In the preceding chapter, we made the decision to leave a unit test failing in the views layer while we proceeded to write more tests and more code at the models layer to get it to pass.
We got away with it because our app was simple, but I should stress that, in a more complex application, this would be a dangerous decision. Proceeding to work on lower levels while you’re not sure that the higher levels are 'really' finished or not is a risky strategy.
|
Note
|
I’m grateful to Gary Bernhardt, who took a look at an early draft of the previous chapter, and encouraged me to get into a longer discussion of test isolation. |
Ensuring isolation between layers does involve more effort (and more of the dreaded mocks!), but it can also help to drive out improved design, as we’ll see in this chapter.
Let’s
revisit the point we were at halfway through the last chapter, when we
couldn’t get the new_list view to work because lists didn’t have the .owner
attribute yet.
We’ll actually go back in time and check out the old codebase using the tag we saved earlier, so that we can see how things would have worked if we’d used more isolated tests:
$ git checkout -b more-isolation # a branch for this experiment $ git reset --hard revisit_this_point_with_isolated_tests
Here’s what our failing test looks like:
class NewListTest(TestCase):
[...]
def test_list_owner_is_saved_if_user_is_authenticated(self):
user = User.objects.create(email='[email protected]')
self.client.force_login(user)
self.client.post('/lists/new', data={'text': 'new item'})
list_ = List.objects.first()
self.assertEqual(list_.owner, user)And here’s what our attempted solution looked like:
def new_list(request):
form = ItemForm(data=request.POST)
if form.is_valid():
list_ = List()
list_.owner = request.user
list_.save()
form.save(for_list=list_)
return redirect(list_)
else:
return render(request, 'home.html', {"form": form})And at this point, the view test is failing because we don’t have the model layer yet:
self.assertEqual(list_.owner, user) AttributeError: 'List' object has no attribute 'owner'
|
Note
|
You won’t see this error unless you actually check out the old code and revert 'lists/models.py'. You should definitely do this; part of the objective of this chapter is to see whether we really can write tests for a models layer that doesn’t exist yet. |
Lists don’t have owners yet, but we can let the views layer tests pretend they do by using a bit of mocking:
from unittest.mock import patch
[...]
@patch('lists.views.List') #(1)
@patch('lists.views.ItemForm') #(2)
def test_list_owner_is_saved_if_user_is_authenticated(
self, mockItemFormClass, mockListClass #(3)
):
user = User.objects.create(email='[email protected]')
self.client.force_login(user)
self.client.post('/lists/new', data={'text': 'new item'})
mock_list = mockListClass.return_value #(4)
self.assertEqual(mock_list.owner, user) #(5)-
We mock out the
Listclass to be able to get access to any lists that might be created by the view. -
We also mock out the
ItemForm. Otherwise, our form will raise an error when we callform.save(), because it can’t use a mock object as the foreign key for the Item it wants to create. Once you start mocking, it can be hard to stop! -
The mock objects are injected into the test’s arguments in the opposite order to which they’re declared. Tests with lots of mocks often have this strange signature, with the dangling
):. You get used to it! -
The list instance that the view will have access to will be the return value of the mocked
Listclass. -
And we can make assertions about whether the
.ownerattribute is set on it.
If we try to run this test now, it should pass:
$ python manage.py test lists [...] Ran 37 tests in 0.145s OK
If you don’t see a pass, make sure that your views code in 'views.py' is
exactly as I’ve shown it, using List(), not List.objects.create.
|
Note
|
Using mocks does tie you to specific ways of using an API. This is one of the many trade-offs involved in the use of mock objects. |
The trouble with this test is that it can still let us get away with writing the wrong code by mistake. Imagine if we accidentally call save before we we assign the owner:
if form.is_valid():
list_ = List()
list_.save()
list_.owner = request.user
form.save(for_list=list_)
return redirect(list_)The test, as it’s written now, still passes:
OK
So strictly speaking, we need to check not just that the owner is assigned, but that it’s assigned 'before' we call save on our list object.
Here’s how we could test the sequence of events using mocks—you can mock out a function, and use it as a spy to check on the state of the world at the moment it’s called:
@patch('lists.views.List')
@patch('lists.views.ItemForm')
def test_list_owner_is_saved_if_user_is_authenticated(
self, mockItemFormClass, mockListClass
):
user = User.objects.create(email='[email protected]')
self.client.force_login(user)
mock_list = mockListClass.return_value
def check_owner_assigned(): #(1)
self.assertEqual(mock_list.owner, user)
mock_list.save.side_effect = check_owner_assigned #(2)
self.client.post('/lists/new', data={'text': 'new item'})
mock_list.save.assert_called_once_with() #(3)-
We define a function that makes the assertion about the thing we want to happen first: checking that the list’s owner has been set.
