collection of python libs developed for testing purposes
Alois Mahdal db594fe84d Convert to Python 3: `dict.iteritems()` -> `dict.items()` 2 лет назад
doc/examples Add basic README and a "live" example 10 лет назад
sznqalibs Convert to Python 3: `dict.iteritems()` -> `dict.items()` 2 лет назад
tests rename to sznqalibs - stupid dash, test imported 10 лет назад
.gitignore Initial commit 10 лет назад
LICENSE Initial commit 10 лет назад
README.md Add basic README and a "live" example 10 лет назад
setup.py PEP8/Pyflakes clean up 10 лет назад

README.md

sznqalibs

Collection of python libs developed for testing purposes.

hoover

hoover is a testing framework built with following principles in mind:

  • data-driven testing,
  • easy test data definition (even with huge sets),
  • helpful reporting.

Typical use case is that you have a tested system, another "reference" system and knowledge about testing input. You then create drivers for both systems that will parse and prepare output in a way that it can be compared to eaach other.

Examples

An example is worth 1000 words:

from sznqalibs import hoover


class BaloonDriver(hoover.TestDriver):
    """
    Object enclosing SUT or one of its typical use patterns
    """

    _get_data(self):
        # now do something to obtain results from the SUT
        # using self._argset dictionary
        self.data['sentence'] = subprocess.check_output(
            ['sut', self.args['count'], self.args['color']]
        )

class OracleDriver(hoover.TestDriver):
    """
    Object providing Oracle (expected output) for test arguments
    """

    _get_data(self):
        # obtain expected results, for example by asking
        # reference implementation (or by reimplementing
        # fraction of the SUT, e.g. only for the expected
        # data)
        self.data['sentence'] = ("%(count)s %(color)s baloons"
                                 % self._args)

class MyTest(unittest.TestCase):

    def test_valid(self):
        # as alternative to defining each _args separately,
        # Cartman lets you define just the ranges
        argsrc = hoover.Cartman({
            # for each parameter define iterator with
            # values you want to combine in this test
            'count': xrange(100),
            'color': ['red', 'blue']
        })
        # regression_test will call both drivers once with
        # each argument set, compare results and store results
        # along with some statistics
        tracker = hoover.regression_test(
            argsrc=argsrc,
            tests=[(operator.eq, OracleDriver, BaloonDriver)]
        )
        if tracker.errors_found():
            print tracker.format_report()

But that's just to get the idea. For a (hopefully) working example, look at doc/examples subfolder, there's a "calculator" implemented in Bash and Perl/CGI, and a hoover test that compares these two implementations to a Python implementation defined inside the test.

pFAQ (Potentially FAQ)

The truth is that nobody asked any questions so far, so I can't honestly write FAQ (or even AQ, for that matter) ;). So at least I'll try to answer what I feel like people would ask:

  • Q: What do you mean by implementing "reference", or "oracle" driver? Am I supposed to re-implement the system? Are you serious?

    A: Yes, I am serious. But consider this:

    First, not all systems are necessarily complicated. Take GNU cat. All it does is print data. Open, write, close. The added value is that it's insanely good at it. However, your oracle driver does not need to be that good. Even if it only was able to check the length or MD5 of data, it would be more than nothing.

    Also if you are creative enough, you can select the data in a clever way so that you can develop tricks that can help your driver on the way. For example you could "inject" the data with hints on how the result should be.

    Next, you don't need to actually re-implement the driver. As the most "brute" strategy, instead of using hoover to generate the data, you might want to just go and generate the data somehow manually (as you might have done it so far), verify it and feed it to your drivers. including expected results. This might not be the most viable option for a huge set, but at least what hoover will give you is the running and reporting engine.

    Then there might be cases when the system actually is trivial and you can re-implement it, but for some reason you don't have a testing framework on the native platform. For example, embedded system, or a library that needs to be in specific language like bash. In case it has trivial parts, you can test them in hoover and save yourself some headache with maintenance.

    Last, but not least--this was actually the story behind hoover being born--there are cases whan you already have reference implementation and new implementation, you just need to verify that behavior is the same. So you just wrap both systems in drivers, tweak them so that they can return the same data (if they already don't) or at least data you can write a comparison function for, squeeze them all into hoover.regression_test and hit the big button. Note that you can even have 1 reference driver and N SUT drivers, which can save you kajillions of machine seconds if your old library is slow or resource-hungry but you have more ports of the new one.

    As a bonus, note that hoover can also provide you with some performance stats. Well, there's absolutely no intent to say that this is a proper performance measurement tool (it's actually been designed to assess performance of the drivers), but on the other hand, it comes with the package, so it might be useful for you.

  • Q: Is it mature?

    A: No and a tiny yes.

    Yes, because it has already been used in real environment and it succeeded. But then again, it has been deployed by author, and he has no idea if that's actually doable for any sane person. You are more than welcome to try it and provide me with feedback and I can't provide any kind of guarranteees whatsoever.

    No, because there are parts that are still far from being polished, easy to use or even possible to understand. (Heck, at this moment even I don't understand what RuleOp is or was for :D). And there are probably limitations that could be removed.

    That said, the code is not a complete utter holy mess, though.

    But the API will change. Things will be re-designed and some even removed or split to other modules.

    My current "strategy", however, is to do this on the run, probably based on real experience when trying to use it in real testing scenarios.