This page contains information related to upcoming products, features, and functionality. It is important to note that the information presented is for informational purposes only. Please do not rely on this information for purchasing or planning purposes. The development, release, and timing of any products, features, or functionality may be subject to change or delay and remain at the sole discretion of GitLab Inc.
Status Authors Coach DRIs Owning Stage Created
proposed devops verify -

Use case study: using secrets in a CI job

Objectives

  • To map out how users can use their native GitLab secrets in their CI jobs.
  • Given OpenBao is a fork of HashiCorp Vault, we want to confirm its compatibility with our Vault integration in Runner.
  • At a high level, gain a better understanding of how to structure OpenBao policies and JWT roles to be compatible with a project’s varied permissions per GitLab user role.

Prerequisites

The workflow requires that the templated policies for each combination of capabilities (e.g. read+update, read+update+create) are predefined. For example, consider the following templated policy that allows full access to a project’s secrets:

bao policy write project_full_access - <<EOF
path "kv-v2/data/projects/{{identity.entity.aliases.auth_jwt_02163755.metadata.project_id}}/*" {
  capabilities = [ "read", "create", "update", "delete", "list" ]
}
EOF

The policies are associated to JWT roles on authorization. The project_full_access policy is particularly important for the initial project owner role:

bao write auth/jwt/role/project_owner - <<EOF
{
  "role_type": "jwt",
  "policies": ["project_full_access"],
  "token_explicit_max_ttl": 60,
  "user_claim": "user_id",
  "claim_mappings": {
    "project_id": "project_id"
  },
  "bound_audiences": "secrets.gitlab.com",
  "bound_claims_type": "glob",
  "bound_claims": {
    "user_access_level": "owner"
  }
}
EOF

Given OpenBao policies are deny by default, this initial JWT role is necessary to grant project owners full access to read and write secrets.

Initial setup workflow

Details the steps and technical information for when the project’s native secrets are set up for the first time.

  1. Project owner enables GitLab Secrets manager through the GitLab UI.
  2. Project owner defines additional permissions on which GitLab user roles can read, write, or create secrets through the GitLab UI.
    • By default, project owners have full access and other roles are denied.
    • For example, if the owner allows read-only access for developer role then, through the OpenBao API, the Rails backend defines project_88_developer:

      # The format of the role name is `project_<project-id>_<user-role>`
      bao write auth/jwt/role/project_88_developer - <<EOF
       {
         "role_type": "jwt",
         "policies": ["project_read_only"],
         "token_explicit_max_ttl": 60,
         "user_claim": "user_id",
         "claim_mappings": {
           "project_id": "project_id"
         },
         "bound_audiences": "secrets.gitlab.com",
         "bound_claims_type": "glob",
         "bound_claims": {
           "user_access_level": "developer"
         }
       }
       EOF
      
    • Unlike the project_owner generic role, we have to define other non-owner roles tied to the project because projects may have different combinations of permissions per user role.
  3. Project owner defines secrets through the GitLab UI.
    • User defines details such as name, key, and value. Sample input:
      • name: Production Database Password
      • key: DB_PASS
      • value: mydbpass
    • The secret is stored in OpenBao under kv-v2/data/projects/88/ci/DB_PASS, with the JSON data:

      {
        "data": "mydbpass"
      }
      
    • The user doesn’t need to enter the secret value in JSON format. The Rails backend transforms the input into JSON object with the data key before sending it to OpenBao.
  4. Developer uses the secrets keyword in the .gitlab-ci.yml.
    • Sample configuration:

      job-with-secrets:
        secrets:
          MY_SECRET_ON_OPENBAO:
            key: DB_PASS # Translates to kv-v2/data/projects/88/DB_PASS, field `data`
      
    • There is no need to specify id_tokens:VAULT_ID_TOKEN as aud defaults to https://secrets.gitlab.com where OpenBao service is.
    • Unlike with HashiCorp Vault, there is no need to define CI/CD variables.
      • The VAULT_SERVER_URL defaults to https://secrets.gitlab.com where OpenBao service is.
      • The VAULT_AUTH_ROLE defaults to project_<project_id>_<job_user_role> to match the JWT role in OpenBao.
  5. The CI job runs and MY_SECRET_ON_OPENBAO is available as an environment variable.
    • OpenBao verifies the integrity of the ID token and validates the bound_claims if it matches the custom claims, specially the user_access_level which contains the GitLab user role of the user.
    • Similar to HashiCorp Vault secrets, this is a file variable.

Technical implementation findings

High-level technical implementation details pertaining to OpenBao and Rails to support the workflow.

