit if requirements are limited and someone maintains its configuration, controls and tests. Control Tower is worth evaluating when creating and governing accounts repeatably is a concrete need. Before choosing, review how it will integrate existing accounts, who will handle configuration drift and what operational changes it introduces.

There is no team-size threshold or guarantee of lower cost here. The documented Control Tower structure depends on version and configuration; do not adopt an old diagram as a universal requirement. Neither option migrates workloads on its own.

DecisionWhat it provides or limitsWork it addsEvidence for acceptance
Separate production and non-productionDistinct account and access boundariesReview resources, permissions and connectionsDevelopment deploys to its own account and does not modify production without explicit authorisation
Federate human accessCommon management of identities and account accessDefine permissions, onboarding, offboarding and access recoveryEach person accesses only the intended accounts and functions
Apply limits with SCPsMaximum permissions for the identities subject to themDesign, test and maintain policies and exceptionsExpected restrictions work without blocking authorised operations
Automate the landing zoneRepeatable configuration and controlsIntegrate accounts and maintain the chosen mechanismThe foundation can be reproduced and the expected drift is detected
Transition an existing workloadBring resources and data within the intended boundaryReprovision or transfer according to the service, test and change dependenciesThe application works at the destination and recovery has a defined test

Separate human permissions, pipelines and organisational limits

For people, AWS recommends federation through IAM Identity Center. Define routine development access, deliberate production access and who administers the organisation. Also validate how to restore operations if the person who normally handles them is unavailable.

CI/CD and application identities are a separate decision. A pipeline needs its own destination and permissions: it does not automatically inherit the human access model. Check the actual account used by each deployment and cross-account authorisations. Renaming a variable does not demonstrate that the pipeline has stopped targeting production.

Service control policies, or SCPs, set maximum limits; they do not grant permissions. They require all features to be enabled in Organizations, not just consolidated billing. IAM and resource policies still have their roles.

In addition, SCPs do not restrict users and roles in the management account or service-linked roles. They do not directly affect resource policies or restrict identities from accounts outside the organisation. They are not a universal barrier around every resource.

AWS recommends testing their impact before applying them at the organisation root. Keep the pilot scoped and check both what is allowed and what is denied. An unexpected block on a necessary service is not a passed security test.

Make dependencies visible before moving workloads

If the current account is not well understood, start by reviewing an inherited AWS account. Separation needs an operational inventory, not just a list of identifiers.

For each workload, record:

  • Component or data, source account and region, intended destination and owner.
  • Identities that use it, cross-account permissions and encryption keys or dependencies.
  • Connectivity, endpoints and DNS that must change or be retained.
  • Logs, alerts, backups and recovery procedure.
  • Functional test, stop criterion and rollback conditions.

Do not copy production data to development by default. If tests need representative data, decide specifically how to obtain it, protect it and limit its use. Account separation alone does not amount to compliance with GDPR, ENS or NIS2, nor does it determine data residency in Spain. Region, backups and logs must be evaluated against the workload's requirements.

Also include Terraform remote state and the roles that read it or apply changes. Getting Terraform organised can make the transition easier, but a workspace, a provider change or state manipulation does not automatically move resources and data to another account.

A phased transition, with stop conditions

The AWS resource migration guide distinguishes options by service and explains that IaC helps reprovision an environment in another account. Not all resources are transferred in the same way. Each workload's plan must check the applicable mechanism, especially for data and encryption.

A useful sequence of work, rather than a universal procedure, would be:

  1. Inventory workloads and agree on what is in scope and what remains shared.
  2. Design accounts, access and responsibilities; define tests before making changes.
  3. Prepare the destination's governance foundation, permissions, logs and budgets.
  4. Validate a non-critical pilot, including deployment, restrictions and recovery.
  5. Reprovision or move the workload and its data using the mechanism appropriate to the service.
  6. Test the application and switch traffic or dependencies where appropriate.
  7. Remove old access and resources only after acceptance, respecting recovery needs.

Stop the transition if a critical dependency has no owner, the destination cannot read encrypted data, the required restore has not been tested, or the rollback plan does not account for writes made after the switch.

Pointing DNS back to the source does not, on its own, resolve a database that has already received writes at the destination. It must be clear which data will be retained and how operations will be restored. Depending on the workload, there may be downtime; this separation should not be commissioned under a generic zero-downtime promise.

Budget for both duplicated resources during the transition and subsequent operations: networking, backups, logs and control maintenance. Separating accounts makes attribution clearer, but does not demonstrate savings.

How to accept the work, not just the diagram

Before commissioning or carrying out the work, agree on acceptance evidence. This checklist proposes a verifiable scope; it does not describe results already achieved or a fixed commercial package.

  • A representative deployment reaches the intended account and leaves evidence of its actual identity.
  • The routine development identity receives the expected denial when attempting an agreed production action, without making a destructive change.
  • Authorised production access works and is traceable; the access recovery procedure is validated.
  • Shared flows have a documented owner, purpose and permissions.
  • Required logs can be accessed by authorised people, and alerts reach the intended owner.
  • Recovery has been tested against defined workload criteria and explicitly handles writes made after the switch.
  • Each account has an owner for operations and spending, with agreed budgets and notifications.
  • Changes remain reviewable, the team receives a runbook, and exceptions and outstanding risks are recorded.

The negative test must identify the identity, action, resource and expected denial. A network error or an expired credential does not demonstrate that the permission is properly limited. The handover is also incomplete if only the person who implemented the separation knows how to maintain it.

Questions before deciding

Is one VPC per environment enough?

It can separate part of the network, but it does not create an account boundary. If you need distinct access and governance limits, design that separation as well as the connectivity.

Do I need Control Tower?

Not for every multi-account structure. Evaluate whether its automation and controls are a better fit than a foundation of your own that the team can maintain. Organizations does not require it.

Should I move the entire current account?

Not necessarily. An existing member account can retain workloads if the design allows it. What you cannot consider complete is a separation objective while production and development remain mixed without resolution. If it is the management account, keeping product workloads there leaves a specific outstanding risk.

What changes for Terraform and CI/CD?

Target accounts, roles, permissions and state and deployment dependencies must be made explicit. Separating directories without validating the actual identity is not enough.

How much work does it involve?

It depends on the workloads, their data, encryption, cross-account access and recovery requirements. Ask for a scope based on that inventory, with a pilot, exclusions and acceptance. Without it, a fixed timeline or price for “separating AWS” hides too many decisions.

Choose a separation the team can sustain

If production and development share permissions or dependencies the team cannot explain, the next step is to clarify those boundaries, not buy a landing zone because it is fashionable.

The AWS foundations with Terraform service covers account and environment patterns, IAM, networking, reviewable changes and handover documentation. You can request the free call advertised on that page to review where the blocker is and what separation your team can maintain. The scope of an engagement is agreed after understanding the situation, without assuming a complete migration or guaranteeing savings.

AR

Alejandro Rodríguez

Freelance Cloud & DevOps Consultant, Madrid