Smithy4s Conductor

Conductor is a companion library to smithy4s providing a type-safe and boilerplate-free API to add authentication and authorization logic to your smithy4s-generated services.

In particular, its API enables to:

Define authentication logic in one place as a Request[F] => F[A] function, and seamlessly reuse it across multiple services. The logic could be something like "look up for Authorization header and find out what UserId it matches in a Redis instance"
Authorize above A (which becomes Option[UserId] in our example) in a very customizable way to perform certain operation. Think of "allow a user to perform DeleteTask operation if the user is an author of the task"
Perform some post-processing of the operation output, for example if you need to erase some information in O when user doesn't have permission to view it
Invoke callbacks, for example to implement event-driven architecture or track user activity

The above use cases probably remind you of typical HTTP Middleware and Conductor is middleware in some sense, but given that it's attached on per-endpoint basis - it's completely aligned with types of your endpoint, O and E represent output and error exactly of that endpoint. And very much like with middleware - this logic is detached from logic of your service implementation.

Again, but briefly

It gives a handy framework for authenticating your users and authorizing their requests to the service generated by Smithy4s.

Here's nitpick of its primary API:

Auth((req: Request[IO]) => authUser(req, sessionsDb), User.Anonymous)
  .ofService(Tasks)
  .withHandler(_.getTask.authorizeWithInput((user, task) => user.isAllowedToGet(task.id)))
  .withHandler(_.getTasks.authorize((user) => user.isAllowedToListTasks))
  .build

Prerequisites

It's Scala 3 only
All services you want to use this library for must have generateServiceProduct trait
Your Smithy definition must have metadata smithy4sErrorsAsScala3Unions = true

Example

You can see the whole code in example directory.

Intro

In nutshell it's a service with the definition similar to the following:

object Tasks extends TasksService[IO]:
  def getTasks(id: TaskId): IO[GetTaskOutput] = ???
  def postTask(body: TaskBase): IO[PostTaskOutput] = ???
  def deleteTask(id: TaskId): IO[DeleteTaskOutput] = ???

Now, imagine this simple service needs to have the following logic:

Anyone can perform GetTasks operation, but if user is anonymous they cannot see the assignee of any of the tasks (we erase assigneeId field)
Only admins can perform PostTask
Only authors of the task can perform DeleteTask

The Challenge

The first problem arising here is that we don't have any notion of authentication - every operation could be called by admin as well as anonymous user. Smithy4s helps you to solve this problem with Server-side middleware, but despite being a very generic mechanism it has its own drawbacks:

If authentication is successful - all we know that the operation can be performed, but all information about who performs the request is lost
It's attached to the whole service and it's very hard to apply it on per-endpoint basis

Addressing the first issue typically involves leveraging IOLocal alongside middleware, a solution that, while effective, tends to require a substantial amount of boilerplate code. So, Conductor encapsulates most of that boilerplate (yes, Conductor is nothing more than IOLocal plus Middleware wrapped in a nice fluent API).

Authentication and Handlers

First you define your own Authentication domain (see Database):

enum Authentication:
  case Authenticated(user: User)
  case Anonymous

This information will be derived for every incoming HTTP Request via so called Auth logic:

val auth = Auth[IO, Authentication](authenticate(db), Authentication.Anonymous)

The next stage is to define so called "handlers" - functions that process authorize the input and transform the output.

def getTasksTransform(input: Unit, auth: Authentication, out: GetTasksOutput) =
  auth match
    case Authentication.Authenticated(_) =>
      IO.pure(out)
    case Authentication.Anonymous =>
      val erased = out.body.map(task => task.copy(origin = task.origin.copy(assigneeId = None)))
      IO.pure(GetTasksOutput(erased))

In this transformer:

First argument is an input of the endpoint. Empty in this case.
Second argument is our domain-specific Authetntication
Third argument is the endpoint output, that got evaluated and which we can modify
The output is of the same type as third argument, but potentially a) changed; b) having some side-effects (so getting back to DB is possible)

Note, we call this handler a transformer - it will get executed only after the operation succeeds. Although it's more often to use handlers that executed before the request even reaches the endpoint logic to perform authorization.

