Main image of article Python Unit Testing: Best Practices to Follow

What are some of the best practices for Python unit testing? Before we break it down, let's distinguish unit testing from integration testing:

  • A unit test verifies a unit of code in isolation.
  • An integration test verifies several units of code in conjunction.

Crucially, unit tests should not make network requests, modify database tables, or alter files on disk unless there's an obvious reason for doing so. Unit tests use mocked or stubbed dependencies, optionally verifying that they are called correctly, and make assertions on the results of a single function or module. Integration tests, on the other hand, use concrete dependencies and make assertions on the results of a larger system.

You might disagree on the finer points when it comes to Python, but this distinction works well for me. I won't go over unit testing basics, like how to run pytest or how to use the @patch decorator, assuming you're familiar already.

Avoid Module-Level Globals

Imagine we have a class defined in validator.py:

    import os

    import fastjsonschema
    import requests

    schema_url = os.environ['SCHEMA_URL']

    class Validator:
        def __init__(self):
            schema = requests.get(schema_url).json()
            self.validator = fastjsonschema.compile(schema)

        def validate(self, event: dict) -> dict:
            return self.validator(event)

And we have a module named my_lambda.py that uses the class:

    validator = Validator()

    def handle_event(event: dict, context: object) -> dict:
        validator.validate(event)
        return {'body': 'Goodbye!', 'status': 200}

And we want to test the module:

    from unittest.mock import Mock, patch

    from my_lambda import handle_event

    @patch('my_lambda.Validator', autospec=True)
    def test_handle_event(mock_validator: Mock):
        result = handle_event({'message': 'Hello!'}, None)
        assert result == {'body': 'Goodbye!', 'status': 200}

We're patching the Validator class and not the validator property because the former will allow us to prevent the network request when Validator() is called, whereas the latter would only allow us to act afterwards.

As it's written, this test will fail during collection when my_lambda.py is imported, before the patch is even applied:

    test_my_lambda.py:3: in <module>
        from my_lambda import handle_event
    my_lambda.py:1: in <module>
        from validator import Validator
    validator.py:6: in <module>
        schema_url = os.environ['SCHEMA_URL']
    /usr/local/Cellar/python@3.9/3.9.12/Frameworks/Python.framework/Versions/3.9/lib/python3.9/os.py:679: in __getitem__
        raise KeyError(key) from None
    E   KeyError: 'SCHEMA_URL'

We could set `SCHEMA_URL` in our .env file, but that's an extra step that would be required for every developer to run the test. Instead, let's try setting it right before the import:

    import os

    from unittest.mock import Mock, patch

    os.environ['SCHEMA_URL'] = 'foo'

    from my_lambda import handle_event

This will also fail during collection:

    test_my_lambda.py:7: in <module>
        from my_lambda import handle_event
    my_lambda.py:3: in <module>
        validator = Validator()
    validator.py:10: in __init__
        schema = requests.get(schema_url).json()
    ...
    E   requests.exceptions.MissingSchema: Invalid URL 'foo': No scheme supplied. Perhaps you meant http://foo?

Even though we're patching the Validator class, it's still trying to make a network request because Validator() is called when my_lambda.py is imported, before the patch is even applied. If I sound like a broken record, it's because this can be confusing:

> Module-level globals are assigned when the module is imported, which makes them difficult to patch.

The correct way to fix this problem is to eliminate the global from my_lambda.py:

    from validator import Validator

    def handle_event(event: dict, context: object):
        validator = Validator()
        validator.validate(event)

Now the test passes, but it's kind of ugly. Let's remove another global, this time from validator.py:

    import os

    import fastjsonschema
    import requests

    class Validator:
        def __init__(self):
            schema_url = os.environ['SCHEMA_URL']
            schema = requests.get(schema_url).json()
            self.validator = fastjsonschema.compile(schema)

        def validate(self, event: dict) -> dict:
            return self.validator(event)

This allows us to simplify the test, restoring it to its original form:

    from unittest.mock import Mock, patch

    from my_lambda import handle_event

    @patch('my_lambda.Validator', autospec=True)
    def test_handle_event(mock_validator: Mock):
        result = handle_event({'message': 'Hello!'}, None)
        assert result == {'body': 'Goodbye!', 'status': 200}

Globals Aren't Always Bad

If you must use a global for some reason, set it to None initially:

    from typing import Optional

    from validator import Validator

    validator: Optional[Validator] = None

    def handle_event(event: dict, context: object):
        global validator

        if not validator:
            validator = Validator()

        validator.validate(event)

The key is to prevent any expensive computation or network requests when the module is imported to facilitate patching.

