Backend Django tests
Zulip uses the Django framework for its Python backend. We use the testing framework from django.test to test our code. We have over a thousand automated tests that verify that our backend works as expected.
All changes to the Zulip backend code should be supported by tests. We enforce our testing culture during code review, and we also use coverage tools to measure how well we test our code. We mostly use tests to prevent regressions in our code, but the tests can have ancillary benefits such as documenting interfaces and influencing the design of our software.
If you have worked on other Django projects that use unit testing, you will probably find familiar patterns in Zulip’s code. This document describes how to write tests for the Zulip backend, with a particular emphasis on areas where we have either wrapped Django’s test framework or just done things that are kind of unique in Zulip.
Our tests live in
zerver/tests/. You can run them with
./tools/test-backend. The tests run in parallel using multiple
threads in your development environment, and can finish in under 30s
on a fast machine. When you are in iterative mode, you can run
individual tests or individual modules, following the dotted.test.name
cd /srv/zulip ./tools/test-backend zerver.tests.test_queue_worker.WorkerTest
There are many command line options for running Zulip tests, such
--verbose option. The
best way to learn the options is to use the online help:
We also have ways to instrument our tests for finding code coverage,
URL coverage, and slow tests. Use the
-h option to discover these
features. We also have a
--profile option to facilitate profiling
test-backend will run all requested tests, and report
all failures at the end. You can configure it to stop after the first
error with the
--stop option (or
Another useful option is
--rerun, which will rerun just the tests
that failed in the last test run.
Webhook integrations. For performance,
test-backend with no
arguments will not run webhook integration tests (
which would otherwise account for about 25% of the total runtime.
When working on webhooks, we recommend instead running
test-backend zerver/webhooks manually (or better, the direction for the specific
webhooks you’re working on). And of course our CI is configured to
test-backend --include-webhooks and run all of the tests.
Before you write your first tests of Zulip, it is worthwhile to read the rest of this document.
To get a hang of commonly used testing techniques, read
You can also read some of the existing tests in
to get a feel for other patterns we use.
A good practice is to get a “failing test” before you start to implement your feature. First, it is a useful exercise to understand what needs to happen in your tests before you write the code, as it can help drive out simple design or help you make incremental progress on a large feature. Second, you want to avoid introducing tests that give false positives. Ensuring that a test fails before you implement the feature ensures that if somebody accidentally regresses the feature in the future, the test will catch the regression.
Setting up data for tests
All tests start with the same fixture data. (The tests themselves
update the database, but they do so inside a transaction that gets
rolled back after each of the tests complete. For more details on how the
fixture data gets set up, refer to
The fixture data includes a few users that are named after Shakesepeare characters, and they are part of the “zulip.com” realm.
Generally, you will also do some explicit data setup of your own. Here are a couple useful methods in ZulipTestCase:
More typically, you will use methods directly from the backend code. (This ensures more end-to-end testing, and avoids false positives from tests that might not consider ancillary parts of data setup that could influence tests results.)
Here are some example action methods that tests may use for data setup:
Testing code that accesses the filesystem
Some tests need to access the filesystem (e.g.
LocalUploadBackend and the data import tests). Doing
this correctly requires care to avoid problems like:
Leaking files after every test (which are clutter and can eventually run the development environment out of disk) or
Interacting with other parallel processes of this
test-backendrun (or another
test-backendrun), or with later tests run by this process.
To avoid these problems, you can do the following:
Use a subdirectory of
settings.TEST_WORKER_DIR; this is a subdirectory of
/var/<uuid>/test-backendthat is unique to the test worker thread and will be automatically deleted when the relevant
Delete any files created by the test in the test class’s
tearDownmethod (which runs even if the test fails); this is valuable to avoid conflicts with other tests run later by the same test process.
Our common testing infrastructure handles some of this for you,
e.g. it replaces
settings.LOCAL_UPLOADS_DIR for each test process
with a unique path under
UploadSerializeMixin manages some of the cleanup work for
Testing with mocks
This section is a beginner’s guide to mocking with Python’s
unittest.mock library. It will give you answers to the most common
questions around mocking, and a selection of commonly used mocking
What is mocking?
When writing tests, mocks allow you to replace methods or objects with fake entities suiting your testing requirements. Once an object is mocked, its original code does not get executed anymore.
Rather, you can think of a mocked object as an initially empty shell: Calling it won’t do anything, but you can fill your shell with custom code, return values, etc. Additionally, you can observe any calls made to your mocked object.
Why is mocking useful?