Here's an example of authorization handler:

def getTaskAuthorize(auth: Authentication) =
  auth match
    case Authentication.Anonymous =>
      AuthResult.forbid(Unauthorized("Anonymous user cannot see task details!"))
    case Authentication.Authenticated(Database.User(_, "admin")) =>
      AuthResult.allow
    case Authentication.Authenticated(_) =>
      AuthResult.forbid(Unauthorized(s"Non-admin user cannot post new tasks"))

This particular handler:

Uses only authentication info in order to make the decision (although using input is also possible)
Doesn't perform any side-effects (although they're also possible)
Type-safe! You use only AuthResult[E] to make the decision and E must match the error type of the endpoint

Here's another, more complicated authorization handler:

def deleteTaskAuthorize(db: Ref[IO, Database])(input: TaskIdInput, auth: Authentication): IO[AuthResult[TasksServiceOperation.DeleteTaskError]] =
  auth match
    case Authentication.Anonymous =>
      IO.pure(AuthResult.forbid(Unauthorized(s"Anonymous user cannot delete tasks (in fact we don't know if there's such task with id ${input.id})")))
    case Authentication.Authenticated(Database.User(userId, _)) =>
      db.get.map(_.tasks.find(_.id == input.id)).map:
        case Some(task) if task.createdBy == userId =>
          AuthResult.allow
        case Some(task) =>
          AuthResult.forbid(Unauthorized(s"Only author of the task (which is ${task.createdBy} in this case) can delete the task"))
        case None =>
          AuthResult.forbid(NotFound("No such task"))

In this case:

We perform IO to access the database - note, you're responsible for caching mechanism, Conductor doesn't attempt to reduce IO
We have both input and authentication info available. We can use the input to a) give more precise errors; b) to check something in database prior executing the operation
Despite having two independent error (Unauthorized and NotFound) - it's still type-safe (this is where smithy4sErrorsAsScala3Unions used)

This is all cool and type-safe, but what if we want to define a dynamic authorization for our endpoint. For example, with above logic an anonymous user will never be able to perform GetTask without recompiling the whole service. What if we want to store and change that information in database? In this case we can use the "default authorization" logic, which is a function of following form:

def defaultAuthorize(info: EndpointInfo, auth: Authentication): IO[AuthResult[?]] =
  info match
    case info @ EndpointInfo(ShapeId(_, service), ShapeId(_, name), _) =>
      Database.isAllowed(auth.id, info.operationId)   // pseudo-code

In this case EndpointInfo is just a ShapeId and Hints of an operation being performed. And we also know who attempts to perform that operation. And finally, we can perform an arbitrary IO with that information.

Note: you need to be super-careful with errors you're throwing here. Because unlike per-endpoint handlers, this one is not type-safe because we cannot enforce the error type when endpoint is not statically known.

Note: default authorization handler runs only when there's no specific endpoint, so it's OR relationship rather than AND.

Tying it all together

We defined:

Authentication domain and logic
A transformer
A couple of authorization handlers
The Default Authorization handler

Now it's time to attach it to our service

val mkRoutes: Resource[IO, HttpRoutes[IO]] =
  auth
    .ofService(Tasks, TasksServiceGen.serviceProduct)
    .withUnsafeDefault(defaultAuthorize)
    .withHandler(_.getTasks   .transform(getTasksTransform))
    .withHandler(_.getTask    .authorizePure(getTaskAuthorize))
    .withHandler(_.deleteTask .authorizeWithInput(deleteTaskAuthorize(db)))
    .build

We start with our own auth and attach it to Tasks service. TasksServiceGen.serviceProduct is what comes in when you add generateServiceProduct trait to your service and eventually it will become implicit, but now you need to pass it in manually in order to enable Conductor's fluent API.

withUnsafeDefault attached our default authorization hanlder (it's "unsafe" because errors aren't checked)
withHandler enables you to add different type-safe handlers to each of endpoints. This is where we attach most of previously defined handlers.
- transform for the handler that will be executed after operation
- authorizePure for the authorization handler that checks only auth info and performs no IO
- authorize for the authorization handler that checks only auth info and can perform IO
- authorizeWithInputPure for the authorization handler that checks both auth info and the input and performs no IO
- authorizeWithInput for the authorization handler that checks both auth info and the input and can perform IO
withCallback adds a side-effecting fire-and-forget function that for example can put an output to some kind of log
Finally, build transforms it all into HttpRoutes

Using auth info inside the Service

Last, but not least - all our endpoints remain without information about who executes the operation. Yes, we authenticated and authorized the request, but what if someone needs that info as part of the operation logic (handlers are pre- and post-operation on purpose)? For example, in PostTask operation we might want to save the task with createdBy property stating who creates the task.

In that case, we need to add IO[Authentication] argument to the constructor of the service:

class Tasks(getAuth: IO[Authentication]) extends TasksService[IO]:
  def postTask(task: TaskNew): IO[PostTaskOutput] =
    for
      user <- getAuth
      output <- database.add(Task(user, task))
    yield output

Now, when you attach the service to your authentication domain, you need to use not ofService factor (which accepts ready-to use service), but ofServiceConstructor factory, which accepts IO[Authentication] => Tasks constructor.

Bear in mind however, that it's better to keep all authorization logic either in handlers or in endpoints, but not to mix it. If you use getAuth IO - make sure it's for attaching info, not for authorizing requests.

Copyright and License

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this software except in compliance with the License.

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

foldables-io / conductor