Story #11453
closedFederated user identity which works across a network of Arvados clusters
Added by Tom Morris over 7 years ago. Updated almost 7 years ago.
100%
Description
Basic elements:
- a single login server which provides authentication for all clusters in the network
- a single user UUID is used across all nodes in the cluster.
API server needs two additional features:
1. Validate salted token by contacting origin cluster
2. As an origin cluster, validate a received token from a remote cluster
Validation requests return the user record which is used to populate the local user table, along with an expiration time after which revalidation should occur.
Draft: Federated identity
Migration process from local identity to network identity is separate
Updated by Tom Morris over 7 years ago
- Target version set to Arvados Future Sprints
Updated by Tom Morris about 7 years ago
- Description updated (diff)
- Story points set to 2.0
Updated by Tom Morris about 7 years ago
- Assigned To set to Tom Clegg
- Target version changed from Arvados Future Sprints to 2017-10-25 Sprint
Updated by Peter Amstutz about 7 years ago
- Target version changed from 2017-10-25 Sprint to 2017-11-08 Sprint
Updated by Peter Amstutz about 7 years ago
- When validating a remote user,
is_admin
is forced to be false, but is_active is accepted from the remote site. I understand if the remote user is not active we shouldn't accept the token, however this also ignores the local configurationRails.configuration.new_users_are_active
when the user is initially created, which prevents the cluster admin from requiring manual activation of remote users.
- It looks like once a user account is created, a user could create a new token one the remote cluster and start using that, bypassing validation with the home cluster. Is that desirable / intentional?
test "list readable groups with salted token"
checks for the presence of "all users" and absence of "system group" but not any other group fixtures. Since these are special group ids they could conceivably have special behavior that subverts the test.
RemoteUserAccountTest
would benefit from some comments. I think what it is doing is setting up a web server in a separate thread so that the "auth" method can call "ArvadosApiToken" which calls "ApiClientAuthorization.validate" which contacts the stub web server? And the behavior of the stub web server is predetermined, so we're only testing the client side?
- In validate() you call
uuid[0..4]
in a couple of places, it might be easier to understand if it was assigned to a variable with a descriptive name like "token_uuid_prefix".
Updated by Peter Amstutz about 7 years ago
test 'remote api server is not an api server'
should return status 200.
should bail out if 'uuid' or 'secret' is empty or nil (in case of maelformed v2/ tokens)
should test maelformed v2/ tokens.
is there a test that re-validates an expired token?
what happens if a remote user has an email address or username that conflicts with a local user?
what happens if the remote user's email address or username changes?
(currently, it looks like usernames are required to be unique, email is not).
it looks like the remote auth token caching doesn't set the "secret" field, so subsequent requests won't use it but (possibly) re-validate instead.
need a test for auth token caching.
what's the rationale for putting the "validate token from remote cluster" logic in load_read_auths of application_controller instead of in the arvados_api_token middleware?
The "validate token from remote cluster" logic should ensure that the HTTP method is GET.
Updated by Peter Amstutz about 7 years ago
All new tokens given to the client should be v2/ format tokens so the client can tell if it is talking to the home cluster or a remote cluster and determine if it needs to salt the token or not.
Updated by Tom Clegg about 7 years ago
Peter Amstutz wrote:
- When validating a remote user,
is_admin
is forced to be false, but is_active is accepted from the remote site. I understand if the remote user is not active we shouldn't accept the token, however this also ignores the local configurationRails.configuration.new_users_are_active
when the user is initially created, which prevents the cluster admin from requiring manual activation of remote users.
Right. As it stands, if you accept users from remote site X, then you rely on site X's activation policy. Since the default config accepts users from *.arvadosapi.com, this seems rather permissive.
(todo?) set is_active=remote_is_active if (!remote_is_active or local_auto_activate); otherwise, leave it alone so local admin can activate
(todo) consider other activation stuff, like adding to "all users" group
- It looks like once a user account is created, a user could create a new token one the remote cluster and start using that, bypassing validation with the home cluster. Is that desirable / intentional?