-
We assign that check function as a
side_effectto the thing we want to check happened second. When the view calls our mocked save function, it will go through this assertion. We make sure to set this up before we actually call the function we’re testing. -
Finally, we make sure that the function with the
side_effectwas actually triggered—that is, that we did.save(). Otherwise, our assertion may actually never have been run.
|
Tip
|
Two common mistakes when you’re using mock side effects are assigning the side effect too late (i.e., 'after' you call the function under test), and forgetting to check that the side-effect function was actually called. And by common, I mean, "I made both these mistakes several times while writing this chapter.” |
At this point, if you’ve still got the "broken" code from earlier, where we assign the owner but call save in the wrong order, you should now see a fail:
FAIL: test_list_owner_is_saved_if_user_is_authenticated
(lists.tests.test_views.NewListTest)
[...]
File "...python-tdd-book/lists/views.py", line 17, in new_list
list_.save()
[...]
File "...python-tdd-book/lists/tests/test_views.py", line 74, in
check_owner_assigned
self.assertEqual(mock_list.owner, user)
AssertionError: <MagicMock name='List().owner' id='140691452447208'> != <User:
User object>
Notice how the failure happens when we try to save, and then go inside
our side_effect function.
We can get it passing again like this:
if form.is_valid():
list_ = List()
list_.owner = request.user
list_.save()
form.save(for_list=list_)
return redirect(list_)…
OK
But, boy, that’s getting to be an ugly test!
Whenever you find yourself having to write a test like this, and you’re finding it hard work, it’s likely that your tests are trying to tell you something. Eight lines of setup (two lines for mocks, three to set up a user, and three more for our side-effect function) is way too many.
What this test is trying to tell us is that our view is doing too much work, dealing with creating a form, creating a new list object, 'and' deciding whether or not to save an owner for the list.
We’ve already seen that we can make our views simpler and easier to understand
by pushing some of the work down to a form class. Why does the view need to
create the list object? Perhaps our ItemForm.save could do that? And why
does the view need to make decisions about whether or not to save the
request.user? Again, the form could do that.
While we’re giving this form more responsibilities, it feels like it should
probably get a new name too. We could call it NewListForm instead, since
that’s a better representation of what it does…something like this?
# don't enter this code yet, we're only imagining it.
def new_list(request):
form = NewListForm(data=request.POST)
if form.is_valid():
list_ = form.save(owner=request.user) # creates both List and Item
return redirect(list_)
else:
return render(request, 'home.html', {"form": form})That would be neater! Let’s see how we’d get to that state by using fully isolated tests.
Our first attempt at a test suite for this view was highly 'integrated'. It needed the database layer and the forms layer to be fully functional in order for it to pass. We’ve started trying to make it more isolated, so let’s now go all the way.
Let’s rename our old NewListTest class to NewListViewIntegratedTest,
and throw away our attempt at a mocky test for saving the owner, putting
back the integrated version, with a skip on it for now:
import unittest
[...]
class NewListViewIntegratedTest(TestCase):
def test_can_save_a_POST_request(self):
[...]
@unittest.skip
def test_list_owner_is_saved_if_user_is_authenticated(self):
user = User.objects.create(email='[email protected]')
self.client.force_login(user)
self.client.post('/lists/new', data={'text': 'new item'})
list_ = List.objects.first()
self.assertEqual(list_.owner, user)|
Tip
|
Have you heard the term "integration test" and are wondering what the difference is from an "integrated test"? Go and take a peek at the definitions box in [chapter_hot_lava]. |
$ python manage.py test lists [...] Ran 37 tests in 0.139s OK
Let’s start with a blank slate, and see if we can use isolated tests to drive
a replacement of our new_list view. We’ll call it new_list2, build it
alongside the old view, and when we’re ready, swap it in and see if
the old integrated tests all still pass:
def new_list(request):
[...]
def new_list2(request):
passIn order to rewrite our tests to be fully isolated, we need to throw out our old way of thinking about the tests in terms of the "real" effects of the view on things like the database, and instead think of it in terms of the objects it collaborates with, and how it interacts with them.
In the new world, the view’s main collaborator will be a form object, so we mock that out in order to be able to fully control it, and in order to be able to define, by wishful thinking, the way we want our form to work:
from unittest.mock import patch
from django.http import HttpRequest
from lists.views import new_list2
[...]
@patch('lists.views.NewListForm') #(2)
class NewListViewUnitTest(unittest.TestCase): #(1)
def setUp(self):
self.request = HttpRequest()
self.request.POST['text'] = 'new list item' #(3)
def test_passes_POST_data_to_NewListForm(self, mockNewListForm):
new_list2(self.request)
mockNewListForm.assert_called_once_with(data=self.request.POST) #(4)-
The Django
TestCaseclass makes it too easy to write integrated tests. As a way of making sure we’re writing "pure", isolated unit tests, we’ll only useunittest.TestCase. -
We mock out the NewListForm class (which doesn’t even exist yet). It’s going to be used in all the tests, so we mock it out at the class level.
