Skip to content

Authorization

Overview

Authentication verifies who you are, while authorization determines what you can do.

Authorization can only be enabled if Authentication is enabled. Please check the Authentication Docs for more information.

Lakekeeper currently supports the following Authorizers:

  • AllowAll: A simple authorizer that allows all requests. This is mainly intended for development and testing purposes.
  • OpenFGA: A fine-grained authorization system based on the CNCF project OpenFGA. OpenFGA requires an additional OpenFGA service to be deployed (this is included in our self-contained examples and our helm charts). See the Authorization with OpenFGA guide for details.
  • Cedar: An enterprise-grade policy-based authorization system based on Cedar. The Cedar authorizer is built into Lakekeeper and requires no additional external services. See the Authorization with Cedar guide for details.
  • Custom: Lakekeeper supports custom authorizers via the Authorizer trait.

Check the Authorization Configuration for setup details.

Instance Admins

Available since Lakekeeper 0.12.1.

"Instance admins" are principals granted privileged access via deployment configuration rather than through the selected Authorizer. They exist so that deployment-time operators — typically a Kubernetes Operator reconciling Lakekeeper resources — can always administer the catalog, even if users misconfigure their authz backend (for example by deleting the last OpenFGA admin tuple or deploying a restrictive Cedar policy that denies everything).

Scope

Instance admins bypass authorization for control-plane operations: - Bootstrap. - Project, role, warehouse, namespace management. - Table / view metadata operations, including GetMetadata, Commit, Drop, Rename, property changes. - User management.

Instance admins do not bypass authorization for: - Data-plane operations — CatalogTableAction::ReadData and CatalogTableAction::WriteData still route through the configured Authorizer. If the instance admin does not hold the relevant grants, reads and writes of table row data are denied. - Role assumption (x-assume-role header) — an instance admin must act with their own identity. Assuming a role opts into that role's narrower scope. - Permission-management endpoints exposed by the active Authorizer (for example /management/v1/permissions/... under OpenFGA; Cedar exposes its own set) — the instance-admin bypass does not apply to these. Writes go through the Authorizer's own grant-check path, so an instance admin cannot directly make Alice a project_admin. Ongoing permission administration stays with a principal that holds real grants in the configured Authorizer.

This split keeps a leaked operator credential from being trivially used either to exfiltrate data or to escalate arbitrary principals to admin.

Configuration

Set LAKEKEEPER__INSTANCE_ADMINS to a TOML inline array of user IDs. For simple string arrays this is syntactically identical to a JSON array:

# e.g. in a Kubernetes deployment's env block
env:
  - name: LAKEKEEPER__INSTANCE_ADMINS
    value: '["kubernetes~system:serviceaccount:lakekeeper:operator","oidc~alice"]'

Each entry is a Lakekeeper user ID of the form <idp_id>~<subject>. The idp_id matches the identifier of a configured Authenticator (for example, kubernetes or oidc). The subject is the resolved subject claim — for Kubernetes ServiceAccount tokens that is system:serviceaccount:<namespace>:<sa-name>; for OIDC it is whatever the configured subject claim produces.

A bare string (e.g. oidc~alice) is rejected — even a single admin must be wrapped in brackets: ["oidc~alice"]. The indexed-variable pattern that some other config systems accept (LAKEKEEPER__INSTANCE_ADMINS__0=...) is not supported.

Operational notes

  • Not a recovery mechanism. If OpenFGA is unreachable or the authn layer is misconfigured such that the instance admin's identity cannot be resolved, the bypass does not engage. Instance admins are for day-to-day operator access, not break-glass recovery.
  • Rotation. The admin list is read once at process startup. Adding or removing an admin requires a redeploy. This is intentional: the mechanism is a deployment-config concern, not a runtime one.
  • Audit. Authorization events include a privilege_source field indicating how the decision was reached: "internal" (in-process call), "instance_admin" (config-granted bypass), or "authorizer" (configured Authorizer backend decision). See the Logging guide for the event schema.
  • Role-assumed requests. Setting x-assume-role on a request from an instance admin drops the bypass for that request — the effective scope is whatever the assumed role holds.
  • Permission administration. Because instance admins cannot write to the OpenFGA permission-management endpoints, day-to-day management of role grants and assignments is done by a human (or service) principal that was bootstrapped through OpenFGA. The operator use case is provisioning (creating projects/warehouses, initial bootstrap), not ongoing user administration.