This does sound dubious in that it could effectively bypass the home cluster's token expiry time.
One such procedure (login with google at remote's SSO → create session at remote's API based on matching identity_url) can be fixed (in user_sessions_controller#create) by matching existing users on identity_url only if uuid[0..4] matches Rails.configuration.uuid_prefix.
For UX (avoiding accidentally creating a local account instead of using a remote account), for now we're assuming nobody will point their web browser at a "remote" api server.
Do you have other "create a new token" procedures in mind?
test "list readable groups with salted token"
checks for the presence of "all users" and absence of "system group" but not any other group fixtures. Since these are special group ids they could conceivably have special behavior that subverts the test.
(todo) add checks for other fixtures
RemoteUserAccountTest
would benefit from some comments. I think what it is doing is setting up a web server in a separate thread so that the "auth" method can call "ArvadosApiToken" which calls "ApiClientAuthorization.validate" which contacts the stub web server? And the behavior of the stub web server is predetermined, so we're only testing the client side?
(todo) add comments here
Yes. So far we just have separate tests for the server side and the client side. An integration test will require multiple uuid-prefixes = multiple apiserver instances = multiple databases = ... some more work.
- In validate() you call
uuid[0..4]
in a couple of places, it might be easier to understand if it was assigned to a variable with a descriptive name like "token_uuid_prefix".
(todo)
Updated by Tom Clegg about 7 years ago
- Target version changed from 2017-11-08 Sprint to 2017-11-22 Sprint
Updated by Tom Clegg about 7 years ago
- Target version changed from 2017-11-22 Sprint to 2017-12-06 Sprint
Updated by Peter Amstutz about 7 years ago
Related to #12627 it might be a good idea to move more of the token handling into the middleware instead of the application controller (at least from an code auditing standpoint it makes it easier to understand the request handler process).
Updated by Peter Amstutz about 7 years ago
- Blocks Story #12705: Documentation/helper scripts for migrating users to federated identity added
Updated by Tom Clegg about 7 years ago
11453-federated-tokens @ f3dc89653597f7f6de480850231ea1f6b991c8aa
changes:- activation
- if remote user is active, and local mirror entry already exists and is active, user is active locally
- if remote user is active, and local config is auto-activate, then remote user is activated immediately
- if remote user is inactive, remote user is deactivated if necessary
- identity_url is not copied from remote (so remote user can't bypass remote auth -- or get a "trusted client" token -- by logging in via local SSO)
- email and username are not copied from remote (so they don't conflict with local)
- salted token is cached locally so it can be validated on subsequent calls without repeating the remote verification
- "verify salted token" checks http method == GET
- token checks moved to middleware
- test remote-auth cache + expiry + un-expiry, malformed tokens, ignore extra token fields
- other cleanup/comments noted above
All new tokens given to the client should be v2/ format tokens so the client can tell if it is talking to the home cluster or a remote cluster and determine if it needs to salt the token or not.
I'm not sure I follow this. I see how this could be a convenience for a client (just store a list of tokens, instead of a mapping of cluster→token). Is that what you mean, or are you saying it's necessary?
Updated by Tom Clegg about 7 years ago
- Target version changed from 2017-12-06 Sprint to 2017-12-20 Sprint
Updated by Tom Clegg about 7 years ago
- fix tls cert to un-break python tests.
- passed https://ci.curoverse.com/job/developer-run-tests/534/
Updated by Peter Amstutz about 7 years ago
Updates look good. I did some manual testing between two arvbox instances.
