Back in part 1, we had a look at some of the new features of RSpec, and we used those features to create a query-style controller action. In particular, it listed out all the widgets in our inventory management system. This time around, we’re going to look at a command-style action: creating a new widget.

What do I mean by a “command”-style action, though? I think of these as the kinds of actions where the user is telling the system to do something: create a widget, delete a widget, or send an email, for example. There might be a more complex workflow, too: creating a new user, adding them to an announcement mailing list, and sending them an activation email to confirm their address. That sort of behaviour could be encapsulated by a service object which performs all the actions associated with the business requirements for creating a new user.

They still follow a common pattern:

  • the controller looks up the model on which an action has to be performed, if appropriate;

  • then it tells the model to perform that action, noting the success or failure of the action.

Then we branch in two, depending on whether the model operation was successful. If it was:

  • redirect to another action – often the canonical url for the resource; and

  • set a flash message to let the user know that the operation completed successfully.

If the operation wasn’t successful, then:

  • respond with an appropriate HTTP response code. Somewhat counterintuitively, if this is a web-based form a human is filling in, it will be a 200 OK. If it’s an API, then return an appropriate 4xx response code.

  • re-render the template the user didn’t fill in correctly, passing the appropriate object(s) to that template.

Let’s see this in action. We’ll start out by writing an integration test for creating widgets. First of all, let’s make sure we can find the button:

scenario 'finding the "new widget" button' do
  visit '/widgets'

  click_on 'New widget'

  expect(current_path).to eq('/widgets/new')
end

and implement that by adding the appropriate link on the widget listing page:

<%= link_to 'New widget', new_widget_path %>

Now our tests are failing all over the place, complaining about new_widget_path not existing. We should write a couple of routing specs to check that works:

it 'routes GET /widgets/new to widgets#new' do
  expect(get: '/widgets/new').to route_to('widgets#new')
end

it 'generates /widgets from widgets_path' do
  expect(widgets_path).to eq('/widgets')
end

it 'generates /widgets/new from new_widget_path' do
  expect(new_widget_path).to eq('/widgets/new')
end

and, finally, changing the line in config/routes.rb to include the new action, too:

resources :widgets, only: [:index, :new]

Running our tests reveals that all the other tests are back to passing again (phew!), but that our scenario is still failing, this time because the new action is missing. Let’s start test-driving the implementation of that. The new action is just another query-style action, so I’ll gloss over it quickly:

describe 'GET new' do
  let(:widget) { instance_spy('Widget') }

  before(:each) do
    allow(widget_class).to receive(:new) { widget }
  end

  def do_get
    get :new
  end

  it 'builds a new widget from the model' do
    do_get
    expect(widget_class).to have_received(:new)
  end

  it 'responds with http success' do
    do_get
    expect(response).to have_http_status(:success)
  end

  it 'renders the new template' do
    do_get
    expect(response).to render_template('widgets/new')
  end

  it 'passes the new widget to the template' do
    do_get
    expect(assigns(:widget)).to eq(widget)
  end
end

You start to get a feel for the pattern, right? It all fits in a screenful of code, too. The one other change I made was to pull the stubbing of the Widget class out into the enclosing describe block, so it’s available to every nested describe block.

I have a confession to make here. The point of test-driven development is that you write a single failing test, then write just enough code to make that test pass. Wash, rinse, repeat. However, I did just write all four of these tests at once, resulting in four failing tests. The thing with rules is knowing when it’s OK to break them, right? In this case, I already know the implementation is trivial, with the following new action:

def new
  @widget = Widget.new
end

and the view in app/views/widgets/new.html.erb:

<h1>New Widget</h1>

<%= render 'form', widget: @widget %>

and, finally, the form in app/views/widgets/_form.html.erb:

<%= form_for widget do |f| %>
  <%= f.text_field :name %>
  <%= f.text_field :size %>

  <%= f.submit %>
<% end %>

(I still love that you can express forms so simply, while using a FormBuilder subclass to express the tricky HTML gymnastics required to make forms look pretty.) If it turns out that I’d been over-eager, and I still had failing tests at this point, I’d backtrack a bit, remove (or comment out) some of the tests, and choose a smaller thing to make the tests pass. But it turns out that’s enough to make our controller tests pass, and the initial scenario passes, too. Time for a coffee, then onto the real meat: the create action.

We’ll start with the happy case: a correctly entered, valid, widget:

scenario 'creating a new widget with valid inputs' do
  visit '/widgets/new'
  within '#new_widget' do
    fill_in 'Name', with: 'Bazzle'
    fill_in 'Size', with: '54'
  end
  click_on 'Create Widget'

  expect(current_path).to eq('/widgets')
  expect(page).to have_content('New widget successfully created.')
end

Capybara has a nice, expressive, DSL doesn’t it? I’d like to think that a non-programmer could – with some effort – read and understand what’s going on. (But I’m a programmer, and it’s pretty much my natural language, so I might be way off!)