-
We set up a basic POST request in
setUp, building up the request by hand rather than using the (overly integrated) Django Test Client. -
And we check the first thing about our new view: it initialises its collaborator, the
NewListForm, with the correct constructor—the data from the request.
That will start with a failure, saying we don’t have a NewListForm in
our view yet:
AttributeError: <module 'lists.views' from '...python-tdd-book/lists/views.py'> does not have the attribute 'NewListForm'
Let’s create a placeholder for it:
from lists.forms import ExistingListItemForm, ItemForm, NewListForm
[...]and:
class ItemForm(forms.models.ModelForm):
[...]
class NewListForm(object):
pass
class ExistingListItemForm(ItemForm):
[...]Next we get a real failure:
AssertionError: Expected 'NewListForm' to be called once. Called 0 times.
And we implement like this:
def new_list2(request):
NewListForm(data=request.POST)$ python manage.py test lists [...] Ran 38 tests in 0.143s OK
Let’s continue. If the form is valid, we want to call save on it:
from unittest.mock import patch, Mock
[...]
@patch('lists.views.NewListForm')
class NewListViewUnitTest(unittest.TestCase):
def setUp(self):
self.request = HttpRequest()
self.request.POST['text'] = 'new list item'
self.request.user = Mock()
def test_passes_POST_data_to_NewListForm(self, mockNewListForm):
new_list2(self.request)
mockNewListForm.assert_called_once_with(data=self.request.POST)
def test_saves_form_with_owner_if_form_valid(self, mockNewListForm):
mock_form = mockNewListForm.return_value
mock_form.is_valid.return_value = True
new_list2(self.request)
mock_form.save.assert_called_once_with(owner=self.request.user)That takes us to this:
def new_list2(request):
form = NewListForm(data=request.POST)
form.save(owner=request.user)In the case where the form is valid, we want the view to return a redirect,
to send us to see the object that the form has just created. So we mock out
another of the view’s collaborators, the redirect function:
@patch('lists.views.redirect') #(1)
def test_redirects_to_form_returned_object_if_form_valid(
self, mock_redirect, mockNewListForm #(2)
):
mock_form = mockNewListForm.return_value
mock_form.is_valid.return_value = True #(3)
response = new_list2(self.request)
self.assertEqual(response, mock_redirect.return_value) #(4)
mock_redirect.assert_called_once_with(mock_form.save.return_value) #(5)-
We mock out the
redirectfunction, this time at the method level. -
patchdecorators are applied innermost first, so the new mock is injected to our method before themockNewListForm. -
We specify that we’re testing the case where the form is valid.
-
We check that the response from the view is the result of the
redirectfunction. -
And we check that the redirect function was called with the object that the form returns on save.
That takes us to here:
def new_list2(request):
form = NewListForm(data=request.POST)
list_ = form.save(owner=request.user)
return redirect(list_)$ python manage.py test lists [...] Ran 40 tests in 0.163s OK
And now the failure case—if the form is invalid, we want to render the home page template:
@patch('lists.views.render')
def test_renders_home_template_with_form_if_form_invalid(
self, mock_render, mockNewListForm
):
mock_form = mockNewListForm.return_value
mock_form.is_valid.return_value = False
response = new_list2(self.request)
self.assertEqual(response, mock_render.return_value)
mock_render.assert_called_once_with(
self.request, 'home.html', {'form': mock_form}
)That gives us:
AssertionError: <HttpResponseRedirect status_code=302, "te[114 chars]%3E"> != <MagicMock name='render()' id='140244627467408'>
|
Tip
|
When using assert methods on mocks, like assert_called_​once_with,
it’s doubly important to make sure you run the test and see it fail.
It’s all too easy to make a typo in your assert function name and
end up calling a mock method that does nothing (mine was to write
asssert_called_once_with with three essses; try it!).
|
We make a deliberate mistake, just to make sure our tests are comprehensive:
def new_list2(request):
form = NewListForm(data=request.POST)
list_ = form.save(owner=request.user)
if form.is_valid():
return redirect(list_)
return render(request, 'home.html', {'form': form})That passes, but it shouldn’t! One more test then:
def test_does_not_save_if_form_invalid(self, mockNewListForm):
mock_form = mockNewListForm.return_value
mock_form.is_valid.return_value = False
new_list2(self.request)
self.assertFalse(mock_form.save.called)Which fails:
self.assertFalse(mock_form.save.called) AssertionError: True is not false
And we get to to our neat, small finished view:
def new_list2(request):
form = NewListForm(data=request.POST)
if form.is_valid():
list_ = form.save(owner=request.user)
return redirect(list_)
return render(request, 'home.html', {'form': form})…
$ python manage.py test lists [...] Ran 42 tests in 0.163s OK
So
we’ve built up our view function based on a "wishful thinking" version
of a form called NewListForm, which doesn’t even exist yet.
