CephX Capabilities Need Separate Review Across Monitors, OSDs, and Managers
A Ceph keyring often looks like a single credential. Operationally, it is an identity plus a set of service-specific authorities. CephX authentication issues tickets that carry capabilities, and each Ceph service interprets those capabilities in its own API. A review that checks only whether a keyring works can miss a broad manager or monitor grant hidden behind an otherwise ordinary storage client.
Authentication is not the same as authorization
The Ceph documentation describes monitors as the central authorization authority. A principal authenticates to a monitor with a shared secret, then receives tickets that describe what it may do. The monitor, manager, OSD, and metadata server each enforce the portions of that authority that apply to them.
That split matters because one entity name can hold several capability strings. A client may have narrow access to one CephFS subtree, broad read access to a pool, and an unrelated manager capability. The first two grants may be intentional. The third may have been added to make a dashboard work and never removed.
Do not treat a successful mount or ceph status response as evidence that the keyring is least privileged. Record the entity name, its keyring location, its intended workload, and every capability plane attached to it.
Inspect each service plane
Start with the monitor capability. It controls administrative operations such as authentication state and cluster configuration. A health-checking tool usually needs less authority than an operator account, while an automated orchestrator may need more. The distinction should come from the documented commands the workload actually performs.
Next, inspect OSD capabilities. Pool-level read and write grants can be appropriate for an application, but a wildcard OSD capability should have a clear owner and an expiry or review date. Metadata server capabilities deserve a separate review because they govern CephFS paths and administrative operations, not object storage operations.
Manager capabilities are especially easy to overlook. They expose module and cluster-management functions that are not visible in ordinary object reads. A monitoring integration should not inherit allow * merely because the first connection failed.
Compare intent with commands
Build the review from evidence rather than names. Capture the service, user, automation job, or human role that uses each entity. Capture the exact command families that role needs. Then map those needs to documented capability profiles.
For example, an application that only reads objects from one pool should not use the client.admin key by default. A backup process may need read access across selected pools, but it probably does not need to create users or change cluster authentication settings. A dashboard may need read-only status data, not the ability to rewrite a module configuration.
After mapping, rotate credentials where access is excessive. The safe order is to create the replacement identity, test its narrow operations, deploy it, confirm that the old identity is no longer in use, and then remove the old keyring. Keep a rollback path during the change, but do not leave both identities valid indefinitely.
Verify effective authority
A capability review should produce evidence that can be repeated. Query the selected identity, log the output in a controlled administrative system, and compare it with the reviewed baseline. Test a denied operation that the identity should not perform. The denial matters as much as the allowed operation because it demonstrates that the boundary is real.
Check for capability changes outside the review process. Configuration management, operator scripts, and troubleshooting sessions can all add authority. Alert on differences between the approved entity definitions and the current cluster state. If policy allows emergency grants, require an owner, a reason, and a time limit.
Finally, separate CephX authorization from transport security. Ceph documentation states that CephX does not provide transport encryption such as TLS. A strong capability model does not protect credentials or object data in transit by itself. Network isolation and the deployment's supported transport protections remain separate checks.
The practical question is not whether a keyring authenticates. It is which control planes that identity can reach after authentication, who approved each grant, and how quickly an unnecessary grant can be removed.
References
- CephX protocol and capability model, Ceph documentation. Retrieved 2026-11-18.
- Ceph user management and capabilities, Ceph documentation. Retrieved 2026-11-18.
- Ceph scalability and high availability, Ceph documentation. Retrieved 2026-11-18.
Top comments (0)