A tale of two exceptions, continued

Thursday 23 February 2017

In my last blog post, A tale of two exceptions, I laid out the long drawn-out process of trying to get a certain exception to make tests skip in my test runner. I ended on a solution I liked at the time.

But it still meant having test-specific code in the product code, even if it was only a single line to set a base class for an exception. It didn’t feel right to say “SkipTest” in the product code, even once.

In that blog post, I said,

One of the reasons I write this stuff down is because I’m hoping to get feedback that will improve my solution, or advance my understanding. ... a reader might object and say, “you should blah blah blah.”

Sure enough, Ionel said,

A better way is to handle this in coverage’s test suite. Possible solution: wrap all your tests in a decorator that reraises with a SkipException.

I liked this idea. The need was definitely a testing need, so it should be handled in the tests. First I tried doing something with pytest to get it to do the conversion of exceptions for me. But I couldn’t find a way to make it work.

So: how to decorate all my tests? The decorator itself is fairly simple. Just call the method with all the arguments, and return its value, but if it raises StopEverything, then raise SkipTest instead:

def convert_skip_exceptions(method):
    """A decorator for test methods to convert StopEverything to SkipTest."""
    def wrapper(*args, **kwargs):
        """Run the test method, and convert exceptions."""
        try:
            result = method(*args, **kwargs)
        except StopEverything:
            raise unittest.SkipTest("StopEverything!")
        return result
    return wrapper

But decorating all the test methods would mean adding a @convert_skip_exceptions line to hundreds of test methods, which I clearly was not going to do. I could use a class decorator, which meant I would only have to add a decorator line to dozens of classes. That also felt like too much to do and remember to do in the future when I write new test classes.

It’s not often I say this, but: it was time for a metaclass. Metaclasses are one of the darkest magics Python has, and they can be mysterious. At heart, they are simple, but in a place you don’t normally think to look. Just as a class is used to make objects, a metaclass is used to make classes. Since there’s something I want to do everytime I make a new class (decorate its methods), a metaclass gives me the tools to do it.

class SkipConvertingMetaclass(type):
    """Decorate all test methods to convert StopEverything to SkipTest."""
    def __new__(mcs, name, bases, attrs):
        for attr_name, attr_value in attrs.items():
            right_name = attr_name.startswith('test_')
            right_type = isinstance(attr_value, types.FunctionType)
            if right_name and right_type:
                attrs[attr_name] = convert_skip_exceptions(attr_value)

        return super(SkipConvertingMetaclass, mcs).__new__(mcs, name, bases, attrs)

There are details here that you can skip as incantations if you like. Classes are all instances of “type”, so if we want to make a new thing that makes classes, it derives from type to get those same behaviors. The method that gets called when a new class is made is __new__. It gets passed the metaclass itself (just as classmethods get cls and instance methods get self), the name of the class, the tuple of base classes, and a dict of all the names and values defining the class (the methods, attributes, and so on).

The important part of this metaclass is what happens in the __new__ method. We look at all the attributes being defined on the class. If the name starts with “test_”, and it’s a function, then it’s a test method, and we decorate the value with our decorator. Remember that @-syntax is just a shorthand for passing the function through the decorator, which we do here the old-fashioned way.

Then we use super to let the usual class-defining mechanisms in “type” do their thing. Now all of our test methods are decorated, with no explicit @-lines in the code. There’s only one thing left to do: make sure all of our test classes use the metaclass:

CoverageTestMethodsMixin = SkipConvertingMetaclass('CoverageTestMethodsMixin', (), {})

class CoverageTest(
    ... some other mixins ...
    CoverageTestMethodsMixin,
    unittest.TestCase,
):
    """The base class for all coverage.py test classes."""

Metaclasses make classes, just the way classes make instances: you call them. Here we call our with the arguments it needs (class name, base classes, and attributes) to make a class called CoverageTestMethodsMixin.

Then we use CoverageTestMethodsMixin as one of the base classes of CoverageTest, which is the class used to derive all of the actual test classes.

Pro tip: if you are using unittest-style test classes, make a single class to be the base of all of your test classes, you will be glad.

After all of this class machinations, what have we got? Our test classes all derive from a base class which uses a metaclass to decorate all the test methods. As a result, any test which raises StopEverything will instead raise SkipTest to the test runner, and the test will be skipped. There’s now no mention of SkipTest in the product code at all. Better.

Comments

Add a comment:

Ignore this:
Leave this empty:
Name is required. Either email or web are required. Email won't be displayed and I won't spam you. Your web site won't be indexed by search engines.
Don't put anything here:
Leave this empty:
URLs auto-link and some tags are allowed: <a><b><i><p><br><pre>.