Adventures In The Hexagonal: Part 1

One of the key challenges I’ve consistently faced in keeping my code lean and my tests mean in Rails applications has been the controller. Wether refactoring a fat controller down or a dealing with a skinny boiler plate controller, I find they suffer from cruft, issues with integration testing and worries about repetition. As I’ve discussed before it’s possible to unit test controllers, but that doesn’t eliminate the problems with architecture design; it only helps to sooth the aches and pains that arise.

Traditionally I attempted to remove business logic from my controllers by segregating that logic into a service layer. All the better for segregation of responsibilities, and for testing responsibility; even hopefully improving design along the but I had a problem. How do you handle the different resulting conditions? I found even simple CRUD could require a response beyond the boolean true/false. Have even semi-complex business logic and I became stuck, how do you handle the response when you have more than that pass/fail matrix. You could return an object with state, return symbols/strings to represent the result, but that would probably leave you with a mess of complex conditionals, yuk! I settled on throwing exceptions, using the controller to catch them and perform the relevant task. It worked well, despite violating the rule of “exceptions should be truly exceptional”, essentially the services were telling the controller what to do by the means of all hell breaking loose. However, back in the summer, I attended the Scottish Ruby Conference, and saw @MattWynne’s hexagonal rails talk; in that talk Matt proposed using the controllers themselves as response objects. Injecting themselves as a dependency to the service layer turning this:

Into something more like this:

I became a big fan of this idea, by giving the business layer something concrete to attach to, you are definitively telling the controller what to do. Additionally you end up defining an API for talking between the controller and the business layer, meaning that the business layer knows nothing of the how the response is returned, only which action it has triggered. This means you can stack different layers together, you can add in wrappers, all without changing underlying code. You can use this interface in your integration tests, or from a command line interface or anything you want.

However there are still some issues with this approach, it feels wrong to have public methods like this on Rails controllers, especially given their special treatment by Rails. At the post-talk BOF session, a few of us experimented with creating response objects to solve this problem, but didn’t come across a neat solution. However recently I came across this refactoring by @GrahamAshton with input from @MattWynne :

Suddenly I see a clearer path towards nice hexagonal controllers, these simple response objects make for simple testing, nicely segregated responsibility and can be easily reused if you so desired. I haven’t settled on a pattern for the placement of these responses, do they belong to their parents controllers? Do they get placed somewhere in the main namespace on there own? Are they worth reusing at all?

I still see problems with the controllers too, testing them (when they have large level of constants and class creation) still (to me) seems worthwhile, but is fraught with things I don’t like doing (like mocking/stubbing new to produce a fake instance). I’d rather inject the dependent services but Rails gives me no means to do so, if Rails did it would also give me the means to build single CRUD controllers with no magic/inheritance to handle multiple types of process, inject different services and responses to do different things. This is something that Raptor does and while I feel uncomfortable with no controllers I feel the router and injection is moving along the right lines.

So… Who wants to help make this possible with Rails::Metal and the Rails router ;)

As usual… Thoughts? Questions? Queries? Trolling? Find me on Twitter