shoc-backend/infra/cdk/README.md
Alexandre Brandizzi 658bae77d3
Some checks are pending
Validate and deploy dev / Validate deployable source bundle (push) Waiting to run
Validate and deploy dev / Deploy shoc-backend to Elastic Beanstalk dev (push) Blocked by required conditions
fix(cdk): grant observed Elastic Beanstalk deploy reads (#38)
* fix(cdk): grant observed EB deploy reads

* fix(cdk): model EB deployment capability

* fix(cdk): scope EB platform ACL read

* fix(deploy): verify exact EB release

* docs(deploy): record unresolved EB ACL gate
2026-07-28 15:04:48 -03:00

195 lines
10 KiB
Markdown

# shoc-backend CDK (dev deployment IAM)
This CDK v2 app owns exactly one thing in the `shoc-backend` AWS account
(`396287094661`, `us-east-1`): the **GitHub OIDC deploy role** used by the
`dev` deployment workflow in `.github/workflows/deploy.yml`.
## Ownership boundary (deliberate)
CDK owns:
- The IAM role `githubdeploy-shoc-backend-dev`.
- Its OIDC trust relationship to
`arn:aws:iam::396287094661:oidc-provider/token.actions.githubusercontent.com`
scoped to `repo:Sea-Haven-Industries/shoc-backend:environment:dev`.
- Its least-privilege inline permissions policy.
CDK does **not** own, create, import, replace, or modify any of the following.
They are referenced by exact identifier only and remain owned by their original
provisioning path:
- Elastic Beanstalk application `shoc-backend`
- Elastic Beanstalk environment `shoc-backend-dev`
- DNS, VPC, EC2, RDS, and existing service/instance roles
- S3 bucket `elasticbeanstalk-us-east-1-396287094661`
- Environment configuration / option settings
- Database schema (migrations are applied by Elastic Beanstalk at deploy time,
not by CDK)
The role is retained on stack deletion (`DeletionPolicy=Retain`,
`UpdateReplacePolicy=Retain`) so an accidental teardown cannot orphan the trust
or lock out deployments.
## Least-privilege policy summary
The role grants only:
- The three read-only Elastic Beanstalk actions used by deploy, wait, and
rollback (`DescribeApplicationVersions`, `DescribeEnvironments`, and
`DescribeEvents`). These use `Resource: "*"` because Elastic Beanstalk
describe actions are not reliably constrained by resource ARN.
- `elasticbeanstalk:CreateApplicationVersion` on application `shoc-backend`.
- `elasticbeanstalk:UpdateEnvironment` on environment `shoc-backend-dev` only.
- `s3:ListBucket` and `s3:GetBucketLocation` on
`elasticbeanstalk-us-east-1-396287094661` (the official action's
ownership-safe bucket checks), plus `s3:CreateBucket` and
`s3:PutBucketOwnershipControls` on the same bucket-level ARN. Under the
`shoc-backend/` object prefix only:
`s3:PutObject`, `s3:GetObject`, `s3:GetObjectAcl`, and
`s3:GetObjectVersion`, which the pinned official deployment action and
Elastic Beanstalk require to validate the `CreateApplicationVersion` source
bundle after upload.
`s3:CreateBucket` is part of the pinned
`aws-actions/aws-elastic-beanstalk-deploy@cfad3e5e...` (v1.0.6) IAM
contract even though the workflow sets
`create-s3-bucket-if-not-exists: "false"`. That input prevents the
action's explicit bucket-creation helper; it does not remove the permission
required by the subsequent Elastic Beanstalk update path. A live deployment
confirmed this boundary: `CreateApplicationVersion` succeeded, then
`UpdateEnvironment` was denied because the caller lacked
`s3:CreateBucket` on the service bucket. The permission is scoped to that
exact bucket-level ARN only (no object prefix, no wildcard resource), so it
cannot create any other bucket.
`s3:PutBucketOwnershipControls` was added after a second live deployment
(run 30375409934) failed at `UpdateEnvironment` with `AccessDenied` for
`s3:PutBucketOwnershipControls` on the same service bucket. That call is
emitted by Elastic Beanstalk's `UpdateEnvironment` path after the source
bundle upload succeeds; AWS classifies it as a bucket-level permission, so
it is scoped to the same exact bucket-level ARN (no object prefix, no
wildcard resource). It does not widen object-prefix permissions, does not
grant `PutBucketPolicy`, `PutBucketPublicAccessBlock`, or any object-level write,
and does not change `create-s3-bucket-if-not-exists: "false"`.
`s3:GetBucketLocation` was added after CloudTrail showed that run
`30375409934` attempt 4 invoked it as
`githubdeploy-shoc-backend-dev/GitHubActions` and was denied. It is scoped to
the exact bucket-level ARN and grants no object access.
- The six CloudFormation discovery calls observed across the failed OIDC and
successful administrator deployments (`DescribeStackEvents`,
`DescribeStackResource`, `DescribeStackResources`, `DescribeStacks`,
`GetTemplate`, and `ListStackResources`) on the Elastic Beanstalk-managed
stack `awseb-e-hehnrqjjrt-stack`, scoped to
`arn:aws:cloudformation:us-east-1:396287094661:stack/awseb-e-hehnrqjjrt-stack/*`.
These read-only calls are emitted by Elastic Beanstalk's
`UpdateEnvironment` path under the GitHub deploy role. `GetTemplate` was
added after run `30375409934` attempt 2 advanced past the S3
ownership-controls step and was denied on the EB-managed stack instance
`awseb-e-hehnrqjjrt-stack/112f77c0-7718-11f1-a1a9-0e48750aef13`.
CloudTrail then showed attempt 4 denied `DescribeStackResources` and