We’ll need the form’s save method to create a new list, and a new item based on the text from the form’s validated POST data. If we were to just dive in and use the ORM, the code might look something a bit like this:
class NewListForm(models.Form):
def save(self, owner):
list_ = List()
if owner:
list_.owner = owner
list_.save()
item = Item()
item.list = list_
item.text = self.cleaned_data['text']
item.save()This implementation depends on two classes from the model layer, Item and
List. So, what would a well-isolated test look like?
class NewListFormTest(unittest.TestCase):
@patch('lists.forms.List') #(1)
@patch('lists.forms.Item') #(1)
def test_save_creates_new_list_and_item_from_post_data(
self, mockItem, mockList #(1)
):
mock_item = mockItem.return_value
mock_list = mockList.return_value
user = Mock()
form = NewListForm(data={'text': 'new item text'})
form.is_valid() #(2)
def check_item_text_and_list():
self.assertEqual(mock_item.text, 'new item text')
self.assertEqual(mock_item.list, mock_list)
self.assertTrue(mock_list.save.called)
mock_item.save.side_effect = check_item_text_and_list #(3)
form.save(owner=user)
self.assertTrue(mock_item.save.called) #(4)-
We mock out the two collaborators for our form from the models layer below.
-
We need to call
is_valid()so that the form populates the.cleaned_datadictionary where it stores validated data. -
We use the
side_effectmethod to make sure that, when we save the new item object, we’re doing so with a savedListand with the correct item text. -
As always, we double-check that our side-effect function was actually called.
Yuck! What an ugly test! Let’s not even bother saving that to disk, we can do better.
Again, these tests are trying to tell us something: the Django ORM is hard to mock out, and our form class needs to know too much about how it works. Programming by wishful thinking again, what would be a simpler API that our form could use? How about something like this:
def save(self):
List.create_new(first_item_text=self.cleaned_data['text'])Our wishful thinking says: how about a helper method that
would live on the List
class[1]
and encapsulate all the logic of saving a new list object and
its associated first item?
So let’s write a test for that instead:
import unittest
from unittest.mock import patch, Mock
from django.test import TestCase
from lists.forms import (
DUPLICATE_ITEM_ERROR, EMPTY_ITEM_ERROR,
ExistingListItemForm, ItemForm, NewListForm
)
from lists.models import Item, List
[...]
class NewListFormTest(unittest.TestCase):
@patch('lists.forms.List.create_new')
def test_save_creates_new_list_from_post_data_if_user_not_authenticated(
self, mock_List_create_new
):
user = Mock(is_authenticated=False)
form = NewListForm(data={'text': 'new item text'})
form.is_valid()
form.save(owner=user)
mock_List_create_new.assert_called_once_with(
first_item_text='new item text'
)And while we’re at it, we can test the case where the user is an authenticated user too:
@patch('lists.forms.List.create_new')
def test_save_creates_new_list_with_owner_if_user_authenticated(
self, mock_List_create_new
):
user = Mock(is_authenticated=True)
form = NewListForm(data={'text': 'new item text'})
form.is_valid()
form.save(owner=user)
mock_List_create_new.assert_called_once_with(
first_item_text='new item text', owner=user
)You can see this is a much more readable test. Let’s start implementing our new form. We start with the import:
from lists.models import Item, ListNow mock tells us to create a placeholder for our create_new method:
AttributeError: <class 'lists.models.List'> does not have the attribute 'create_new'
class List(models.Model):
def get_absolute_url(self):
return reverse('view_list', args=[self.id])
def create_new():
passAnd after a few steps, we should end up with a form save method like this:
class NewListForm(ItemForm):
def save(self, owner):
if owner.is_authenticated:
List.create_new(first_item_text=self.cleaned_data['text'], owner=owner)
else:
List.create_new(first_item_text=self.cleaned_data['text'])And passing tests:
$ python manage.py test lists Ran 44 tests in 0.192s OK
One of the techniques that emerged from our use of isolated tests was the "ORM helper method".
Django’s ORM lets you get things done quickly with a reasonably readable syntax (it’s certainly much nicer than raw SQL!). But some people like to try to minimise the amount of ORM code in the application—particularly removing it from the views and forms layers.
One reason is that it makes it much easier to test those layers. But another is that it forces us to build helper functions that express our domain logic more clearly. Compare:
list_ = List()
list_.save()
item = Item()
item.list = list_
item.text = self.cleaned_data['text']
item.save()With:
List.create_new(first_item_text=self.cleaned_data['text'])This applies to read queries as well as write. Imagine something like this:
Book.objects.filter(in_print=True, pub_date__lte=datetime.today())Versus a helper method, like:
Book.all_available_books()When we build helper functions, we can give them names that express what we are doing in terms of the business domain, which can actually make our code more legible, as well as giving us the benefit of keeping all ORM calls at the model layer, and thus making our whole application more loosely coupled.
