oso helps developers build authorization into their applications. Authorization typically starts simple – perhaps a few
if statements in your code – but can grow complex as you add:
External identity data, e.g., from LDAP or OAuth
These can be hard to express concisely, and over time what started as a small number of simple
if statements can become a large amount of custom authorization logic spread throughout a codebase, which can be hard to maintain, modify, debug, and secure.
oso is designed to solve these problems based on three principles, which we’ll describe briefly here, then in more detail below.
1. Separation of concerns, but not data. Authorization logic is distinct from business logic, and by separating the two, you can make changes to your policies that apply across your application, write reusable patterns, and get one place to control, test and visualize access. But authorization decisions always rely on application context, like who a user is and what their relationship is to the data they’re trying to access. So without being part of the application, policy code should to be able to call into the application and use application objects and data.
2. Right tool for the job. Authorization deals in facts and rules about who is allowed to do what in a system. Solutions to describe authorization logic ought to be declarative and have semantics that map to common domain concepts – like roles and relationships, or boolean conditions over the input attributes.
3. Freedom to extend. No two authorization problems are the same, because no two applications are the same. And so while many authorization problems can be made to fit a general pattern like roles, the model’s fit typically degrades as you add more – and more complex – requirements. An authorization system should provide simple, opinionated building blocks to start but should not force developers to bend their requirements to the capabilities of the system. Instead, it should give them the ability to extend the system to solve the use case at hand.
Separation of Concerns, but not Data¶
Let’s imagine we’re building a SaaS app that allows organizations to manage their employees’ expenses, and that our authorization policy needs to express (at least) the following logic:
Employees can only view expenses they submitted.
A manager can view and approve their team’s expenses.
We might start by embedding this logic directly into the relevant application methods, e.g.
if user.email == expense.submitted_by: ...
But then our policy logic is intertwined with application and business logic, and diffuses throughout the application. Policy changes, audits, etc. become complicated ordeals, because there is no single place where “the policy” lives.
The maintainable solution to this problem is to factor out the authorization logic from that of the application, and make a single, uniform call when we need to authorize a request. Here’s what that might look like with oso:
Want to see how this policy works? Check out the guide for writing policies.
In this example, we’ve factored out the authorization logic into an oso policy file, and inserted calls to
oso.is_allowed in its place. All of the actual logic now resides in oso, which means that changing permissions, auditing, etc. can all happen in one place.
The key thing we did not do, however, was to separate the authorization logic from the objects it is about. Because oso operates as a library embedded within your application, it has direct access to application data, objects, and methods. For instance, in the last line of the policy above, the term
expense.submitted_by means just what you’d think: it looks up the
submitted_by attribute on the
expense object, and returns the value of that field. But the
expense object is passed directly into oso from your application; it “lives” in the application. If that attribute happened to name a method instead of a field, it would be called (with no arguments) within your application’s runtime context, and the result passed back to oso. Thus, oso can use your application’s native objects to make its authorization decisions, while at the same time keeping authorization logic separate from application logic.
Right Tool for the Job¶
If you ask someone to describe the permissions a user should have in a system using natural language, you will generally find they have no problem doing so. What often happens, however, is that authorization systems make it hard to take an intuitive concept and implement it as a concrete security policy.
oso policies are written using a declarative language designed specifically for expressing authorization logic in applications. This means that you write permissions as simple logical statements, and oso performs the necessary inferences to go from what you have (application objects and information about the request you’re trying to authorize) to a yes/no authorization decision. Rule ordering, access to application objects, and other such ancillary tasks are handled transparently by the system.
Let’s illustrate this by continuing our example from above. Suppose that we now have two different user types who can approve expenses: direct managers, and project managers. With oso, that might look like this:
# managers can approve their employees' expenses allow(user, "approve", expense) if manages(user, employee) and submitted(employee, expense); # project managers can approve project expenses allow(user, "approve", expense) if role(user, "manager", Project.lookup_by_id(expense.project_id));
# manages user or manages users' manager manages(manager, user) if employee in manager.employees() and employee = user or manages(employee, user); # user is in the list of project managers role(user, "manager", project: Project) if user in project.managers();
For full examples of the patterns used here, see the following guides:
The policy stays short and relatively flat because oso handles the evaluation. You don’t need to specify how to apply these rules. If we query oso using the above policy to see if a user can read an expense, oso will handle everything from determining which rules it needs to apply and their relative ordering, to calling into the host application to lookup the email field on the user object. You give oso all the ingredients, then oso searches through everything and puts them together in the necessary order to make a decision.
Freedom to Extend¶
Some applications may never need to go beyond basic role-based access control (RBAC). You can express that in oso easily. And likewise ABAC, and inheritance, etc. oso is purposefully agnostic to the kind of authorization logic that you need; its job is to make expressing simple policies easy, and complex policies possible.
Because the oso policy engine is an interpreter for a Turing-complete domain specific language, it is not limited to a fixed set of configuration parameters, or prescribed authorization structures. And because it offers direct integration with your application’s data and methods, it is not limited to just the data you choose to “package up” for it and ship across a wire, nor does it force you to duplicate application logic in policy code. Instead, it acts as an extension of your application that encapsulates, but does not limit, your authorization logic.
As we developed oso, we talked to a lot of organizations with a lot of different kinds of authorization requirements. Internally-facing, customer-facing, subject to stringent regulations, dependent on data that lives in a foreign system, etc. Endless variations. Most of the ones with even moderately complex requirements ended up investing heavily in custom code and frameworks, either up front, before the complexity exploded (rare) or after the fact (much more common, and much more costly).
oso helps you tame complex authorization problems by abstraction and extension. By abstracting away from, and yet fully supporting:
specific application languages and frameworks
specific authorization schemes
rigid network-based interfaces
You can adapt oso to meet even the most complex authorization requirements, because you extend the built-in system to encapsulate them, and then embed the whole engine in your application – extending your application – so that it can make decisions that are intrinsically coupled to the data and behaviors that reside there.