Arvados

h1. Federated identity

See

* #11453

* #11874

A person should be able to create an account and get a token from a single identity provider, and use that token to access private/protected resources on multiple Arvados clusters.

Motivating use cases:

* A user on cluster B shares a project with a user on cluster A.

* A container running on cluster A reads and writes data on cluster B.

* A user logged in to Workbench A can search/view/download/upload collections at cluster B.

Configuration examples:

* An organization has 5 clusters, but only one of them has user accounts and roles in its database.

* An on-premise cluster runs containers that use public data stored in the cloud (without mirroring the data locally).

h2. Design sketch

Each Arvados client must be able to prove to cluster B that it is authorized by cluster A to act on behalf of a user account which is controlled by cluster A. This must not involve giving enough information to cluster B to act on behalf of the user account: for example, the client cannot simply give cluster B its cluster A token for the purpose of doing a canary query: doing so would allow cluster B to exercise the client's authority on cluster C, D, and E as well.

h2. Protocol idea

"Salted tokens": instead of passing its literal token, the client passes the token UUID and @HMAC(token, "bbbbb")@ when sending a request to cluster B (where "bbbbb" is cluster B's cluster ID / UUID prefix). Cluster B validates the request by passing those two parameters untouched to a "verify request" ("no-op") endpoint at cluster A.

* API server hands out tokens in the form "tokenUUID <delimiter> secret" instead of just the secret part.

* Cluster B figures out cluster A's API endpoint by looking at the "site ID prefix" of the token UUID.

* Cluster B can be configured with a lookup table (clusterID→apiHost) to override the implicit {id}.arvadosapi.com

* Cluster B can be configured to _only_ use the lookup table, i.e., to never use implicit {id}.arvadosapi.com endpoints

(PA) an even simpler approach would be be to contact cluster A to get a scoped token which only allows "GET /users/current" on cluster A but is accepted by cluster B as an [all] token for that user.

h2. Adding permissions

There are a few permission-granting cases to consider.

|grantor|grantee|object|notes|

|user on site A|user on site A|object on site A|(existing permission system)|

|user on site A|group on site A|object on site A|(existing permission system)|

|user on site A|user or group on site A|object on site B|Client creates a link at site B. Site B asks site A whether the grantee user/group is visible to user A.|

|user on site A|user or group on site B|object on site B|Client creates a link at site B. Site B asks site A for a list of groups user A can see, then checks whether (possibly via one of those groups) user A can read the grantee user/group according to site B's local database.|

|user on site A|user or group on site B|object on site A|Client creates a link at site A. Site A generates a salted token and uses it to ask site B whether user A can read the grantee user/group.|

(PA) The principal is that the site that owns the object is the source of truth about who can access the object.  Permissions on site B  will dictate whether user from A has manage permission on object on B in the first place.

 (PA) what token is used to make the request on site B?  What token is used by B to contact A?

h2. TODO

Things to address

* how to sync groups

* diagrams

* mnemonic cluster names / more concrete examples (including who is reachable on the internet)

* [how] do you get a list of users/groups you can share stuff with?

* clarify what UUIDs look like (some people have A uuids, some have B uuids)

* [[Cross-cluster delegation]]