At the models layer, we no longer need to write isolated tests—the whole point of the models layer is to integrate with the database, so it’s appropriate to write integrated tests:
class ListModelTest(TestCase):
def test_get_absolute_url(self):
list_ = List.objects.create()
self.assertEqual(list_.get_absolute_url(), f'/lists/{list_.id}/')
def test_create_new_creates_list_and_first_item(self):
List.create_new(first_item_text='new item text')
new_item = Item.objects.first()
self.assertEqual(new_item.text, 'new item text')
new_list = List.objects.first()
self.assertEqual(new_item.list, new_list)Which gives:
TypeError: create_new() got an unexpected keyword argument 'first_item_text'
And that will take us to a first cut implementation that looks like this:
class List(models.Model):
def get_absolute_url(self):
return reverse('view_list', args=[self.id])
@staticmethod
def create_new(first_item_text):
list_ = List.objects.create()
Item.objects.create(text=first_item_text, list=list_)Notice we’ve been able to get all the way down to the models layer,
driving a nice design for the views and forms layers, and the List
model still doesn’t support having an owner!
Now let’s test the case where the list should have an owner, and add:
from django.contrib.auth import get_user_model
User = get_user_model()
[...]
def test_create_new_optionally_saves_owner(self):
user = User.objects.create()
List.create_new(first_item_text='new item text', owner=user)
new_list = List.objects.first()
self.assertEqual(new_list.owner, user)And while we’re at it, we can write the tests for the new owner attribute:
class ListModelTest(TestCase):
[...]
def test_lists_can_have_owners(self):
List(owner=User()) # should not raise
def test_list_owner_is_optional(self):
List().full_clean() # should not raiseThese two are almost exactly the same tests we used in the last chapter, but I’ve re-written them slightly so they don’t actually save objects—just having them as in-memory objects is enough for this test.
|
Tip
|
Use in-memory (unsaved) model objects in your tests whenever you can; it makes your tests faster. |
That gives:
$ python manage.py test lists [...] ERROR: test_create_new_optionally_saves_owner TypeError: create_new() got an unexpected keyword argument 'owner' [...] ERROR: test_lists_can_have_owners (lists.tests.test_models.ListModelTest) TypeError: 'owner' is an invalid keyword argument for this function [...] Ran 48 tests in 0.204s FAILED (errors=2)
We implement, just like we did in the last chapter:
from django.conf import settings
[...]
class List(models.Model):
owner = models.ForeignKey(settings.AUTH_USER_MODEL, blank=True, null=True)
[...]That will give us the usual integrity failures, until we do a migration:
django.db.utils.OperationalError: no such column: lists_list.owner_id
Building the migration will get us down to three failures:
ERROR: test_create_new_optionally_saves_owner TypeError: create_new() got an unexpected keyword argument 'owner' [...] ValueError: Cannot assign "<SimpleLazyObject: <django.contrib.auth.models.AnonymousUser object at 0x7f5b2380b4e0>>": "List.owner" must be a "User" instance. ValueError: Cannot assign "<SimpleLazyObject: <django.contrib.auth.models.AnonymousUser object at 0x7f5b237a12e8>>": "List.owner" must be a "User" instance.
Let’s deal with the first one, which is for our create_new method:
@staticmethod
def create_new(first_item_text, owner=None):
list_ = List.objects.create(owner=owner)
Item.objects.create(text=first_item_text, list=list_)Two of our old integrated tests for the views layer are failing. What’s happening?
ValueError: Cannot assign "<SimpleLazyObject: <django.contrib.auth.models.AnonymousUser object at 0x7fbad1cb6c10>>": "List.owner" must be a "User" instance.
Ah, the old view isn’t discerning enough about what it does with list owners yet:
if form.is_valid():
list_ = List()
list_.owner = request.user
list_.save()This is the point at which we realise that our old code wasn’t fit for purpose. Let’s fix it to get all our tests passing:
def new_list(request):
form = ItemForm(data=request.POST)
if form.is_valid():
list_ = List()
if request.user.is_authenticated:
list_.owner = request.user
list_.save()
form.save(for_list=list_)
return redirect(list_)
else:
return render(request, 'home.html', {"form": form})
def new_list2(request):
[...]|
Note
|
One of the benefits of integrated tests is that they help you to catch less predictable interactions like this. We’d forgotten to write a test for the case where the user is not authenticated, but because the integrated tests use the stack all the way down, errors from the model layer came up to let us know we’d forgotten something: |
$ python manage.py test lists [...] Ran 48 tests in 0.175s OK
So let’s try switching out our old view, and activating our new view. We can make the swap in 'urls.py':
[...]
url(r'^new$', views.new_list2, name='new_list'),We should also remove the unittest.skip from our integrated test class, to
see if our new code for list owners really works:
class NewListViewIntegratedTest(TestCase):
def test_can_save_a_POST_request(self):
[...]
def test_list_owner_is_saved_if_user_is_authenticated(self):
[...]
self.assertEqual(list_.owner, user)So what happens when we run our tests? Oh no!
