In a github issue reported to the rspec-mocks project, the user had run into a problem in a Rails’ controller spec in which an RSpec-generated test double didn’t behave as expected. What follows is an edited version of the issue and my response, with the hope that it reaches a wider audience and/or sparks some conversation.

The reported problem: ActiveSupport::JSON::Encoding::CircularReferenceError using doubles

This spec …

require 'spec_helper'

describe ListsController do
  let(:list) { double("List") }

  describe "GET 'index'" do
    let(:expected) { [{id: "1", name: "test"}] }

    before do
      list.stub(:id){ "1" }
      list.stub(:name){ "test" }
      List.stub(:select){ [ list ] }
    end

    it "should return the list of lists" do
      get :index, format: :json
      response.body.should == expected.to_json
    end
  end
end

… plus this implementation …

class ListsController < ApplicationController
  respond_to :json

  expose(:lists) { List.select("id, name") }

  def index
    respond_with(lists)
  end
end

… produces this failure:

  Failure/Error: get :index, format: :json
     ActiveSupport::JSON::Encoding::CircularReferenceError:
       object references itself

The deeper problem: this is a great example of when not to use stubs.

Here’s why: there are three incorrect assumptions hiding behind the stubs!

1. select takes an Array: List.select(["id","name"]), but the example stubs it incorrectly.
2. the id is numeric, but the example uses String.
3. the json is wrapped: {"list":{"id":1,"name":"test"}}, but the example doesn’t wrap it.

Even if the stubs were properly aligned with reality, the reason for the error is that respond_with(lists) eventually calls as_json on the list object, which, in this example, is an RSpec double that doesn’t implement as_json. We need to either use a stub_model (which does implement as_json), or explicitly stub it in the example:

list.stub(:as_json) { { list: {id: 1, name: "test"} } }

But I’d avoid stubs altogether in this case. Stubs are great for well defined (and understood) public APIs which are invoked by the code being specified. In this case, we’re stubbing an API (as_json) that is invoked by the Rails framework, not the code being specified. If the Rails framework ever changes how it renders json, this example would continue to pass, but it would be a false positive.

One possible remedy

Here’s how I’d approach this outside-in (based on my own flow, design preferences, and target outcomes. YMMV.)

Start with a request spec:

require 'spec_helper'

describe "Lists" do
  describe "GET 'index.json'" do
    it "returns the list of lists" do
      list = List.create!(name: "test")
      get "/lists.json"
      response.body.should == [{list: {id: list.id, name: "test"}}].to_json
    end
  end
end

This shows exactly what to expect, so when working on clients we can refer directly to this without having to dig into internals.

Run this and it fails with uninitialized constant List, so generate the list resource:

rails generate resource list name:string
rake db:migrate
rake db:test:prepare

Run it again and it fails with ActionView::MissingTemplate. Now we have a couple of choices. The purist view says “write a controller spec”, but some people say controller specs are unnecessary if there are already request specs (or cukes) as they just add duplication.

For me, the answer depends upon the complexity of the requirement as it compares to what we get for free from Rails. In this case, the only difference between the requirement and what Rails gives us for free is that we constrain the fields to id and name This is something we can implement in the model, so I’d just implement this very simple controller code and move on:

class ListsController < ApplicationController
  respond_to :json

  def index
    respond_with List.all
  end
end

Now the request spec fails with:

expected: "[{\"list\":{\"id\":1,\"name\":\"test\"}}]"
     got: "[{\"list\":{\"created_at\":\"2011-08-27T14:56:19Z\",\"id\":1,\"name\":\"test\",\"updated_at\":\"2011-08-27T14:56:19Z\"}}]"

We’re getting more key/value pairs than we want. I want the model responsible for constraining the keys in the json (Rails implements json transformations in the context of the model, so why shouldn’t we?), so I’d add a model spec:

require 'spec_helper'

describe List do
  describe "#as_json" do
    it "constrains keys to id and name" do
      list = List.new(:name => "things")
      list.as_json['list'].keys.should eq(%w[id name])
    end
  end
end

This fails with:

expected ["id", "name"]
     got ["created_at", "name", "updated_at"]

I expect to see created_at and updated_at, but I’m surprised (initially) to see that id is missing. Thinking this through, it makes sense because the example generates the list using new, so no id is generated. To get id to show up in the list of keys, we can use create instead of new, or we can explicitly set it. I’m going to go with setting the id explicitly to avoid the db hit, accepting the self-imposed leaky abstraction. It’s all trade-offs.

it "constrains fields to id and name" do
  list = List.new(:name => "things")
  list.id = 37
  list.as_json['list'].keys.should eq(%w[id name])
end

Now it fails with:

expected ["id", "name"]
     got ["created_at", "id", "name", "updated_at"]

Now we can implement the constraint:

class List < ActiveRecord::Base
  def as_json
    super({ only: %w[id name]})
  end
end

Now the model spec passes, but the request spec fails with:

ArgumentError:
  wrong number of arguments (1 for 0)

This is because the as_json implementation fails to honor the Rails API:

as_json(options = nil)

as_json is called by the Rails framework with an options hash. Had we done this without the request spec and weren’t aware of this information, we’d have a bunch of passing specs but the app would blow up. Hooray for testing at multiple levels!

So we add a new example to the model spec:

it "honors the submitted options hash" do
  list = List.new(:name => "things")
  list.id = 37
  list.as_json(:only => :name)['list'].keys.should eq(%w[name])
end

This fails with wrong number of arguments (1 for 0) as well, so now we adjust the model implementation:

def as_json(opts={})
  super({ only: %w[id name]}.merge(opts))
end

Now the model spec passes again, and so does the request spec! DONE!

The result is a very nice balance of clarity, speed (in spite of the one db hit in the request spec) and flexibility. Any new endpoints we add will get the same json representation because it is expressed in the model (heeding the principle of least surprise). The model spec not only specifies how the model should represent itself as json, but it helps to explain how the rails framework uses the model. All of this with no stubbing at all, and especially no stubbing of APIs our code isn’t invoking.