Feature #14200
open
[API] Reduce privilege exposure via API tokens in multi-cluster workflows
Added by Peter Amstutz over 5 years ago.
Updated 2 months ago.
Release relationship:
Auto
Description
A container running on cluster A might have inputs located on cluster B. Therefore, it must have a runtime token capable of authorizing API calls to cluster B. However, the container does not need all of the privileges on cluster B that it needs on cluster A: for example, it does not need to create a log collection on cluster B.
Proposal:
- Additional "cluster_scope" column restricting which clusters should accept it? If cluster B tries do use with cluster C, cluster A will tell cluster C not to use it.
- "cluster_ scope" could also instruct remote clusters to limit their scope (so token used on cluster C still only has access to read-only collections).
- Proposed format: {cluster1: [scope1, scope2], cluster2: [scope3, scope4]}
- Description updated (diff)
It would be good to have a way to expire the token when the container ends.
Couldn't the existing "salt" mechanism be used when cluster B wants to give cluster C a cluster-C-only token?
As with the single-cluster case, using scopes to limit access to input collections sounds good, but how do we express "permitted to create and update log/output/temp collections"?
Tom Clegg wrote:
It would be good to have a way to expire the token when the container ends.
Thoughts:
- Cluster A can set a default token expiration date
- Cluster A can poll cluster B to see when the container request is finalized (not sure what component would do this)
- Cluster B can push cluster A to expire the token when the container request is finalized
The 2nd and 3rd bullet points assumes non-malicious cooperation on the part of cluster B (but that's probably fine, since we're already trusting it to do compute for us).
Couldn't the existing "salt" mechanism be used when cluster B wants to give cluster C a cluster-C-only token?
Yes, this would prevent cluster C from using the token to access other clusters, which seems fine and desirable. However, to prevent cluster B from using the token it on clusters other than those intended, I think we still want per-cluster scope.
As with the single-cluster case, using scopes to limit access to input collections sounds good, but how do we express "permitted to create and update log/output/temp collections"?
The assumption here is that the token scope limits what can be done on cluster A and C, but not cluster B.
- Subject changed from [API] can delegate permissions to remote container requests to [API] Reduce privilege exposure via API tokens in multi-cluster workflows
- Description updated (diff)
- Related to Feature #14262: [Controller] Specify runtime_token when creating container requests on a remote cluster added
- Target version deleted (
To Be Groomed)
- Target version set to Future
Also available in: Atom
PDF
Updated by 
Updated by