Comment by IanCal | Hacker News Reader

Comment by IanCal | Hacker News Reader

Parent Full thread

I kind of agree with this, but then I see the same tests I've seen elsewhere:

      Scenario: Editing the nickname of a credential does not change the version
        When the nickname is modified
        Then the edit version of the credential should not change

This looks like an awesome test. Simple, clear, anyone can read it. What's wrong? One of the following:

* It's based on timing and unreliable

* It stores state and the steps can't be reused.

You simply cannot put the line "Then the edit version of the credentials should not change" elsewhere because the step involves the previous state of the system. This step can only come after this step:

    Given a credential for a supported vendor

These tests all look like they're testing really nice general properties, but they aren't. They're almost always just checking little properties here and there, and rarely doing just what the step says. Which could be fine, but the language doesn't support any idea of specified dependencies. I cannot change a test without understanding the code behind it.

What you end up with is a big codebase based on regexes, with very heavy (non-specified) dependencies. You have to be extremely careful not to end up like this, although the tests look more complex afterwards (but you have re-usable steps).

I think there's a great place for a quickcheck & cucumber combination.

I don't quite get how any when/then could NOT be dependent upon previous givens.

Also, despite the job-centric nature of the .feature, it wasn't timing-dependent because we mocked Resque to keep an in-process queue that we could dispatch in whatever order we want. So out-of-order arrival was deterministically achievable, etc.

And yes, that test is testing a simple property -- it's effectively a unit test, with greater clarity of intent.

Not familiar with quickcheck, but I will look into it, thanks!

> I don't quite get how any when/then could NOT be dependent upon previous givens.

    Given I am called Bob
    And I change my name to Steve
    Then my name is not Bob

None of these steps are dependent on the ones previous to it, and can be reused elsewhere. The check for "my name is bob" can be used whether or not the previous step has been called. Your test won't pass, which is fine, but it will run and do exactly what it says. The step does what it says even if none of the others are used. It also means I can write a new test using bits from others without worrying about what's in the source code.

A test that looks like this:

    Given I am a person
    And I change my name
    Then my name is different

Cannot be re-used. The "my name is different" can only be used if you know the ruby code underneath. "my name is different" may not do what you expect unless it comes after "I change my name".

What I meant with timing is the only way you can get around storing state is to have the "my name is different" watch for a change in the name, which brings in timing issues. I've seen both implementations used, and both caused problems (the timing one for obvious reasons).

> Not familiar with quickcheck, but I will look into it, thanks!

I thoroughly recommend it. The haskell version is probably the most advanced, but there are similar versions for most languages (and you can write your own, I had to for AS3). The idea is you express general properties about your system, and then it auto-generates thousands of examples (and if you have a nice library, automatically shrink failing cases for you). For example, the test above would be nicer as something like:

    X is a string
    Y is a string
    X =/= Y

    Given I am called X
    And I change my name to Y
    Then my name is not X
    And my name is Y

Or something like that. This would then generate examples with no-length strings, crazy unicode characters, long strings, different mixings of RTL sections, etc. Much more likely to drive out bugs than a test for Bob and Steve.

Some more useful ones would look like this (the first is a test I've written before, but not in this format):

    X is a number >= 1
    Y is an interface element
    Given I am on element Y
    When I press Tab X times
    And I press Shift-Tab X times
    Then I am on element Y

A similar version for navigating in a website and pressing "back" to get back to where you were (to ensure you're not breaking the back button).

These are really simple, but powerful tests. I was most sold on the idea when I wrote this (not in this format, but this logic):

    X is a positive integer
    Y is a positive integer
    INSTRUCTION is one of [addElement, removeElement(X), setFocus(X)]
    MENU is an interface
    When I perform Y INSTRUCTIONS
    Then MENU has one focused item or no items at all

This then generated thousands of menus of each valid type (vertical, horizontal, grids, etc) and then called library functions to add or remove elements, or move focus. Millions of tests overall. I was using this as a test for my quickcheck implementation, and found it failed. If I set the focus, deleted all the elements and then added a single new one it wasn't focused.

When I fixed it, a unit test failed. We has previously specified that was to be the behaviour, but also specified that no matter what there would always be a focused element (if there were any elements at all). The general test drew out an inconsistency in our spec because we were forced to write general rules.

Mixing quickcheck and cucumber has been one of my "Some weekend I'll do it" projects for a couple of years now.

So your example still requires that state exist, it's just far more global. I.E. there must be some prerequisite code that sets up the thing to which "I" refers such that it's available to any step at any time. There may be some good patterns for that -- and that approach may be an excellent way to keep step definitions clean and simple, but I think it's somewhat orthogonal to my point.

I'm still not with you on the timing thing. By the time that step is reached, either the state has changed, or it hasn't. We address the need for it to be able to see "back in time" by storing a history of the state of the object in a way that is accessible to the steps. I.E. For steps like this:

  Given a credential for a supported vendor
  When I change the nickname
  Then the edit version should not be changed

The step definitions might look like:

  Given /a credential for a supported vendor/ do
    @thing = FactoryGirl.create(:'credential/amazon', :valid)
    @thing_history = [@thing.attributes.dup]
  end

  When /I change the nickname/ do
    @thing.nickname += " meh"
    @thing.save!
    @thing_history << @thing.attributes.dup
  end

  Then /the edit version should not be changed/ do
    @thing_history[-1][:edit_version].should == @thing_history[0][:edit_version]
  end

(The use of "@thing" is a way to encourage myself to only be talking about one object at a time...)

