[unittest] raise error if @skip is used with an argument that looks like a test method

[Naitreey]

BPO	34596
Nosy	@rbtcollins, @ezio-melotti, @stevendaprano, @voidspace, @berkerpeksag, @zware, @serhiy-storchaka, @Naitreey
PRs	bpo-34596: Fallback to a default reason when @unittest.skip is uncalled #9082 [3.8] bpo-34596: Fallback to a default reason when @unittest.skip is uncalled (GH-9082) #15781

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = 'https://github.com/voidspace'
closed_at = <Date 2019-09-09.14:44:10.065>
created_at = <Date 2018-09-06.13:05:21.123>
labels = ['extension-modules', 'type-feature', '3.8']
title = '[unittest] raise error if @skip is used with an argument that looks like a test method'
updated_at = <Date 2019-09-09.15:01:17.873>
user = 'https://github.com/Naitreey'

bugs.python.org fields:

activity = <Date 2019-09-09.15:01:17.873>
actor = 'python-dev'
assignee = 'michael.foord'
closed = True
closed_date = <Date 2019-09-09.14:44:10.065>
closer = 'michael.foord'
components = ['Extension Modules']
creation = <Date 2018-09-06.13:05:21.123>
creator = 'Naitree Zhu'
dependencies = []
files = []
hgrepos = []
issue_num = 34596
keywords = ['patch']
message_count = 18.0
messages = ['324688', '324689', '324691', '324692', '324697', '324699', '324700', '324701', '324703', '324836', '324853', '324903', '325308', '325417', '351431', '351444', '351471', '351493']
nosy_count = 9.0
nosy_names = ['rbcollins', 'ezio.melotti', 'steven.daprano', 'michael.foord', 'python-dev', 'berker.peksag', 'zach.ware', 'serhiy.storchaka', 'Naitree Zhu']
pr_nums = ['9082', '15781']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'enhancement'
url = 'https://bugs.python.org/issue34596'
versions = ['Python 3.8']

[Naitreey]

When using @Skip decorator, reason argument is required. But one could easily overlook that and use it like so:

    class SomeTest(TestCase):

        @skip
        def test_method(self):
            # ...

The test actually passes when running with unittest, because it's equivalent to

    class SomeTest(TestCase):

        def test_method(self):
            # ...
        test_method = skip(test_method)

This gives the illusion that @Skip decorator worked as expected.

I propose we should check the passed in reason, and raise error if it looks like a function/method, therefore prevent this erroneous usage.

In practice, this behavior has misled people using the decorator in the unintended way. For example: In this chapter of Test-Driven Development with Python (http://www.obeythetestinggoat.com/book/chapter_organising_test_files.html), the author used @Skip decorator without reason argument.

[stevendaprano]

Is there a use-case for reason to be anything but a string?

[Naitreey]

Well, I personally can not think of any. I think reason should normally just be string.

If that is ok, I'll be happy to submit a PR that restricts reason to be a string.

[zware]

It could be interesting to enable uncalled skip by setting a default reason of "Unconditional skip" when the argument is a function.

Do note that decorating with an uncalled skip does actually work to skip the test currently, but the test is marked as success rather than skipped (see example pasted below). For this reason, I don't think we should turn it into an error in maintenance releases, as anyone using this accidentally will suddenly have many failing tests.

$ cat test.py
from unittest import TestCase, skip

class Test(TestCase):

    def test_good(self):
        self.assertTrue(1.0)

    def test_bad(self):
        self.assertFalse(1.0)

    @skip
    def test_bad_skip(self):
        self.assertFalse(1.0)

@skip('always skipped')
def test_good_skip(self):
    self.assertFalse(1.0)

$ ./python.exe -m unittest test.py -v
test_bad (test.Test) ... FAIL
test_bad_skip (test.Test) ... ok
test_good (test.Test) ... ok
test_good_skip (test.Test) ... skipped 'always skipped'

======================================================================
FAIL: test_bad (test.Test)
----------------------------------------------------------------------

Traceback (most recent call last):
  File "/.../test.py", line 10, in test_bad
    self.assertFalse(1.0)
AssertionError: 1.0 is not false

---

Ran 4 tests in 0.002s

FAILED (failures=1, skipped=1)