  1. The OpenBao service needs to be properly configured to make it compatible with the workflow.
    • Configure JWT authentication to make it work with ID tokens authentication.
    • The documentation shows instructions using the vault CLI, but it should work similarly for bao.
    • The OpenBao API is reachable through https://secrets.gitlab.com.
    • To reference the project_id in the templated policy, it was needed to get the value of the JWT auth mount accessor (auth_jwt_02163755 from the result of bao auth list). This has to be automated during deployments so that the templated policies remain up-to-date with the correct accessor. The mount accessor value is persisted in storage and keeps its value even when the OpenBao server is restarted and sealed.
  2. The Rails backend needs the accompanying implementations to support the workflow.
    • ActiveRecord model for the secrets. Listing secrets and viewing details in the UI shouldn’t make a request to OpenBao.
    • ActiveRecord model for the permissions. Listing permissions in the UI shouldn’t make a request to OpenBao.
    • Update ID tokens related implementation to support the use of ID tokens without the need to define id_tokens in the CI configuration.
    • Proper mapping of defaults for VAULT_SERVER_URL and VAULT_AUTH_ROLE.

How to test locally

The policies and roles structure presented here was first tested locally on a GDK setup and OpenBao server running on dev mode.

Here’s a step-by-step guide on how to test this locally:

  1. Make sure GDK is properly set up with runner.
    • Tested on a GDK with Docker executor and pointed gdk.test to 172.16.123.1 but this should also work with a shell executor.
    • Confirm that you can successfully run a CI pipeline on a test project.
  2. Create the test project for fetching the secrets from OpenBao later.
    • Track its project ID. In this example, the project ID was 53.
  3. Start-up the OpenBao in dev mode.

    bao server -dev -dev-root-token-id="dev-only-token"
    
    • This makes OpenBao reachable at http://127.0.0.1:8200.
    • You might need to run export BAO_ADDR='http://127.0.0.1:8200' for the bao CLI commands below to work.
  4. Enable kv-v2 secrets engine.

    bao secrets enable kv-v2 # By default mounts to `kv-v2/data`
    
  5. Configure OpenBao JWT authentication.

    bao write auth/jwt/config \
      oidc_discovery_url="http://gdk.test:3000" \
      bound_issuer="http://gdk.test:3000"
    
  6. To test the policy and role generated for a project owner with the GitLab user role owner, create the templated policy and the JWT role for the specific owner role. The JWT role was based on the GitLab Vault sample server role.
    • Take note of the value of the JWT auth mount accessor when you run bao auth list:

      Path      Type     Accessor               Description                Version
      ----      ----     --------               -----------                -------
      jwt/      jwt      auth_jwt_02163755      n/a                        n/a
      token/    token    auth_token_90d6d0c1    token based credentials    n/a
      
    • define the templated policy and reference the project_id through the metadata of the mounted JWT auth plugin:

      bao policy write project_full_access - <<EOF
      
      # owners have full read-write access to their project's secrets
      # copy over the `auth_jwt_02163755` mount accessor value
      path "kv-v2/data/projects/{{identity.entity.aliases.auth_jwt_02163755.metadata.project_id}}/*" {
        capabilities = [ "read", "create", "update", "delete", "list" ]
      }
      EOF
      
    • define the JWT role and associate the project_full_access policy:

      bao write auth/jwt/role/project_owner - <<EOF
      {
        "role_type": "jwt",
        "policies": ["project_full_access"],
        "token_explicit_max_ttl": 60,
        "user_claim": "user_id",
        "claim_mappings": {
          "project_id": "project_id"
        },
        "bound_audiences": "secrets.gitlab.com",
        "bound_claims_type": "glob",
        "bound_claims": {
          "user_access_level": "owner"
        }
      }
      EOF
      
  7. Create a sample secret that we want to fetch in the CI job.

    bao kv put -mount=kv-v2 projects/53/foo val=my-long-passcode
    
  8. On the test project, configure the .gitlab-ci.yml to fetch secrets from OpenBao using the existing Vault integration.

    test_openbao:
      variables:
        VAULT_SERVER_URL: http://127.0.0.1:8200
        VAULT_AUTH_ROLE: project_owner
      id_tokens:
        VAULT_ID_TOKEN:
        aud: secrets.gitlab.com
      secrets:
        SECRET:
          vault: projects/53/foo/val  # translates to secret `kv-v2/data/projects/53/foo`, field `val`
          token: $VAULT_ID_TOKEN
      script:
        - echo "testing..."
        - cat $SECRET
        - echo "done."
    
    • VAULT_AUTH_ROLE matches the JWT role we created earlier.
    • aud matches the role’s bound_audiences.
    • The ID token generated in this job is matched by OpenBao using the bound_claims, specifically the user_access_level which is included in the custom claims of the ID token.
  9. Run a pipeline and confirm that in the job trace there’s a masked output of the secret that it fetched from OpenBao.