Perhaps that's a somewhat obtuse way of handling things -- and there's likely ways of DRYing up that pattern, if it's worth preserving, but it's proven to be fairly effective for me in the past.

And yeah, Cucumber might make a good vehicle for that sort of testing. Quickcheck sounds like a brilliant idea -- although a bit terrifying in the context of a slow language like Ruby. Would love to get my hands on / build a tool like that...

> So your example still requires that state exist, it's just far more global.

The state exists only in your application, not in the test. Your given sets up the environment in the way you want, the "when" manipulates the environment and the "then" checks that the environment exists in a particular setting.

> I'm still not with you on the timing thing. By the time that step is reached, either the state has changed, or it hasn't.

This doesn't apply to your tests. What I've seen before is a step that says

    Then X changes

Which is implemented as

    previous = X.state
    wait 5 seconds
    X.state != previous

This relies on the previous step taking long enough to actually change the state that the change happens during the wait. All of the tests I was dealing with were asynchronous.

Your pattern does look quite nice but there are still dependencies that aren't specified in code. For example, assuming you have a complementary step about the edit version changing written in the same way.

    Given a credential for a supported vendor
    When I do something that changes the edit version
    Then the edit version should be changed
    When I do something irrelevant
    Then the edit version should not be changed

This would fail, saying the edit version has changed. Contrived, I know, but the problem with dependencies that aren't specified is that you start having valid tests that fail because they simply do not do what they say. I'm strongly in the camp of "If it shouldn't work, it shouldn't build".

My core suggestion for all of this is to write your behavioural tests as a test script. What does the user do to get to that point, how do they interact and what's the result. The state exists either in your application (which is fine) or is explicit in the writing of the tests (Given I log in as person X rather than Given I log in).

Cucumber is a great tool, but it leaves a lot of things implicit, which to be fair could probably be solved as a library. I'd have much less of a problem with all of this if when I wrote an invalid test, something warned me. I've spent far too much time battling with cucumber tests which lied about what they were actually testing.

> The state exists only in your application, not in the test. > Your given sets up the environment in the way you want, the > "when" manipulates the environment and the "then" checks > that the environment exists in a particular setting.

Err, are you asserting that that is what I am doing, or what I ought to be doing?

And yes, time-dependent code is evil. I should probably add commentary to my style-guide to explicitly call that out, but thankfully we never ran into that despite testing of distributed-job-queue functionality, by virtue of having a queue so simple stubbing its main loop to work in-process in a deterministic way was trivial.

Your example should be in violation of the style guide for precisely the reason you state, among several others (blurring of concerns, etc). If my style guide isn't clear on that point -- that a When should NEVER follow a Then -- I need to clarify that. :)

I recall a Gherkin-based testing framework that handled the actual step definitions MUCH differently and much more cleanly but I never got around to fully switching us over and don't recall the name... :-S

> Err, are you asserting that that is what I am doing, or what I ought to be doing?

Ought to be. Your Given and When manipulate the environment and store state in the test. Your then doesn't check the environment at all, it checks it's own internal state. I think in your example this is less of an issue than the cases I've worked with before (usually testing remote running apps).

> And yes, time-dependent code is evil. I should probably add commentary to my style-guide to explicitly call that out, but thankfully we never ran into that despite testing of distributed-job-queue functionality, by virtue of having a queue so simple stubbing its main loop to work in-process in a deterministic way was trivial.

Yes, as I say I wasn't claiming that was what you were doing, but it's the only other way I've seen the same kind of tests written. Before the first reply, I didn't know which it would be. I'm very glad it wasn't :)

> If my style guide isn't clear on that point -- that a When should NEVER follow a Then -- I need to clarify that. :)

My example was contrived, but I think you can probably see the point I'm making. What the test does is not clear without understanding the ruby underneath, and by storing state you can have tests which don't do what you expect because they aren't just querying the state of your application.

I'm aware I've been quite ranty about this, it's mostly I've had to deal with bad tests written in this style. Good tests written in this style seem fine generally. I think this is more of an issue with cucumber than the tests themselves.

> I recall a Gherkin-based testing framework that handled the actual step definitions MUCH differently and much more cleanly but I never got around to fully switching us over and don't recall the name... :-S

I might have a look around. It's something I've probably spent more time writing about it than it would have spent trying to write something better (/at least with dependency docs, maybe quickcheck style).

Given that the system these scenarios test also manage state, it's not surprising that the step definitions used in the scenario also need to manage some form of state.

While it's advisable to avoid unnecessary state between steps in a scenario like this, removing all state limits the ability for a scenario to express the intent the author speaks about.

One option is to move the state to be explicit in the call to the step, which means the test becomes an example rather than a general rule (although that's what it always was, now it's just a bit clearer). It also means the steps can be reused, rather than having steps with dependencies (which aren't specified anywhere or statically enforced).

> removing all state limits the ability for a scenario to express the intent the author speaks about.

However, writing them like that means each step moves further away from what it says it's doing. The more you do this, the more you might as well not execute them and go and write more examples.

To use your example -- how desirable is it to generalize "I change my name" to something "reusable"?

I'm finding these days that a pairing of .feature to corresponding step definition file allows for fluid language, simple step definition code, and well-compartmentalized messes (as opposed to the big ball of mud that Cucumber step definitions often become). Code reuse can happen through test helper classes that are called from step definitions as needed. (FactoryGirl would be an example of a very generalized test helper class, in my mind -- but more system-specific ones can be handy, albeit one wants to avoid making stateful / complex test helper code!)

Reply on news.ycombinator.com