When writing tests, it often occurs that you make calls to functions taking complex arguments. Creating a real instance of such an argument would require the use of various different libraries, a lot of boilerplate code, etc. Another scenario is that the tested code accesses files or objects that don’t exist at testing time. Finally, it is good practice to keep tests independent from others. Mocks help you to isolate test cases by simulating objects and methods irrelevant to a test’s goal.
In all of these cases, you can “mock out” the function calls / objects and replace them with fake instances that only implement a limited interface. On top of that, these fake instances can be easily analyzed.
Say you have a module
greetings defining the following functions:
def fetch_database(key: str) -> str: # ... # Do some look-ups in a database return data def greet(name_key: str) -> str: name = fetch_database(name_key) return "Hello" + name
You want to test
In your test, you want to call
greet("Mario")and verify that it returns the correct greeting:
from greetings import greet def test_greet() -> str: greeting = greet("Mario") assert greeting == "Hello Mr. Mario Mario"
-> You have a problem:
fetch_database() does some look-ups in
a database. You haven’t created that database for your tests, so your test would fail, even though
the code is correct.
Luckily, you know that
fetch_database("Mario")should return “Mr. Mario Mario”.
Hint: Sometimes, you might not know the exact return value, but one that is equally valid and works with the rest of the code. In that case, just use this one.
-> Solution: You mock
fetch_database(). This is also referred to as “mocking out”
from unittest.mock import patch def test_greet() -> None: # Mock `fetch_database()` with an object that acts like a shell: It still accepts calls like `fetch_database()`, # but doesn't do any database lookup. We "fill" the shell with a return value; This value will be returned on every # call to `fetch_database()`. with patch("greetings.fetch_database", return_value="Mr. Mario Mario"): greeting = greetings.greet("Mario") assert greeting == "Hello Mr. Mario Mario"
That’s all. Note that this mock is suitable for testing
greet(), but not for testing
More generally, you should only mock those functions you explicitly don’t want to test.
How does mocking work under the hood?
Since Python 3.3, the standard mocking library is
unittest.mock implements the basic mocking class
It also implements
MagicMock, which is the same as
Mock, but contains many default magic methods (in Python,
those are the ones starting with with a dunder
__). From the docs:
In most of these examples the Mock and MagicMock classes are interchangeable. As the MagicMock is the more capable class it makes a sensible one to use by default.
Mock itself is a class that principally accepts and records any and all calls. A piece of code like
from unittest import mock foo = mock.Mock() foo.bar('quux') foo.baz foo.qux = 42
is not going to throw any errors. Our mock silently accepts all these calls and records them.
Mock also implements methods for us to access and assert its records, e.g.
unittest.mock also provides a method to mock objects only within a scope:
patch(). We can use
as a decorator or as a context manager. In both cases, the mock created by
patch() will apply for the scope of the decorator /
patch() takes only one required argument
target is a string in dot notation that refers to
the name of the object you want to mock. It will then assign a
MagicMock() to that object.
As an example, look at the following code:
from unittest import mock from os import urandom with mock.patch('__main__.urandom', return_value=42): print(urandom(1)) print(urandom(1)) # No matter what value we plug in for urandom, it will always return 42. print(urandom(1)) # We exited the context manager, so the mock doesn't apply anymore. Will return a random byte.
Note that calling
mock.patch('os.urandom', return_value=42) wouldn’t work here:
os.urandom would be the name of our patched
object. However, we imported
from os import urandom; hence, we bound the
urandom name to our current module
On the other hand, if we had used
import os.urandom, we would need to call
mock.patch('os.urandom', return_value=42) instead.
Including the Python mocking library:
from unittest import mock
Mocking a class with a context manager:
with mock.patch('module.ClassName', foo=42, return_value='I am a mock') as my_mock: # In here, 'module.ClassName' is mocked with a MagicMock() object my_mock. # my_mock has an attribute named foo with the value 42. # var = module.ClassName() will assign 'I am a mock' to var.
Mocking a class with a decorator:
@mock.patch('module.ClassName', foo=42, return_value='I am a mock') def my_function(my_mock): # ... # In here, 'module.ClassName' will behave as in the previous example.
Mocking a class attribute:
with mock.patch.object(module.ClassName, 'class_method', return_value=42) # In here, 'module.ClassName' has the same properties as before, except for 'class_method' # Calling module.ClassName.class_method() will now return 42.
Note the missing quotes around module.ClassName in the patch.object() call.
Zulip mocking practices
For mocking we generally use the “mock” library and use
a context manager or decorator. We also take advantage of some context managers
from Django as well as our own custom helpers. Here is an example:
with self.settings(RATE_LIMITING=True): with mock.patch('zerver.decorator.rate_limit_user') as rate_limit_mock: api_result = my_webhook(request) self.assertTrue(rate_limit_mock.called)
Follow this link for more information on the “settings” context manager.