[Naitreey]

What would be a good default reason? How about the function name?

    if isinstance(reason, types.FunctionType):
        reason = reason.__name__

For example,

    from unittest import TestCase, skip

    class Test(TestCase):

        @skip
        def test_bad_skip(self):
            self.assertFalse(1.0)

    @skip('always skipped')
    def test_good_skip(self):
        self.assertFalse(1.0)

$ python -m unittest -v test.py
test_bad_skip (test.Test) ... skipped 'test_bad_skip'
test_good_skip (test.Test) ... skipped 'always skipped'

----------------------------------------------------------------------
Ran 2 tests in 0.000s

OK (skipped=2)

But this wouldn't be very helpful if test method is a lambda (just saying...)

[zware]

"Unconditionally" :)

[Naitreey]

Hi @zach.ware, Just to make sure I'm getting this right (first time contributing to cpython...)

Now I need to open 4 PRs at GitHub,

• 1 PR to master branch, with following changes: raise TypeError when reason is not a string. (Include unit test.)
• 3 PRs to 3.7, 3.6, and 2.7, with following changes: Setting reason to "Unconditionally" when reason is a function. (Include unit test.)

Looking good?

[zware]

I think we should make the same change on all branches (why would we fix it in maintenance branches just to break it in the next major release?), in which case it's just one PR to master and our backport bot (and/or the merging core dev) will take care of it from there.

[Naitreey]

Ok, I see. Thank you.

[serhiy-storchaka]

It doesn't look like a good design to me if allow a function be a decorator and a decorator fabric at the same time. It always lead to cumbersome and errorprone implementation. Currently there is only one example of such design in the stdlib, other propositions were rejected.

Checking that the argument is a string and raising exception otherwise looks good to me for 3.8. There is no a bug that need to be fixed in maintained versions. If you use unittest.skip() improperly, it is your failure. Helping to catch such mistakes is a new feature.

[zware]

I don't agree that this change makes the implementation significantly more cumbersome. I also think there's a backward compatibility argument to be made for allowing the uncalled usage, particularly considering the OP's published example and other similar ones, such as https://stackoverflow.com/questions/2066508/disable-individual-python-unit-tests-temporarily which was for me the first Google result for "python unittest skip" that was not our own docs.

[stevendaprano]

For what its worth, I'm +1 with Serhiy that we raise an exception
on a non-string argument, including the case where the caller forgets to
provide a reason at all.

I don't think that @Skip with no reason was ever intended to work (it
certainly wasn't documented as working) and as this only effects tests,
not production code, I don't think we ought to be constrained by "bug
compatibility" concerns. Fixing broken tests is easier than fixing
broken production code.

[berkerpeksag]

It would be nice to make *reason* optional. Every time I needed to use @unittest.skip() (even if I wanted to use it temporarily), I ended up passing some random string as reason or use 'raise SkipTest' directly.

If we decide to keep *reason* required, I agree with Serhiy's proposal.

[Naitreey]

Hi guys, what's our consensus on this?

• raise an exception as a fix? or
• fallback to default reason as a new feature?

If we choose to explicitly make reason optional (I mean by documenting it as such), shouldn't we also change @skipIf and @skipUnless to keep consistency?

[zware]

I think we need Michael's ruling on which way to go here. I'm fine with either an exception or a default reason, but still lean towards the default. I don't think skipIf or skipUnless really need a change either way; they both require two arguments and thus can't be misused in the same way as skip.

[voidspace]

I'm in favour of a default and "Unconditionally skipped" is fine with me. Although "Skipped" would also be fine.

Making @Skip work with no arguments is fine. Having to pass in arguments message arguments you don't want is a pain and there's no need to force the user.

Let's add a default message and make the skipped tests report as skipped properly on 3.8 and close it.

[python-dev]

New changeset d5fd75c by Michael Foord (Naitree Zhu) in branch 'master':
bpo-34596: Fallback to a default reason when @unittest.skip is uncalled (bpo-9082)
d5fd75c

[python-dev]

New changeset 3bd4bed by Michael Foord (Miss Islington (bot)) in branch '3.8':
bpo-34596: Fallback to a default reason when @unittest.skip is uncalled (GH-9082) (bpo-15781)
3bd4bed