Since they use self-signed certificates, I added this (you should consider applying it):
--- a/services/api/app/models/api_client_authorization.rb
+++ b/services/api/app/models/api_client_authorization.rb
@@ -133,6 +133,9 @@ class ApiClientAuthorization < ArvadosModel
# [re]validate it.
begin
clnt = HTTPClient.new
+ if Rails.configuration.sso_insecure
+ clnt.ssl_config.verify_mode = OpenSSL::SSL::VERIFY_NONE
+ end
remote_user = SafeJSON.load(
clnt.get_content('https://' + host + '/arvados/v1/users/current',
{'remote' => Rails.configuration.uuid_prefix},
I created a token in Python:
>>> import hmac
>>> import hashlib
>>> h = hmac.new('4yxuv7ra2ge7pndh3a075nwa2nd8endbm1kf7v73dyt0yiws2v', '1bq65', hashlib.sha1)
>>> "v2/1lzl6-gj3su-evhdy1tn20jjb0d/"+h.hexdigest()
'v2/1lzl6-gj3su-evhdy1tn20jjb0d/3586b7802b2a37abafd056a019ba5307636a31b9'
This worked! I was able to use "arv user current".
I was also able to log into workbench by adding "?api_token=..."
Notes:
- Your note has a mistake, in your branch you are copying the email address. However, I don't think Arvados cares if there are two accounts with the same email addresses. We can probably update email address when refreshing the user record.
- We are not assigning a username at all. This repositories. I suggest using the upstream username if available, or creating a unique username if not. We probably don't want to ever change the username once it has been assigned.
- Should we also synchronize the user's public SSH keys from upstream?
All new tokens given to the client should be v2/ format tokens so the client can tell if it is talking to the home cluster or a remote cluster and determine if it needs to salt the token or not.
I'm not sure I follow this. I see how this could be a convenience for a client (just store a list of tokens, instead of a mapping of cluster→token). Is that what you mean, or are you saying it's necessary?
I can't remember now if I had a specific case in mind where it was strictly necessary. There's some benefit to making the tokens more self-describing, but I don't think it blocks anything.
Updated by Peter Amstutz about 7 years ago
Note on #12758 (Ensure federated identity + group sync tool work well together)
11453-federated-tokens does not include synchronizing federated group memberships, although it has some minimal support for queries necessary to support it in the future.
This means the group sync tool will still need to be run on every cluster, but will update group membership for federated user accounts instead of local ones.
Updated by Tom Clegg about 7 years ago
- sso_insecure applies to remote auth too
- unique username is assigned using the usual rules (based on remote username if available, otherwise based on email)
- included "email is preserved" in test case
Copying SSH keys sounds like a good idea. After merging this, right?
Updated by Peter Amstutz about 7 years ago
Tom Clegg wrote:
11453-federated-tokens @ 1b993cdda270016bcf82fcad7f2168659345aa0e
- sso_insecure applies to remote auth too
Thanks.
- unique username is assigned using the usual rules (based on remote username if available, otherwise based on email)
Looks like it doesn't check if we already assigned a username? Which might lead to it assigning a new username on every token refresh!
Copying SSH keys sounds like a good idea. After merging this, right?
Yes, was just a thought, not a blocker.
Updated by Tom Clegg about 7 years ago
Peter Amstutz wrote:
Looks like it doesn't check if we already assigned a username? Which might lead to it assigning a new username on every token refresh!
The find_or_create_by() block only runs in the "create" case. I'll add a test though.
Updated by Tom Clegg about 7 years ago
11453-federated-tokens @ afd47ec89aae6fa96dfedc53420b998b50d48318
Updated by Tom Clegg about 7 years ago
11453-federated-tokens @ 934b2b25199cb4696fe6ad38406b3ec1cec8f9e8
Updated by Peter Amstutz about 7 years ago
934b2b25199cb4696fe6ad38406b3ec1cec8f9e8 LGTM, thanks.
Updated by Tom Clegg almost 7 years ago
- Target version changed from 2017-12-20 Sprint to 2018-01-17 Sprint
Updated by Tom Morris almost 7 years ago
- Status changed from In Progress to Closed