We'll respond shortly.
This is my second post on the trinity of test tools known as ‘test doubles’. The first covered stubs. This one is all about mocks, which are woefully misunderstood and loathed by many.
If you want to know more about the history of mock objects, get a copy of GOOS. It’s my favourite recent work on TDD and software design, and it’s written by folks involved in the invention of mocks.
Unlike my previous post, I’m going to concentrate on pure mock objects. The disadvantages of partial mocks are similar to those of partial stubs.
First, a definition: a mock is an object that expects messages, and consequently verifies the outputs of the system under test.
Mocks were originally designed to avoid the ‘traditional’ TDD practice of adding code to an object purely for state verification. Their use as a design tool is worth getting to grips with, even if you don’t intend to use them in new code, or if you intend to graduate to something else, which we’ll also touch upon.
You want to mock because you want to write code in a Tell-Don’t-Ask style. You want to be hit in the head whenever you write code in one object that obsesses over the state of another. You want to avoid writing code like this:
class Barista def make_coffee(customer) machine = CoffeeMachine.new if machine.in_use? apologize(to: customer) elsif machine.has_water? tray = Tray.new tray.add_grounds(CoffeeGrounds.new) tray.compact! machine.attach_tray(tray) cup = Cup.new machine.place_cup(cup) status = machine.release_steam! if status.successful? serve(cup, to: customer) else apologize(to: customer) end end end end
If that example was too long, and you didn’t read it: great! You have perfected the art of bad-code-blindness. The Barista is obsessed with details about the labourious process of making an espresso. A real-life Barista arguably should have this disposition, but in code this list of conditions and imperatives is clumsy. If #make_coffee had been designed from the beginning with judicious use of mock objects, so the story goes, the designer would have had more feedback about how bad the code was. Let’s try it:
describe Barista do it "hits the correct button on the coffee machine" do machine = double('coffee making machine') order = Order.new barista = Barista.new(machine) expect(machine).to receive(:brew).with(order) barista.order_received(order) end it "serves coffee when it's ready" do americano = Americano.new fred = double('customer') order = Order.new(fred) barista = Barista.new expect(fred).to receive(:serve).with(americano) barista.order_ready(order, americano) end end
After bashing up against the rigidity of the mock (try writing a test like this yourself to see how it feels), we’ve thought about the domain a little bit, and decided that the events in the system are the decision making points. With this in mind, we’ve let the Barista trust the machine to perform its role’s responsibilities. We expect the Barista to be alerted when coffee is ready, so that he/she can focus on taking more orders in the meantime. Let’s see how this could be implemented:
class Barista def initialize(machine = nil) @machine = machine end def order_received(order) @machine.brew(order) end def order_ready(order, product) order.customer.serve(product) end end Order = Struct.new(:customer) Americano = Struct.new(:strength)
The Barista isn’t perfect, but it is far easier to read, and it’s obvious that there are more behavioural components in the collaborators. The example would need further fleshing out if it were to achieve the same effect as the imperative code we began with.
Mocking isn’t always appropriate. In general, mocking makes sense when you are testing an object that depends on another object’s behaviour. If the collaborating object has no behaviour, you probably shouldn’t be expecting that messages are sent to it.
Value objects (the Structs in the above example) should not be mocked, but simply instantiated and used in tests. Value objects are just bags of data, whose equality depends on the values they contain, rather than on some unique ID, as would be the case for an Entity. Your tests should fail if you are using values inconsistently across your suite. Values include Money, Distance, Specification and so on. Some concepts are values in one application and entities in another. See DDD for a detailed study of these definitions.
You may have heard the mantra “Don’t mock types you don’t own”. This is a guideline to help avoid pain around changes in external dependencies, such as libraries. For example, if your application talks to AWS, you should create a wrapper around the external library that does the work of communicating with AWS. This interface is under your control and is ‘safe’ to mock. Mocking the AWS library’s interface directly is likely to result in a lot of broken code when the library changes. GOOS, linked above, has a good treatment of this.
Another mantra from GOOS is “mock roles, not objects”. In practice, for me this means choosing a name for the mock in a test that refers generally to the role a collaborator is playing. So, I would sooner
double('payment gateway') than
double('paypal'). As soon as you mock PayPal, you are in the mindset of pleasing the external service, with its crazy API, rather than providing a neat abstraction whose messages are relevant to your particular domain.
Gary Bernhardt is leading a charge against mock objects that is much more informed than the usual arguments you may hear (e.g. “I don’t mock because it’s brittle” – there are ways to mitigate this). In a poorly paraphrased nutshell, I believe he’s saying that we can use value objects as the inputs and outputs of objects that we’d otherwise mock. If these objects are immutable, then we get a cheap way to enforce an interface, and we can ‘isolate’ units of code by testing against these values instead of expecting messages. In effect, I think the values become the messages.