/nose2pytest

Scripts to convert Python Nose tests to PyTest

Primary LanguagePythonOtherNOASSERTION

Contents

Overview

This package provides a Python script and pytest plugin to help convert Nose-based tests into pytest-based tests. Specifically, the script transforms nose.tools.assert_* function calls into raw assert statements, while preserving the format of original arguments as much as possible. For example, the script:

assert_true(a, msg)
assert_greater(a, b, msg)

gets converted to:

assert a, msg
assert a > b, msg

A small subset of nose.tools.assert_* function calls are not transformed because there is no raw assert statement equivalent, or the equivalent would be hard to maintain. They are provided as functions in the pytest namespace via pytest's plugin system.

Running

For a one-time conversion use the shell command

pipx run --python 3.11 nose2pytest path/to/dir/with/python_files

This will find all .py files in the folder tree starting at path/to/dir/with/python_files and overwrite the original (assuming most users will be running this on a version-controlled code base, this is almost always what would be most convenient). Type nose2pytest -h for other options, such as -v.

Installation

For doing multiple conversions use the shell command

pipx install --python 3.11 nose2pytest

For each conversion use the shell command

nose2pytest path/to/dir/with/python_files

Motivation

I have used Nose for years and it is a great tool. However, to get good test failure diagnostics with Nose you ought to use the assert_*() functions from nose.tools. Although they provide very good diagnostics, they are not as convenient to use as raw assertions, since you have to decide beforehand what type of assertion you are going to write: an identity comparison to None, a truth check, a falseness check, an identity comparison to another object, etc. Just being able to write a raw assertion, and still get good diagnostics on failure as done by pytest, is really nice. This is a main reason for using pytest for me. Another reason is the design of fixtures in pytest.

Switching an existing test suite from Nose to pytest is feasible even without nose2pytest, as it requires relatively little work: relatively as in, you will probably only need a few modifications, all achievable manually, to get the same test coverage and results. A few gotchas:

  • test classes that have __init__ will be ignored, those will have to be moved (usually, into class's setup_class())
  • the setup.cfg may have to be edited since test discovery rules are slightly more strict with pytest
  • the order of tests may be different, but in general, that should not matter
  • all test modules are imported up-front, so some test modules may need adjustment such as moving some code from the top of the test module into its setup_module()

Once the above has been done to an existing code base, you don't really have to do anything else. However, your test suite now has an additional third-party test dependency (Nose), just because of those assert_* functions used all over the place. Moreover, there is no longer one obvious way to do things in your test suite: existing test code uses nose.tools.assert_* functions, yet with pytest you can use raw assertions. If you add tests, which of these two approaches should a developer use? If you modify existing tests, should new assertions use raw assert? Should the remaining test method, test class, or test module be updated? A test module can contain hundreds of calls to nose.tools.assert_* functions, is a developer to manually go through each one to convert it? Painful and error-prone, in general not feasible to do manually.

This is why I developed nose2pytest: I wanted to migrate my pypubsub project's test suite from Nose to pytest, but also have only pytest as a dependency, and have one obvious way to write assertions in the test suite.

Requirements

I expect nose2pytest script to run with supported versions of CPython <= v3.11, on any OS supported by a version of Python that has lib2to3 compatible with fissix. I expect it to succeed even with quite old versions of Nose (even prior to 1.0 which came out ca. 2010) and with the new Nose2 test driver.

The pytest package namespace will be extended with assert_ functions that are not converted by the script only if, err, you have pytest installed!

Status

The package has been used on over 5000 assert_*() function calls, among which the pypubsub test suite. I consider it stable, but I have only used it on my code, and code by a few other developers. Feedback on results of conversions would be most appreciated (such as version information and number of assert statements converted).

The following conversions have been implemented:

Function Statement
assert_true(a[, msg]) assert a[, msg]
assert_false(a[, msg]) assert not a[, msg]
assert_is_none(a[, msg]) assert a is None[, msg]
assert_is_not_none(a[, msg]) assert a is not None[, msg]
assert_equal(a,b[, msg]) assert a == b[, msg]
assert_not_equal(a,b[, msg]) assert a != b[, msg]
assert_list_equal(a,b[, msg]) assert a == b[, msg]
assert_dict_equal(a,b[, msg]) assert a == b[, msg]
assert_set_equal(a,b[, msg]) assert a == b[, msg]
assert_sequence_equal(a,b[, msg]) assert a == b[, msg]
assert_tuple_equal(a,b[, msg]) assert a == b[, msg]
assert_multi_line_equal(a,b[, msg]) assert a == b[, msg]
assert_greater(a,b[, msg]) assert a > b[, msg]
assert_greater_equal(a,b[, msg]) assert a >= b[, msg]
assert_less(a,b[, msg]) assert a < b[, msg]
assert_less_equal(a,b[, msg]) assert a <= b[, msg]
assert_in(a,b[, msg]) assert a in b[, msg]
assert_not_in(a,b[, msg]) assert a not in b[, msg]
assert_is(a,b[, msg]) assert a is b[, msg]
assert_is_not(a,b[, msg]) assert a is not b[, msg]
assert_is_instance(a,b[, msg]) assert isinstance(a, b)[, msg]
assert_count_equal(a,b[, msg]) assert collections.Counter(a) == collections.Counter(b)[, msg]
assert_not_regex(a,b[, msg]) assert not re.search(b, a)[, msg]
assert_regex(a,b[, msg]) assert re.search(b, a)[, msg]
assert_almost_equal(a,b[, msg]) assert a == pytest.approx(b, abs=1e-7)[, msg]
assert_almost_equal(a,b, delta[, msg]) assert a == pytest.approx(b, abs=delta)[, msg]
assert_almost_equal(a, b, places[, msg]) assert a == pytest.approx(b, abs=1e-places)[, msg]
assert_not_almost_equal(a,b[, msg]) assert a != pytest.approx(b, abs=1e-7)[, msg]
assert_not_almost_equal(a,b, delta[, msg]) assert a != pytest.approx(b, abs=delta)[, msg]
assert_not_almost_equal(a,b, places[, msg]) assert a != pytest.approx(b, abs=1e-places)[, msg]

The script adds parentheses around a and/or b if operator precedence would change the interpretation of the expression or involves newline. For example:

assert_true(some-long-expression-a in
            some-long-expression-b, msg)
assert_equal(a == b, b == c), msg

gets converted to:

assert (some-long-expression-a in
            some-long-expression-b), msg
assert (a == b) == (b == c), msg

Not every assert_* function from nose.tools is converted by nose2pytest:

  1. Some Nose functions can be handled via a global search-replace, so a fixer was not a necessity:

    • assert_raises: replace with pytest.raises
    • assert_warns: replace with pytest.warns

  2. Some Nose functions could be transformed but the readability would be decreased:

    • assert_dict_contains_subset(a,b) -> assert set(b.keys()) >= a.keys() and {k: b[k] for k in a if k in b} == a

    The nose2pytest distribution contains a module, assert_tools.py which defines these utility functions to contain the equivalent raw assert statement. Copy the module into your test folder or into the pytest package and change your test code's from nose.tools import ... statements accordingly. pytest introspection will provide error information on assertion failure.

  3. Some Nose functions don't have a one-line assert statement equivalent, they have to remain utility functions:

    • assert_raises_regex
    • assert_raises_regexp # deprecated by Nose
    • assert_regexp_matches # deprecated by Nose
    • assert_warns_regex

    These functions are available in assert_tools.py of nose2pytest distribution, and are imported as is from unittest.TestCase (but renamed as per Nose). Copy the module into your test folder or into the pytest package and change your test code's from nose.tools import ... statements accordingly.

  4. Some Nose functions simply weren't on my radar; for example I just noticed for the first time that there is a nose.tools.ok_() function which is the same as assert_equal. Feel free to contribute via email or pull requests.

Limitations

  • The script does not convert nose.tools.assert_ import statements as there are too many possibilities. Should from nose.tools import ... be changed to from pytest import ..., and the implemented conversions be removed? Should an import pytest statement be added, and if so, where? If it is added after the line that had the nose.tools import, is the previous line really needed? Indeed the assert_ functions added in the pytest namespace could be accessed via pytest.assert_, in which case the script should prepend pytest. and remove the from nose.tools import ... entirely. Too many options, and you can fairly easily handle this via a global regexp search/replace.

  • Similarly, statements of the form nose.tools.assert_ are not converted: this would require some form of semantic analysis of each call to a function, because any of the following are possible:

    import nose.tools as nt

nt.assert_true(...)

nt2 = nt
nt2.assert_true(...)
nt2.assert_true(...)

import bogo.assert_true
bogo.assert_true(...)  # should this one be converted?