The scenario is failing, because it can’t find the appropriate routes. Let’s write a test for that route (the url generation for widgets_path has already been covered):

it 'routes POST /widgets to widgets#create' do
  expect(post: '/widgets').to route_to('widgets#create')
end

So now we modify the routes to include the create action:

resources :widgets, only: [:index, :new, :create]

and our integration test is now pointing at the next place we need to work on: the create action is missing. Let’s write some tests for the common behaviour between the happy path and the error case:

describe 'POST create' do
  let(:widget) { instance_spy('Widget', persisted?: true) }
  let(:widget_params) { { name: 'Widget', size: '10' } }
  let(:params) { { widget: widget_params } }

  before(:each) do
    allow(widget_class).to receive(:create) { widget }
  end

  def do_post(params = params)
    post :create, params
  end

  it 'attempts to create a new widget' do
    do_post
    expect(widget_class).to have_received(:create).with(widget_params)
  end

  it 'filters invalid data from the request parameters' do
    do_post params.merge(invalid: 'data')
    expect(widget_class).to have_received(:create).with(widget_params)
  end

  it 'checks to see if the operation was a success' do
    do_post
    expect(widget).to have_received(:persisted?)
  end
end

It turns out that the controller has an additional responsibility after all. As of Rails 4.mumble it now figures out which parameters are valid for this request, so we’ve got a wee spec in there to make sure it’s filtering out invalid (or malicious) data.

There’s a (perfectly valid) argument against stubbing out the behaviour underneath the controller here. What happens when you stub out the implementation you expect, write just enough code to make the tests pass, then discover it doesn’t work in production because the behaviour you stubbed out doesn’t actually exist? In the example above, how do I know that there’s a Widget.create method that takes a hash? Fortunately, RSpec 3 has our backs, with “verifying doubles”. These doubles will, if the object they’re standing in for exists, verify that there really is a method which takes the same number of arguments. Lets fat-finger our before block:

before(:each) do
  allow(widget_class).to receive(:carrot) { widget }
end

Running our tests reveals our mistake:

1) WidgetsController POST create checks to see if the operation was a success
   Failure/Error: allow(widget_class).to receive(:carrot) { widget }
     Widget does not implement: carrot
   # ./spec/controllers/widgets_controller_spec.rb:84:in `block (3 levels) in <top (required)>'
   # -e:1:in `<main>'

The slight snag is that the object being stubbed needs to be around when it is being stubbed, so that rspec can verify the methods really exist. In practice, this means that you’ll only see this sort of feedback when you run your entire test suite, but not necessarily when (for example) Guard is only running the tests that have changed. Still, it’s useful feedback, and another great feature from RSpec 3!

So, what happens on the happy path?

describe 'when the widget is valid' do
  before(:each) do
    allow(widget).to receive(:persisted?) { true }
  end

  it 'redirects back to the list of widgets' do
    do_post
    expect(response).to redirect_to(widgets_path)
  end

  it 'sets a flash message' do
    do_post
    expect(flash.notice).to eq('New widget successfully created.')
  end
end

The before block here is unnecessary – we’ve already stubbed out the persisted? method to return true – but it helps me to make the context of the describe block explicit. Let’s be bold and write some tests for the failure case, too:

describe 'when the widget is invalid' do
  before(:each) do
    allow(widget).to receive(:persisted?) { false }
  end

  it 'renders the new template' do
    do_post
    expect(response).to render_template('widgets/new')
  end

  it 'assigns the faulty widget to the view' do
    do_post
    expect(assigns(:widget)).to eq(widget)
  end
end

and we can write a simple implementation:

def create
  @widget = Widget.create(widget_params)
  if @widget.persisted?
    redirect_to widgets_path, notice: 'New widget successfully created.'
  else
    render 'new'
  end
end

private
def widget_params
  params.require(:widget).permit(:name, :size)
end

Done, our unit tests are passing and, better still, it’s been enough to make the integration test pass, too! Winning. :) Time to commit, push to production, and treat ourselve to a takeout pizza for dinner. As with part 1, you can find the code on GitHub, on the part 2 branch. Tomorrow, let’s have a look at how to test our model – right now, since we don’t have an integration test to demonstrate it, we don’t know that widgets without a name do the right thing (I can honestly say I haven’t tried this project in the browser, but I’m confident that submitting a widget without a name is going to raise an exception, which isn’t the sort of behaviour our client would like.)

A Sneak Peek at
The Internet

If you enjoyed this article, you might be interested in my new project, A Sneak Peek at The Internet. What happens when you enter www.facebook.com into your web browser and hit return? A Sneak Peek at The Internet will take you on a deep dive through the network stack, from HTTP, SSL, TCP and IP, all the way down through the data link layer, back up to Facebook's data centres, and then on the return journey back to the browser.

There's more fun, excitement and peril than a Disneyland rollercoaster!