`ListStackResources` on that same stack instance.
CloudFormation stack ARNs carry a random GUID instance suffix, so the
permission is scoped to that one stack-name prefix (`/*`) rather than a
single instance ARN. The statement grants no CloudFormation mutation, no
`Resource: "*"`, and no access to any other stack. CDK does not own or
mutate that stack; it is owned by Elastic Beanstalk and referenced by
identifier only.
- `ec2:DescribeAvailabilityZones`, `ec2:DescribeImages`, and
`ec2:DescribeSubnets` as read-only account-level discovery queries.
CloudTrail identified the GitHub deploy role as the caller during run
`30375409934`; attempt 5 confirmed the first two denials after
`DescribeSubnets` was allowed. EC2 does not support resource-level
constraints for these Describe actions, so IAM requires `Resource: "*"`.
No EC2 mutation action is granted.
- The Auto Scaling discovery calls `DescribeAutoScalingGroups` and
`DescribeScalingActivities` on `Resource: "*"` plus
`PutNotificationConfiguration`, `ResumeProcesses`, and `SuspendProcesses`
on only Auto Scaling groups whose name starts with
`awseb-e-hehnrqjjrt-stack-`. These are the exact calls recorded during the
successful administrator deployment. AWS supports resource-level
constraints for all three mutations, so replacement ASGs remain covered
without granting access to another environment.
- `s3:GetObjectAcl` under the existing
`elasticbeanstalk-us-east-1-396287094661/shoc-backend/*` object prefix.
Elastic Beanstalk emitted this read during the same attempt while validating
the uploaded application bundle. It grants no bucket-wide or cross-prefix
object access.
It grants **no** IAM mutation or `PassRole`, **no** RDS / Secrets Manager
access, no EC2 mutation, and **no** administrator policy. The only non-EB/S3
mutations are the three deployment-process Auto Scaling calls, restricted to
this environment's ASG name pattern. There are no wildcard mutation surfaces.
## Prerequisites
- Node >= 22.22.1 and npm.
- AWS credentials authorized to create/inspect CloudFormation, IAM roles, and
trust policies in account `396287094661`.
- The GitHub OIDC provider
`arn:aws:iam::396287094661:oidc-provider/token.actions.githubusercontent.com`
must already exist in the account (created once, outside this stack).
## Commands
```bash
npm ci # install pinned dependencies
npm run build # type-check / compile to dist/
npm run synth # synthesize the CloudFormation template
npm run diff # diff deployed stack vs local (requires AWS)
npm run deploy # deploy the stack (requires AWS)
```
All commands run from `infra/cdk/`.
## CI integration
`npm run synth` is the deterministic local/CI validation. After synth, inspect
`cdk.out/shoc-backend-deploy-dev.template.json` and verify the synthesized
`AWS::IAM::Role`:
- Trust policy `StringEquals` matches the exact audience and subject above.
- The inline policy contains no `Resource: "*"` mutation action and no service
outside `elasticbeanstalk` / `s3` / `cloudformation` / `ec2` /
`autoscaling`. CloudFormation discovery is limited to the single EB-managed
stack prefix. EC2 and Auto Scaling discovery use `Resource: "*"` only where
the IAM resource model requires it; Auto Scaling mutations are limited to
this environment's ASG name pattern.
The workflow's `AWS_DEPLOY_ROLE_ARN` repository secret (environment `dev`) must
hold the ARN output by this stack (`GithubDeployRoleArn`).
The OIDC deployment is currently fail-closed, not repaired. Elastic Beanstalk
still reports a generic `s3:GetObjectAcl` denial after the account-owned source
prefix and every exact cross-account object referenced by the sanitized live
stack were tested independently. The runtime-prefix, platform-assets, and
launch-control hypotheses were disproved and are intentionally absent from the
policy. Do not broaden `GetObjectAcl` without an exact principal/action/resource
record from AWS Support or the AWS-owned bucket's diagnostic owner.
The pinned deployment action can return success after Elastic Beanstalk emits a
fatal deployment event. The following workflow step therefore verifies that
the exact immutable version label is active and healthy before smoke testing.
Any mismatch fails and invokes rollback. This guard prevents false success; it
does not make the unresolved OIDC deployment path release-ready.
The GitHub `dev` environment is an external release control and must restrict
deployments to the `dev` branch. Required reviewers should be configured when
the repository plan supports environment reviewers. The workflow also checks
the exact branch before requesting an OIDC token.
## Migration and recovery contract
The deployment bundle applies pending EF Core migrations before the new
application starts. Migrations must therefore use an expand/contract sequence:
- Expand changes must remain backward compatible with the previously deployed
application version.
- Destructive contract changes are deployed only after all application versions
relying on the old schema have been retired.
- A failed deployment restores the previous **application version only**.
Database schema is not downgraded, and schema rollback is not claimed.
This contract preserves the usefulness of application-version recovery without
misrepresenting it as a tested database downgrade.