    The possibilities are endless so supporting this would require such a large amount of time that I do not have. As with other limitations in this section

  • Nose functions that can be used as context managers can obviously not be converted to raw assertions. However, there is currently no way of preventing nose2pytest from converting Nose functions used this way. You will have to manually fix.

  • @raises: this decorator can be replaced via the regular expression @raises\((.*)\) to @pytest.mark.xfail(raises=$1), but I prefer instead to convert such decorated test functions to use pytest.raises in the test function body. Indeed, it is easy to forget the decorator and add code after the line that raises, but this code will never be run and you won't know. Using the pytest.raises(...) is better than xfail(raise=...).

  • Nose2pytest does not have a means of determining if an assertion function is inside a lambda expression, so the valid lambda: assert_func(a, b) gets converted to the invalid lambda: assert a operator b. These should be rare, are easy to spot (your IDE will flag the syntax error, or you will get an exception on import), and are easy to fix by changing from a lambda expression to a local function.

I have no doubt that more limitations will arise as nose2pytest gets used on more code bases. Contributions to address these and existing limitations are most welcome.

Other tools

If your test suite is unittest- or unittest2-based, or your Nose tests also use some unittest/2 functionatlity (such as setUp(self) method in test classes), then you might find the following useful:

I have used neither, so I can't make recommendations. However, if your Nose-based test suite uses both Nose/2 and unittest/2 functionality (such as unittest.case.TestCase and/or setUp(self)/tearDown(self) methods), you should be able to run both a unittest2pytest converter, then the nose2pytest converter.

Solution Notes

I don't think this script would have been possible without lib2to3/fissix, certainly not with the same functionality since lib2to3/fissix, due to their purpose, preserves newlines, spaces and comments. The documentation for lib2to3/fissix is very minimal, so I was lucky to find http://python3porting.com/fixers.html.

Other than figuring out lib2to3/fissix package so I could harness its capabilities, some aspects of code transformations still turned out to be tricky, as warned by Regobro in the last paragraph of his Extending 2to3 page.

  • Multi-line arguments: Python accepts multi-line expressions when they are surrounded by parentheses, brackets or braces, but not otherwise. For example, converting:

    assert_func(long_a +
             long_b, msg)

    to:

    assert long_a +
           long_b, msg

    yields invalid Python code. However, converting to the following yields valid Python code:

    assert (long_a +
           long_b), msg

    So nose2pytest checks each argument expression (such as long_a +\n long_b) to see if it has newlines that would cause an invalid syntax, and if so, wraps them in parentheses. However, it is also important for the readability of raw assertions that parentheses only be present if necessary. In other words:

    assert_func((long_a +
             long_b), msg)
assert_func(z + (long_a +
                 long_b), msg)

    should convert to:

    assert (long_a +
           long_b), msg
assert z + (long_a +
                 long_b), msg)

    rather than:

    assert ((long_a +
           long_b)), msg
assert (z + (long_a +
                 long_b)), msg)

    So nose2pytest only tries to limit the addition of external parentheses to code that really needs it.

  • Operator precedence: Python assigns precedence to each operator; operators that are on the same level of precedence (like the comparison operators ==, >=, !=, etc) are executed in sequence. This poses a problem for two-argument assertion functions. Example: translating assert_equal(a != b, a <= c) to assert a != b == a <= c is incorrect, it must be converted to assert (a != b) == (a <= c). However, wrapping every argument in parentheses all the time does not produce easy-to-read assertions: assert_equal(a, b < c) should convert to assert a == (b < c), not assert (a) == (b < c).

    So nose2pytest adds parentheses around its arguments if the operator used between the args has lower precedence than any operator found in the arg. So assert_equal(a, b + c) converts to assert a == b + c whereas assert_equal(a, b in c) converts to assert a == (b in c) but assert_in(a == b, c) converts to assert a == b in c).