Use Dependency Injection

Next, let's test the class:

from validator import Validator

def test_validator_init():
    result = Validator()
    assert result.validator is not None

This test will fail unless SCHEMA_URL is defined in the environment. As previously mentioned, we could add it to our .env file, or we could set it directly like we did before with os.environ['SCHEMA_URL'] = 'foo'. It's better to use the @patch.dict decorator because it will restore the original value after the test exits:

    import os

    from unittest.mock import patch

    from validator import Validator

    @patch.dict(os.environ, {'SCHEMA_URL': 'foo'})
    def test_validator_init():
        result = Validator()
        assert result.validator is not None

Either way, this test will fail because it's trying to make a network request to an invalid URL:

    validator.py:9: in __init__
        schema = requests.get(schema_url).json()
    ...
    E   requests.exceptions.MissingSchema: Invalid URL 'foo': No scheme supplied. Perhaps you meant http://foo?

Again, we want to prevent the network request. This is where dependency injection, also known as inversion of control, comes in handy. Let's refactor validator.py to accept a few arguments:

    import os

    from typing import Callable, Optional

    import fastjsonschema
    import requests

    class Validator:
        def __init__(
            self,
            validator: Optional[Callable[[dict], dict]] = None,
            schema: Optional[dict] = None,
            schema_url: Optional[str] = None,
        ):
            if validator:
                self.validator = validator
                return

            if schema:
                self.validator = fastjsonschema.compile(schema)
                return

            if not schema_url:
                schema_url = os.environ['SCHEMA_URL']

            schema = requests.get(schema_url).json()
            self.validator = fastjsonschema.compile(schema)

        def validate(self, event: dict) -> dict:
            return self.validator(event)

It's worth a moment to understand what's happening here when Validator() is called:

  1. If called with a `validator` argument, that argument is assigned to `self.validator`. Any other arguments are ignored.
  2. If called with a `schema` argument, that argument is used to create `self.validator`.
  3. If called with a `schema_url` argument, the schema is requested from the URL, the response is parsed and used to create `self.validator`.
  4. If called without any arguments, the schema is requested from the URL specified in the environment.

This allows us to easily prevent the network request when running the unit test:

    import os

    from unittest.mock import Mock, patch

    from validator import Validator

    def test_validator_init_with_validator():
        mock_validator = Mock()

        result = Validator(validator=mock_validator)
        assert result.validator == mock_validator

The downside is we need to cover the additional complexity in the constructor. Fortunately, it can be accomplished with some straightforward patching:

    @patch('fastjsonschema.compile', autospec=True)
    def test_validator_init_with_schema(mock_compile):
        mock_validator = Mock()
        mock_compile.return_value = mock_validator

        result = Validator(schema={'type': 'string'})
        assert result.validator == mock_validator

    @patch('fastjsonschema.compile', autospec=True)
    @patch('requests.get', autospec=True)
    def test_validator_init_with_schema_url(mock_get, mock_compile):
        mock_schema = Mock()
        mock_get.return_value.json.return_value = mock_schema

        mock_validator = Mock()
        mock_compile.return_value = mock_validator

        result = Validator(schema_url='foo')
        assert result.validator == mock_validator

        mock_get.assert_called_once_with('foo')

    @patch('fastjsonschema.compile', autospec=True)
    @patch('requests.get', autospec=True)
    @patch.dict(os.environ, {'SCHEMA_URL': 'bar'})
    def test_validator_init_with_no_args(mock_get, mock_compile):
        mock_schema = Mock()
        mock_get.return_value.json.return_value = mock_schema

        mock_validator = Mock()
        mock_compile.return_value = mock_validator

        result = Validator()
        assert result.validator == mock_validator

        mock_get.assert_called_once_with('bar')

    def test_validator_validate():
        mock_validator = Mock()
        mock_validator.return_value = 42

        validator = Validator(mock_validator)

        result = validator.validate({})
        assert result == 42

This last unit test is dirt simple because we don't really care how fastjsonschema is implemented. It has its own unit tests.