ERROR: test_list_owner_is_saved_if_user_is_authenticated
[...]
ERROR: test_can_save_a_POST_request
[...]
ERROR: test_redirects_after_POST
(lists.tests.test_views.NewListViewIntegratedTest)
File "...python-tdd-book/lists/views.py", line 30, in new_list2
return redirect(list_)
[...]
TypeError: argument of type 'NoneType' is not iterable
FAILED (errors=3)
Here’s an important lesson to learn about test isolation: it might help you to drive out good design for individual layers, but it won’t automatically verify the integration 'between' your layers.
What’s happened here is that the view was expecting the form to return a list item:
list_ = form.save(owner=request.user)
return redirect(list_)But we forgot to make it return anything:
def save(self, owner):
if owner.is_authenticated:
List.create_new(first_item_text=self.cleaned_data['text'], owner=owner)
else:
List.create_new(first_item_text=self.cleaned_data['text'])Ultimately, even if we had been writing nothing but isolated unit tests, our functional tests would have picked up this particular slip-up. But ideally we’d want our feedback cycle to be quicker—functional tests may take a couple of minutes to run, or even a few hours once your app starts to grow. Is there any way to avoid this sort of problem before it happens?
Methodologically, the way to do it is to think about the interaction between your layers in terms of contracts. Whenever we mock out the behaviour of one layer, we have to make a mental note that there is now an implicit contract between the layers, and that a mock on one layer should probably translate into a test at the layer below.
Here’s the part of the contract that we missed:
@patch('lists.views.redirect')
def test_redirects_to_form_returned_object_if_form_valid(
self, mock_redirect, mockNewListForm
):
mock_form = mockNewListForm.return_value
mock_form.is_valid.return_value = True
response = new_list2(self.request)
self.assertEqual(response, mock_redirect.return_value)
mock_redirect.assert_called_once_with(mock_form.save.return_value) #(1)-
The mocked
form.savefunction is returning an object, which we expect our view to be able to use.
It’s worth reviewing each of the tests in NewListViewUnitTest and seeing
what each mock is saying about the implicit contract:
def test_passes_POST_data_to_NewListForm(self, mockNewListForm):
[...]
mockNewListForm.assert_called_once_with(data=self.request.POST) #(1)
def test_saves_form_with_owner_if_form_valid(self, mockNewListForm):
mock_form = mockNewListForm.return_value
mock_form.is_valid.return_value = True #(2)
new_list2(self.request)
mock_form.save.assert_called_once_with(owner=self.request.user) #(3)
def test_does_not_save_if_form_invalid(self, mockNewListForm):
[...]
mock_form.is_valid.return_value = False #(2)
[...]
@patch('lists.views.redirect')
def test_redirects_to_form_returned_object_if_form_valid(
self, mock_redirect, mockNewListForm
):
[...]
mock_redirect.assert_called_once_with(mock_form.save.return_value) #(4)
@patch('lists.views.render')
def test_renders_home_template_with_form_if_form_invalid(
[...]-
We need to be able to initialise our form by passing it a POST request as data.
-
It should have an
is_valid()function which returns True or False appropriately, based on the input data. -
The form should have a
.savemethod which will accept arequest.user, which may or may not be a logged-in user, and deal with it appropriately. -
The form’s
.savemethod should return a new list object, for our view to redirect the user to.
If we have a look through our form tests, we’ll see that, actually, only item (3)
is tested explicitly. On items (1) and (2) we were lucky—they’re default
features of a Django ModelForm, and they are actually covered by our
tests for the parent ItemForm class.
But contract clause number (4) managed to slip through the net.
|
Note
|
When doing Outside-In TDD with isolated tests, you need to keep track of
each test’s implicit assumptions about the contract which the next layer
should implement, and remember to test each of those in turn later. You
could use our scratchpad for this, or create a placeholder test with
a self.fail.
|
Let’s add a new test that our form should return the new saved list:
@patch('lists.forms.List.create_new')
def test_save_returns_new_list_object(self, mock_List_create_new):
user = Mock(is_authenticated=True)
form = NewListForm(data={'text': 'new item text'})
form.is_valid()
response = form.save(owner=user)
self.assertEqual(response, mock_List_create_new.return_value)And, actually, this is a good example—we have an implicit contract
with the List.create_new; we want it to return the new list object.
Let’s add a placeholder test for that:
class ListModelTest(TestCase):
[...]
def test_create_returns_new_list_object(self):
self.fail()So, we have one test failure that’s telling us to fix the form save:
AssertionError: None != <MagicMock name='create_new()' id='139802647565536'> FAILED (failures=2, errors=3)
Like this:
class NewListForm(ItemForm):
def save(self, owner):
if owner.is_authenticated:
return List.create_new(first_item_text=self.cleaned_data['text'], owner=owner)
else:
return List.create_new(first_item_text=self.cleaned_data['text'])That’s a start; now we should look at our placeholder test:
[...]