Zulip has several features, like outgoing webhooks or social
authentication, that made outgoing HTTP requests to external
servers. We test those features using the excellent
responses library, which has a
nice interface for mocking
requests calls to return whatever HTTP
response from the exteranl server we need for the test. you can find
git grep responses.add. Zulip’s own
class should be used only for low-level tests for code that expects to
receive Django HttpRequest object.
Zulip testing philosophy
If there is one word to describe Zulip’s philosophy for writing tests, it is probably “flexible.” (Hopefully “thorough” goes without saying.)
When in doubt, unless speed concerns are prohibitive, you usually want your tests to be somewhat end-to-end, particularly for testing endpoints.
These are some of the testing strategies that you will see in the Zulip test suite…
We strive to test all of our URL endpoints. The vast majority of Zulip endpoints support a JSON interface. Regardless of the interface, an endpoint test generally follows this pattern:
Set up the data.
Log in with
self.login()or set up an API key.
Use a Zulip test helper to hit the endpoint.
Assert that the result was either a success or failure.
Check the data that comes back from the endpoint.
Generally, if you are doing endpoint tests, you will want to create a
test class that is a subclass of
ZulipTestCase, which will provide
you helper methods like the following:
For certain Zulip library functions, especially the ones that are not intrinsically tied to Django, we use a classic unit testing approach of calling the function and inspecting the results.
For these types of tests, you will often use methods like
self.assertTrue(), etc., which come with
Particularly for testing Zulip’s integrations with third party systems, we strive to have a highly data-driven approach to testing. To give a specific example, when we test our GitHub integration, the test code reads a bunch of sample inputs from a JSON fixture file, feeds them to our GitHub integration code, and then verifies the output against expected values from the same JSON fixture file.
Our fixtures live in
Mocks and stubs
We use mocks and stubs for all the typical reasons:
to more precisely test the target code
to stub out calls to third-party services
to make it so that you can run the Zulip tests on the airplane without wifi
A detailed description of mocks, along with useful coded snippets, can be found in the section Testing with mocks.
In zerver/tests/test_templates.py we have a test that renders all of our backend templates with a “dummy” context, to make sure the templates don’t have obvious errors. (These tests won’t catch all types of errors; they are just a first line of defense.)
SQL performance tests
A common class of bug with Django systems is to handle bulk data in
an inefficient way, where the backend populates objects for join tables
with a series of individual queries that give O(N) latency. (The
remedy is often just to call
select_related(), but sometimes it
requires a more subtle restructuring of the code.)
We try to prevent these bugs in our tests by using a context manager
queries_captured() that captures the SQL queries used by
the backend during a particular operation. We make assertions about
those queries, often simply asserting that the number of queries is
below some threshold.
The Zulip backend has a mechanism where it will fetch initial data
for a client from the database, and then it will subsequently apply
some queued up events to that data to the data structure before notifying
the client. The
BaseAction.do_test() helper helps tests
verify that the application of those events via apply_events() produces
the same data structure as performing an action that generates said event.
This is a bit esoteric, but if you read the tests, you will see some of the patterns. You can also learn more about our event system in the new feature tutorial.
It is important to verify error handling paths for endpoints, particularly
situations where we need to ensure that we don’t return results to clients
with improper authentication or with limited authorization. A typical test
will call the endpoint with either a non-logged in client, an invalid API
key, or missing input fields. Then the test will call
to verify that the endpoint is properly failing.
Here are some things to consider when writing new tests:
Duplication We try to avoid excessive duplication in tests. If you have several tests repeating the same type of test setup, consider making a setUp() method or a test helper.
Network independence Our tests should still work if you don’t have an internet connection. For third party clients, you can simulate their behavior using fixture data. For third party servers, you can typically simulate their behavior using mocks.
Coverage We have 100% line coverage on several of our backend modules. You can use the
--coverageoption to generate coverage reports, and new code should have 100% coverage, which generally requires testing not only the “happy path” but also error handling code and edge cases. It will generate a nice HTML report that you can view right from your browser (the tool prints the URL where the report is exposed in your development environment).
Console output A properly written test should print nothing to the console; use
with self.assertLogsto capture and verify any logging output. Note that we reconfigure various loggers in
zproject/test_extra_settings.pywhere the output is unlikely to be interesting when running our test suite.
test-backend --ban-console-outputchecks for stray print statements.
test-backend --coverage will assert that
various specific files in the project have 100% test coverage and
throw an error if their coverage has fallen. One of our project goals
is to expand that checking to ever-larger parts of the codebase.