On Tue, Jun 28, 2016 at 7:35 AM, Stian Thorgersen <sthorger@redhat.com> wrote:

On 27 June 2016 at 21:26, John Dennis <jdennis@redhat.com> wrote:
On 06/27/2016 07:48 AM, Marko Strukelj wrote:
> I've started work on Client Registration CLI tool. As a first step, here
> is a design document describing how I imagine the tool would be used.
> https://docs.google.com/document/d/18SoZ34sY_k7N8ae-WDsvo7QeI-cHkpTURIlUk5dpIhU/edit?usp=sharing
> I'll use this document as a spec / guide as I implement the client tool.
> Within days I'll also send a link to initial ideas for Admin Client tool
> which in principle should allow administrator to configure everything
> that can otherwise be done through Admin Console.
> Any feedback welcome.

FWIW we've already written a client registration tool for Keycloak. At
the moment it is specifically targeted for SAML clients (SP, Service
Provider) in Apache HTTPD but we have plans to extend it to OIDC.

It is currently in Fedora and will also ship in OSP.

It is hosted here:

The man page for it (formatted for HTML) can be found here:

The man page discusses 3 different ways you can authenticate and 2
different ways client registration can be performed.

I have a lot of experience with Keycloak client registration tools and
have worked through many issues, I'm happy to share my experience.

Here are some thoughts/issues you may want to take into account:

* The tool must be capable of running without interactivity as part of a
scripted installation task.

* It should not depend on a home directory being available.

* If a home directory is utilized how will you disambiguate any stored
state belonging to a script that is run by different processes but under
the same user (possibly simultaneously)? To clarify, many install tools
run as the root user or some other admin user. Each invocation of these
install tools can be run with entirely different parameters and may
execute either in parallel or partially overlapping in time.

Maybe I should have included this link in the design document to make it clear to everyone what Client Registration this tool is for: http://keycloak.github.io/docs/userguide/keycloak-server/html/client-registration.html

It's a REST API defined by specs, and is separate from Admin REST API.
About using home directory, the way I see it - you either a) specify all the state when executing a command, or b) you have a mechanism that allows the concept of 'session' between command invocations.
If you use the first approach (a) then on each invocation of the command you have to specify either username:password, or a token. The client registration specification defines workflow for Initial Access Tokens, and Registration Access Tokens, which require to automatically intercept a newly issued token after each CRUD operation, and save it for any subsequent operation on the same client resource. I can't see how this could be achieved by using the first approach.

For the second approach (b) you need a way to communicate 'session' state. The state we are saving are just tokens associated with current user or specific clients, or specific grants. Looks to me that if multiple parallel sessions are in collision about these tokens then the cli tool itself might be used the wrong way. Namely, once the client authenticates with a login, access token and refresh token are cached. Multiple client instances can use the same access token, and the same refresh token. A thing to maybe be careful about is just properly locking the file when making changes to it. For initial access token you have to explicitly add it, and assign it an alias - you can use any random value there if you want. For registration access token they are automatically associated with initial token they were initiated from - again there should be no collision.

What alternative mechanism would you suggest for storing 'session' info? We want to support Windows as well so it can't be Unix / Bash specific.


* The tool should be idempotent.

* You suggest storing tokens in a cache, how do you plan on handling the
case where a token expires before all operations are complete?

* We also initially took the approach of caching tokens but discovered
the complexity did not justify the minimal cost of obtaining a new token
for each invocation. This greatly simplified the code with very little
performance impact.

* You do not mention what type of client you're registering. I'm
assuming it's OpenID but SAML clients (SP) are equally important. The
tool must be able to handle both.

Marko is probably referring to the Keycloak client representation, which can be either OpenID or SAML. However, we also need to support OpenID Connect client descriptions as well as SAML entity descriptors as both are supported by client reg services.

The CLI needs to know which of the client registration providers (REST endpoints) to use - there are four as described in the Client Registration documentation (http://keycloak.github.io/docs/userguide/keycloak-server/html/client-registration.html)

Ideally the input format of the file could be recognised as only appropriate for one of these providers, and the correct provider then automatically used. But maybe we need a way to explicitly tell the tool what provider to use. For example:

kc new --type default --name test-app --enabled true --base-url http://localhost:8480/test-app --redirect-uri 'http://localhost:8480/test-app/*' --admin-url http://localhost:8480/test-app/logout --secret password | kc create --type default -f -

Having to set --type for both creating a description (kc new), and pushing it to the server (kc create) is not ideal.

* I don't see anything in your document on how to specify the SAML metadata.

Instead of piping in my-client.json, you would pipe in my-client-saml.xml, possibly requiring an extra --type specifier as described above.

* I don't see anything in your document on how the user modifies the
client. It appears as if you are retrieving a ClientRepresentation JSON
document and expecting the user to edit it in a text editor which will
then be sent back. That won't work for non-interactive installs. It also
presumes the user knows how to read and modify the JSON.

It would be nice to be able to set specific fields without having to modify JSON. We discussed that for the Admin CLI, but we should probably also add it to the client reg CLI

It would be very valuable to see some of the usecases, what kind of changes people do on existing clients.

* Keycloak currently has a few problems with client registration and
it's necessary to modify the client before it will work correctly. We
currently do this via the REST API. How are you planning on handling
these issues in your installer? It would be nice if the installer was
aware of the Keycloak version and could apply "fix-ups" as needed based
on the version.

AFAIK you have one problem? About not all redirect URI included from the SAML entity descriptor. Is that what you are referring to or do you have other problems?

In either case fix-ups should be performed by the client registration services, not in the CLI.

* Keycloak has two ways to register a client (client registration
service vs. REST API). The two methods do not produce the same client
configuration (I suspect because they do not share common code in the
server). How are you planning on addressing the discrepancies?

The task of the CLI is not to address any discrepancies. It's just invoking the client reg services. Any discrepancies should be handled by the client reg services themselves. Have you created JIRA's for these or can you list them to us?

* The tool should be smart enough to produce a working client without
manual intervention (i.e. the need to run admin cli commands afterwards
to fix problems). Most admins won't know how to tweak the configuration.

Can you list any you are aware of? Same comment as above applies though, it's the responsibility of the client reg services to handle this, not the CLI. Otherwise you'd have different behavior if you invoke client reg services directly rather than through the CLI.

* The tool should not have significant dependencies.

It'll be a fat-jar and will have a single dependency on the JVM.

Those are the thoughts off the top of my head, as you fill out the
details I'll continue to review. Recall the plan of record is for
Keycloak to provide such tools which we will then utilize. The
keycloak-httpd-client-install tool is a stop-gap solution until such
time as "offical" tools become available.

keycloak-dev mailing list