FAIL: test_create_returns_new_list_object
self.fail()
AssertionError: None
FAILED (failures=1, errors=3)
We flesh it out:
def test_create_returns_new_list_object(self):
returned = List.create_new(first_item_text='new item text')
new_list = List.objects.first()
self.assertEqual(returned, new_list)…
AssertionError: None != <List: List object>
And we add our return value:
@staticmethod
def create_new(first_item_text, owner=None):
list_ = List.objects.create(owner=owner)
Item.objects.create(text=first_item_text, list=list_)
return list_And that gets us to a fully passing test suite:
$ python manage.py test lists [...] Ran 50 tests in 0.169s OK
That’s our code for saving list owners, test-driven all the way down and working. But our functional test isn’t passing quite yet:
$ python manage.py test functional_tests.test_my_lists selenium.common.exceptions.NoSuchElementException: Message: Unable to locate element: Reticulate splines
It’s because we have one last feature to implement, the .name attribute on list
objects. Again, we can grab the test and code from the last chapter:
def test_list_name_is_first_item_text(self):
list_ = List.objects.create()
Item.objects.create(list=list_, text='first item')
Item.objects.create(list=list_, text='second item')
self.assertEqual(list_.name, 'first item')(Again, since this is a model-layer test, it’s OK to use the ORM. You could conceivably write this test using mocks, but there wouldn’t be much point.)
@property
def name(self):
return self.item_set.first().textAnd that gets us to a passing FT!
$ python manage.py test functional_tests.test_my_lists Ran 1 test in 21.428s OK
Now everything is working, we can remove some redundant tests, and decide whether we want to keep any of our old integrated tests.
We can get rid of the test for the old save method on the ItemForm:
--- a/lists/tests/test_forms.py
+++ b/lists/tests/test_forms.py
@@ -23,14 +23,6 @@ class ItemFormTest(TestCase):
self.assertEqual(form.errors['text'], [EMPTY_ITEM_ERROR])
- def test_form_save_handles_saving_to_a_list(self):
- list_ = List.objects.create()
- form = ItemForm(data={'text': 'do me'})
- new_item = form.save(for_list=list_)
- self.assertEqual(new_item, Item.objects.first())
- self.assertEqual(new_item.text, 'do me')
- self.assertEqual(new_item.list, list_)
-And in our actual code, we can get rid of two redundant save methods in 'forms.py':
--- a/lists/forms.py
+++ b/lists/forms.py
@@ -22,11 +22,6 @@ class ItemForm(forms.models.ModelForm):
self.fields['text'].error_messages['required'] = EMPTY_ITEM_ERROR
- def save(self, for_list):
- self.instance.list = for_list
- return super().save()
-
-
class NewListForm(ItemForm):
@@ -52,8 +47,3 @@ class ExistingListItemForm(ItemForm):
e.error_dict = {'text': [DUPLICATE_ITEM_ERROR]}
self._update_errors(e)
-
-
- def save(self):
- return forms.models.ModelForm.save(self)
-We can now completely remove the old new_list view, and rename new_list2 to
new_list:
-from lists.views import new_list, new_list2
+from lists.views import new_list
class HomePageTest(TestCase):
@@ -75,7 +75,7 @@ class NewListViewIntegratedTest(TestCase):
request = HttpRequest()
request.user = User.objects.create(email='[email protected]')
request.POST['text'] = 'new list item'
- new_list2(request)
+ new_list(request)
list_ = List.objects.first()
self.assertEqual(list_.owner, request.user)
@@ -91,21 +91,21 @@ class NewListViewUnitTest(unittest.TestCase):
def test_passes_POST_data_to_NewListForm(self, mockNewListForm):
- new_list2(self.request)
+ new_list(self.request)
[.. several more]--- a/lists/urls.py
+++ b/lists/urls.py
@@ -3,7 +3,7 @@ from django.conf.urls import url
from lists import views
urlpatterns = [
- url(r'^new$', views.new_list2, name='new_list'),
+ url(r'^new$', views.new_list, name='new_list'),
url(r'^(\d+)/$', views.view_list, name='view_list'),
url(r'^users/(.+)/$', views.my_lists, name='my_lists'),
]def new_list(request):
form = NewListForm(data=request.POST)
if form.is_valid():
list_ = form.save(owner=request.user)
[...]And a quick check that all the tests still pass:
OK
Finally, we have to decide what (if anything) to keep from our integrated test suite.
One option is to throw them all away, and decide that the FTs will pick up any integration problems. That’s perfectly valid.
On the other hand, we saw how integrated tests can warn you when you’ve made small mistakes in integrating your layers. We could keep just a couple of tests around as "sanity checks", to give us a quicker feedback cycle.