Contributing

Patches and extensions are welcome. Please fork, branch, and then submit PR. Nose2pytest uses lib2to3.pytree, in particular the Leaf and Node classes. There are a few particularly challenging aspects to transforming nose test expressions to equivalent pytest expressions:

  1. Finding expressions that match a pattern: If the code you want to transform does not already match one of the uses cases in script.py, you will have to determine the lib2to3/fissix pattern expression that describes it (this is similar to regular expressions, but for AST representation of code, instead of text strings). Various expression patterns already exist near the top of nose2pytest/script.py. This is largely trial and error as there is (as of this writing) no good documentation.
  2. Inserting the sub-expressions extracted by lib2to3/fissix in step 1 into the target "expression template". For example to convert assert_none(a) to assert a is None, the a sub-expression extracted via the lib2to3/fissix pattern must be inserted into the correct "placeholder" node of the target expression. If step 1 was necessary, then step 2 like involves creating a new class that derives from FixAssertBase.
  3. Parentheses and priority of operators: sometimes, it is necessary to add parentheses around an extracted subexpression to protect it against higher-priority operators. For example, in assert_none(a) the a could be an arbitrary Python expression, such as var1 and var2. The meaning of assert_none(var1 and var2) is not the same as assert var1 and var2 is None; parentheses must be added i.e. the target expression must be assert (var1 and var2) is None. Whether this is necessary depends on the transformation. The wrap_parens_* functions provide examples of how and when to do this.
  4. Spacing: white space and newlines in code must be preserved as much as possible, and removed when unnecessary. For example, assert_equal(a, b) convers to assert a == b; the latter already has a a space before the b, but so does the original; the lib2to3.pytree captures such 'non-code' information so that generating Python code from a Node yields the same as the input if no transformations were applied. This is done via the Node.prefix property.

When the pattern is correctly defined in step 1, adding a test in tests/test_script.py for a string that contains Python code that matches it will cause the FixAssertBase.transform(node, results) to be called, with node being the Node for which the children match the defined pattern. The results is map of object names defined in the pattern, to the Node subtree representing the sub-expression matched. For example, a pattern for assert_none(a) (where a could be any sub-expression such as 1+2 or sqrt(5) or var1+var2) will cause results to contain the sub-expression that a represents. The objective of transform() is then to put the extracted results at the correct location into a new Node tree that represents the target (transformed) expression.

Nodes form a tree, each Node has a children property, containing 0 or more Node and/or Leaf. For example, if node represents assert a/2 == b, then the tree might be something like this:

node (Node)
    assert (Leaf)
    node (node)
        node (node)
            a (Leaf)
            / (Leaf)
            2 (Leaf)
        ==  (Leaf)
        b (Leaf)

Sometimes you may be able to guess what the tree is for a given expression, however most often it is best to use a debugger to run a test that attempts to transform your expression of interest (there are several examples of how to do this in tests/test_script.py), break at the beginning of the FixAssertBase.transform() method, and explore the node.children tree to find the subexpressions that you need to extract. In the above example, the assert leaf node is child at index 0 of node.children, whereas child 1 is another Node; the a leaf is child 0 of child 0 of child 1 of node.children, i.e. it is node.children[0].children[0].children[1]. Therefore the "path" from node to reach 'a' is (0, 0, 1).

The main challenge for this step of nose2test extension is then to find the paths to reach the desired "placeholder" objects in the target expression. For example if assert_almost_equal(a, b, delta=value) must be converted to assert a == pytest.approx(b, delta=value), then the nodes of interest are a, b, and delta, and their paths are 0, (2, 2, 1, 0) and (2, 2, 1, 2, 2) respectively (when a path contains only 1 item, there is no need to use a tuple).

Releasing

See RELEASING.rst.

Maintenance

  • Clone or fork the git repo, create a branch
  • Install pytest and nose on your system: python -m pip install pytest nose
  • In the root folder, run pytest
  • Once all tests pass, install tox on your system: on Ubuntu, python -m pip install tox
  • Run tox: tox
  • Add a python version if the latest Python is not in tox.ini

Note

Notes for Ubuntu:

My experience today installing python 3.5 to 3.11 on Ubuntu 18 was surprisingly not smooth. I had to use these commands:

  • sudo apt install python3.5 (ok)
  • sudo apt install python3.x-distutils for x=9,10,11
  • had to use python -m pip intead of just pip otherwise wrong version would get found
  • used sudo update-alternatives --install /usr/bin/python python /usr/bin/python3.x 1 for all x
  • used sudo update-alternatives --config python to choose which python active
  • had to install setuptools from git repo otherwise weird pip error (used https://stackoverflow.com/a/69573368/869951)
  • note however that once the correct tox installed,

Acknowledgments

Thanks to (AFAICT) Lennart Regebro for having written http://python3porting.com/fixers.html#find-pattern, and to those who answered my question on SO and my question on pytest-dev.