It’s just a fact of life, as code grows eventually you will need to start adding mocks to your test suite. What started as a cute little two class project is now talking to external services and you cannot test it comfortably anymore.
That’s why Python ships with
unittest.mock, a powerful part of the standard library for stubbing dependencies and mocking side effects.
unittest.mock is not particularly intuitive.
I’ve found myself many times wondering why my go-to recipe does not work for a particular case, so I’ve put together this cheatsheet to help myself and others get mocks working quickly.
You can find the code examples in the article’s Github repository. I’ll be using Python 3.6, if you’re using 3.2 or below you’ll need to use the
mock PyPI package.
The examples are written using
unittest.TestCase classes for simplicity in executing them without dependencies, but you could write them as functions using
pytest almost directly,
unittest.mock will work just fine. If you are a
pytest user though I encourage you to have a look at the excellent
The Mock class in a nutshell
The centerpoint of the
unittest.mock module is, of course, the
Mock class. The main characteristic of a
Mock object is that it will return another
Mock instance when:
- accessing one of its attributes
- calling the object itself
This is the default behaviour, but it can be overridden in different ways. For example you can assign a value to an attribute in the
- Assign it directly, like you’d do with any Python object.
- Use the
configure_mockmethod on an instance.
- Or pass keyword arguments to the
Mockclass on creation.
To override calls to the mock you’ll need to configure its
return_value property, also available as a keyword argument in the
Mock initializer. The
Mock will always return the same value on all calls, this, again, can also be configured by using the
- if you’d like to return different values on each call you can assign an iterable to
- If you’d like to raise an exception when calling the Mock you can simply assign the exception object to
With all these tools we can now create stubs for essentially any Python object, which will work great for inputs to our system. But what about outputs?
If you’re making a CLI program to download the whole Internet, you probably don’t want to download the whole Internet on each test. Instead it would be enough to assert that
requests.download_internet (not a real method) was called appropriately.
Mock gives you handy methods to do so.
Note in our example
assert_called_once failed, this showcases another key aspect of
Mock objects, they record all interactions with them and you can then inspect these interactions.
For example you can use
call_count to retrieve the number of calls to the
Mock, and use
call_args_list to inspect the arguments to the last or all calls respectively.
If this is inconvenient at any point you can use the
reset_mock method to clear the recorded interactions, note the configuration will not be reset, just the interactions.
Finally, let me introduce
MagicMock, a subclass of
Mock that implements default magic or dunder methods. This makes
MagicMock ideal to mock class behaviour, which is why it’s the default class when patching.
We are now ready to start mocking and isolating the unit under tests. Here are a few recipes to keep in mind:
Patch on import
The main way to use
unittest.mock is to patch imports in the module under test using the
patch will intercept
import statements identified by a string (more on that later), and return a
Mock instance you can preconfigure using the techniques we discussed above.
Imagine we want to test this very simple function:
Note we’re importing
os and calling
getcwd to get the current working directory. We don’t want to actually call it on our tests though since it’s not important to our code and the return value might differ between environments it run on.
As mentioned above we need to supply
patch with a string representing our specific import. We do not want to supply simply
os.getcwd since that would patch it for all modules, instead we want to supply the module under test’s
os , i.e.
work.os . When the module is imported
patch will work its magic and return a
Alternatively, we can use the decorator version of
patch , note this time the test has an extra parameter:
mocked_os to which the
Mock is injected into the test.
patch will forward keyword arguments to the
Mock class, so to configure a
return_value we simply add it as one:
It’s quite common to patch classes completely or partially. Imagine the following simple module:
Now there’s two things we need to test:
Workerreturns the expected path supplied by
In order to test
Worker in complete isolation we need to patch the whole
Note the double
return_value in the example, simply using
MockHelper.get_path.return_value would not work since in the code we call
get_path on an instance, not the class itself.
The chaining syntax is slightly confusing but remember
MagicMock returns another
MagicMock on calls
__init__. Here we’re configuring any fake
Helper instances created by
MockHelper to return what we expect on calls to
get_path which is the only method we care about in our test.
A consequence of the flexibility of
Mock is that once we’ve mocked a class Python will not raise
AttributeError as it simply will return new instances of
MagicMock for basically everything. This is usually a good thing but can lead to some confusing behaviour and potentially bugs. For instance writing the following test,
will silently pass with no warning completely missing the typo in
Additionally, if we were to rename
Helper.get_folder, but forget to update the call in
Worker our tests will still pass:
Mock comes with a tool to prevent these errors, speccing.
Put simply, it preconfigures mocks to only respond to methods that actually exist in the spec class. There are several ways to define specs, but the easiest is to simply pass
autospec=True to the
patch call, which will configure the
Mock to behave as the object being mocked, raising exceptions for missing attributes and incorrect signatures as required. For example:
Partial class mocking
If you’re less inclined to testing in complete isolation you can also partially patch a class using
patch.object is allowing us to configure a mocked version of
get_path only, leaving the rest of the behaviour untouched. Of course this means the test is no longer what you would strictly consider a unit test but you may be ok with that.