How about these three:
class NewListViewIntegratedTest(TestCase):
def test_can_save_a_POST_request(self):
self.client.post('/lists/new', data={'text': 'A new list item'})
self.assertEqual(Item.objects.count(), 1)
new_item = Item.objects.first()
self.assertEqual(new_item.text, 'A new list item')
def test_for_invalid_input_doesnt_save_but_shows_errors(self):
response = self.client.post('/lists/new', data={'text': ''})
self.assertEqual(List.objects.count(), 0)
self.assertContains(response, escape(EMPTY_ITEM_ERROR))
def test_list_owner_is_saved_if_user_is_authenticated(self):
user = User.objects.create(email='[email protected]')
self.client.force_login(user)
self.client.post('/lists/new', data={'text': 'new item'})
list_ = List.objects.first()
self.assertEqual(list_.owner, user)If you’re going to keep any intermediate-level tests at all, I like these three because they feel like they’re doing the most "integration" jobs: they test the full stack, from the request down to the actual database, and they cover the three most important use cases of our view.
Django’s
testing tools make it very easy to quickly put together integrated
tests. The test runner helpfully creates a fast, in-memory version of your
database and resets it for you in between each test. The TestCase class
and the test client make it easy to test your views, from checking whether
database objects are modified, confirming that your URL mappings work, and
inspecting the rendering of the templates. This lets you get started with
testing very easily and get good coverage across your whole stack.
On the other hand, these kinds of integrated tests won’t necessarily deliver the full benefit that rigorous unit testing and Outside-In TDD are meant to confer in terms of design.
If we look at the example in this chapter, compare the code we had before and after:
def new_list(request):
form = ItemForm(data=request.POST)
if form.is_valid():
list_ = List()
if not isinstance(request.user, AnonymousUser):
list_.owner = request.user
list_.save()
form.save(for_list=list_)
return redirect(list_)
else:
return render(request, 'home.html', {"form": form})def new_list(request):
form = NewListForm(data=request.POST)
if form.is_valid():
list_ = form.save(owner=request.user)
return redirect(list_)
return render(request, 'home.html', {'form': form})If we hadn’t bothered to go down the isolation route, would we have bothered to refactor the view function? I know I didn’t in the first draft of this book. I’d like to think I would have "in real life", but it’s hard to be sure. But writing isolated tests does make you very aware of where the complexities in your code lie.
I’d say the point at which isolated tests start to become worth it is to do with complexity. The example in this book is extremely simple, so it’s not usually been worth it so far. Even in the example in this chapter, I can convince myself I didn’t really 'need' to write those isolated tests.
But once an application gains a little more complexity—if it starts growing any more layers between views and models, if you find yourself writing helper methods, or if you’re writing your own classes, then you will probably gain from writing more isolated tests.
We already have our suite of functional tests, which will serve the purpose of telling us if we ever make any mistakes in integrating the different parts of our code together. Writing isolated tests can help us to drive out better design for our code, and to verify correctness in finer detail. Would a middle layer of integration tests serve any additional purpose?
I think the answer is potentially yes, if they can provide a faster feedback cycle, and help you identify more clearly what integration problems you suffer from—their tracebacks may provide you with better debug information than you would get from a functional test, for example.
There may even be a case for building them as a separate test suite—you
could have one suite of fast, isolated unit tests that don’t even use
manage.py, because they don’t need any of the database cleanup and teardown
that the Django test runner gives you, and then the intermediate layer that
uses Django, and finally the functional tests layer that, say, talks to a
staging server. It may be worth it if each layer delivers incremental
benefits.
It’s a judgement call. I hope that, by going through this chapter, I’ve given you a feel for what the trade-offs are. There’s more discussion on this in [chapter_hot_lava].
We’re happy with our new version, so let’s bring it across to master:
$ git add . $ git commit -m "add list owners via forms. more isolated tests" $ git checkout master $ git checkout -b master-noforms-noisolation-bak # optional backup $ git checkout master $ git reset --hard more-isolation # reset master to our branch.
In the meantime—those FTs are taking an annoyingly long time to run. I wonder if there’s something we can do about that?
and Decoupling ORM Code
- Functional tests
-
-
Provide the best guarantee that your application really works correctly, from the point of view of the user
-
But: it’s a slower feedback cycle
-
And they don’t necessarily help you write clean code
-
- Integrated tests (reliant on, for example, the ORM or the Django Test Client)
-
-
Are quick to write
-
Are easy to understand
-
Will warn you of any integration issues
-
But: may not always drive good design (that’s up to you!)
-
And are usually slower than isolated tests
-
- Isolated ("mocky") tests
-
-
Involve the most hard work
-
Can be harder to read and understand
-
But: are the best ones for guiding you towards better design
-
And run the fastest
-
- Decoupling our application from ORM code
-
One of the consequences of striving to write isolated tests is that we find ourselves forced to remove ORM code from places like views and forms, by hiding it behind helper functions or methods. This can be beneficial in terms of decoupling your application from the ORM, but also just because it makes your code more readable. As with all things, it’s a judgement call as to whether the additional effort is worth it in particular circumstances.