brandur.org

Scoping and OAuth 2

Yesterday marked the beta release of OAuth for the Heroku Platform API, which we hope will empower users to develop apps against the API by providing a simple and powerful authentication framework that’s consistent with other providers across the web.

One interesting discussion that developed while we were building this out was around OAuth scoping, the mechanism that allows OAuth clients to tell an authorization server what permissions they’ll need on resources they’re accessing. I thought this might be a good opportunity to talk a little about OAuth scoping, what the spec has to say about it, how it’s implemented elsewhere on the web, and our own design considerations.

RFC 6749 describes how scope should be implemented according to the proposed OAuth 2 standard. I’ve tried to summarize the main points presented in the document:

  • Scope is specified on either the authorization or token endpoints using the parameter scope.
  • Scope is expressed as a set of case-sensitive and space-delimited strings.
  • The authorization server may override the scope request, in this case it must include scope in its response to inform a client of their actual scope.
  • When a scope is not specified, the server may either fallback to a well-documented default, or fail the request.

The spec describes the format that a scope should have and how the server should handle it, but is open-ended in respect to what strings in a scope should actually look like. This decision allows providers to define their own strings, and gives them enough flexibility to ensure that OAuth 2 scoping is a good fit for accessing a wide variety of different resources.

The open-ended spec has resulted in all kinds of creative implementations across the web, with no two being exactly alike. I’ve compiled a few examples to demonstrate the range of ideas out there.

App.net allows developers to define a basic set of scopes in snake_case. This pretty standard scoping implementation is simple and effective.

basic stream update_profile

http://developers.app.net/docs/authentication/#scopes

Facebook deviates from spec a bit by suggesting that scope strings be comma-delimited. The two other interesting characteristics of Facebook scopes are that more specific strings are namespaced under their broader category (e.g. user_actions.video), and that some strings are dynamic (e.g. APP_NAMESPACE scopes to a particular app in user_actions:APP_NAMESPACE). Facebook also offers a very extensive variety of available scopes so that apps can be very precise about what powers they’ll require.

email,read_stream,user_actions.video,user_actions:APP_NAMESPACE

https://developers.facebook.com/docs/reference/login/#permissions

GitHub provides a concise set of scopes with some namespacing using the colon character. For example, user:email is a subset of the permissions allowed by user.

gist repo user user:email

Another interesting innovation here is that for any API requests, GitHub passes back the response headers X-OAuth-Scopes and X-Accepted-OAuth-Scopes to indicate to the user what scope strings their token has, and what strings this endpoint will accept. This makes their APIs self-documenting in that it provides users an easy alternative to looking up documentation when choosing scope for their apps and tokens.

http://developer.github.com/v3/oauth/#scopes

Google mandates that scopes should start with the openid string, then include either or both of email and profile. From there, scope is extended across Google’s flourishing ecosystem by defining other strings as extensible URIs.

openid profile email https://www.googleapis.com/auth/drive.file

https://developers.google.com/accounts/docs/OAuth2Login

Another fairly simple implementation, with the notable use of plus signs rather than spaces for delimitation.

likes+comments

http://instagram.com/developer/authentication

LinkedIn reserves the underscore to separate types of resources from the read/write permissions to that type, with an r specifying read privileges and w write.

r_basicprofile r_emailaddress rw_groups w_messages

https://developer.linkedin.com/documents/authentication#granting

An uncommon trait here is that Salesforce requires a particular scope string for the privilege of being granted a refresh token.

api refresh_token web

http://help.salesforce.com/help/doc/en/remoteaccess_oauth_scopes.htm

Shopify also mixes read and write permissions into scope strings. Their system is fairly intuitive in that write_ also implies read_ permission, so that developers don’t need to specify both.

read_customers write_script_tags, write_shipping

http://docs.shopify.com/api/tutorials/oauth

Defines scope strings that are prefixed with wl. (for Windows Live); presumably so that scopes are unique across Microsoft’s entire product space.

wl.basic wl.offline_access wl.contacts_photos

http://msdn.microsoft.com/en-us/library/live/hh243646.aspx

The end product for Heroku OAuth scope was shaped by a few major product and engineering design goals:

  • The set of scope strings should be minimal so that we still have the power to evolve scoping as we continue to build out our product. Even if we completely redesign our scope strings, all the old strings should be general enough to easily map to the new system. Adding things is easy, but deprecating is hard.
  • Some app resources are sensitive enough that even if a scope grants almost universal permission to manage an app, they still need to be protected. A good example of this are an app’s config vars, which contain secrets like database connection strings. The scoping system must take this into account.
  • We should provide a very minimal scope that provides basic user information and nothing else. This is useful in systems that will use OAuth to identify a user and little else like the Heroku Forums.

Taking these goals into account, along with the spec and the web’s other implementations, we came up with a starting point for our scope system which is what was released yesterday:

  • identity: Allows access to GET /account for basic user info, but nothing else.
  • read: Read access to all a user’s apps and their subresources, except for protected subresources like config vars and releases.
  • write: Write access to apps and unprotected subresources. Superset of read.
  • read-protected: Read including protected subresources. Superset of read.
  • write-protected: Write including protected subresources. Superset of read-protected and write.
  • global: Global access encompassing all other scope.

These strings map to a much more granular set of permissions in the backend, which will allow us to continue evolving the public interface as need be.

Like a few other providers, we also elected for self-documenting API endpoints that help developers along by specifying their accepted scope strings as response headers (we tend to drop any X- prefixes as they’re effectively deprecated):

Oauth-Scope: global
Oauth-Scope-Accepted: global read read-protected write write-protected

This isn’t a finalized design, and are looking forward to collecting requirements from the community and internal consumers and iterating on it, the hope being that we can provider a system which offers a powerful amount of flexibility and granularity, while staying simple to use and true to the original spec.

Did I make a mistake? Please consider sending a pull request.