Mocking built-in functions and environment variables
In the previous examples we neglected to test one particular aspect of our simple class, the
We can write a test like the following:
Note a few things:
- we can mock
patchcall to signal
unittest.mockto create a
Mockeven though no import matches the identifier.
- we’re using
patch.dictto inject a temporary environment variable in
os.environ, this is extensible to any other dictionary we’d like to mock.
- we’re nesting several
patchcontext manager calls but only using
asin the first one since it’s the one we need to use to call
If you’re not fond of nesting context managers you can also write the
patch calls in the decorator form:
Note however the order of the arguments to the test matches the stacking order of the decorators, and also that
patch.dict does not inject an argument.
Mocking context managers
Context managers are incredibly common and useful in Python, but their actual mechanics make them slightly awkward to mock, imagine this very common scenario:
You could, of course, add a actual fixture file, but in real world cases it might not be an option, instead we can mock the context manager’s output to be a
There’s nothing special here except the magic
__enter__ method, we just have to remember the underlying mechanics of context managers and some clever use of our trusty
Mocking class attributes
There are many ways in which to achieve this but some are more fool-proof that others. Say you’ve written the following code:
You could test the code without any mocks in two ways:
- If the code under test accesses the attribute via
self.ATTRIBUTE, which is the case in this example, you can simply set the attribute directly in the instance. This is fairly safe as the change is limited to this single instance.
- Alternatively you can also set the attribute in the imported class in the test before creating an instance. This however changes the class attribute you’ve imported in your test which could affect the following tests, so you’d need to remember to reset it.
The main drawback from this non-Mock approaches is that if you at any point rename the attribute your tests will fail and the error will not directly point out this naming mismatch.
To solve that we can make use of
patch.object on the imported class which will complain if the class does not have the specified attribute.
Here are the some tests using each technique:
Mocking class helpers
The following example is the root of many problems regarding monkeypatching using Mock. It usually shows up on more mature codebases that start making use of frameworks and helpers at class definition. For example, imagine this hypothetical
Field class helper:
Its purpose is to wrap and enhance an attribute in another class, a fairly pattern you commonly might see in ORMs or form libraries. Don’t concern yourself too much with the details of it just note that there is a
default value passed in.
Now take this other sample of production code:
This class makes use of the
Field class by defining its
country attribute as one whose type is
str and a
default as the first element of the
COUNTRIES constant. The logic under test is that depending on the country a discount is applied.
For which we might write the following test:
But that would NOT pass.
Let’s walk through the test:
- First it patches the default countries in
pricerto be a list with a single entry
- This should make the
CountryPricer.countryattribute default to that entry since it’s definition includes
- It then instanciates the
CountryPricerand asks for the discounted price for
So what’s going on?
The root cause of this lies in Python’s behaviour during import, best described in Luciano Ramalho’s excellent Fluent Python on chapter 21:
For classes, the story is different: at import time, the interpreter executes the body of every class, even the body of classes nested in other classes. Execution of a class body means that the attributes and methods of the class are defined, and then the class object itself is built.
Applying this to our example:
Fieldinstance is built before the test is ran at import time,
- Python reads the body of the class, passing the
COUNTRIESthat is defined at that point to the
- Our test code runs but it’s too late for us to patch
COUNTRIESand get a proper assertion.
From the above description you might try delaying the importing of the module until inside the tests, something like:
This, however, will still not pass as
mock.patch will import the module and then monkeypatch it, resulting in the same situation as before.
To work around this we need to embrace the state of the
CountryPricer class at the time of the test and patch
default on the already initialized
This solution is not ideal since it requires knowledge of the internals of
Field which you may not have or want to use if you’re using an external library but it works on this admittedly simplified case.
This import time issue is one of the main problems I’ve encountered while using
unittest.mock. You need to remember while using it that at import time code at the top-level of modules is executed, including class bodies.
If the logic you’re testing is dependent on any of this logic you may need to rethink how you are using
This introduction and cheatsheet should get you mocking with
unittest.mock, but I encourage you to read the documentation thoroughly, there’s plenty of more interesting tricks to learn.
It’s worth mentioning that there are alternatives to
unittest.mock, in particular Alex Gaynor’s
pretend library in combination with
monkeypatch fixture. As Alex points out, using these two libraries you can take a different approach to make mocking stricter yet more predictable. Definitely an approach worth exploring but outside the scope of this article.
unittest.mock is currently the standard for mocking in Python and you’ll find it in virtually every codebase. Having a solid understanding of it will help you write better tests quicker.
🎉 Thanks for reading! If you’ve enjoyed this article you may want to have a look at my series on Python’s new built-in concurrency library, AsyncIO:
- AsyncIO for the Working Python Developer,
- AsyncIO Coroutine Patterns: Beyond await
- and AsyncIO Coroutine Patterns: Errors and Cancellation.