Tiny nitpick on something you haven't written: we should not test for an emtpy list using len(list) but just list.
An empty list evaluates as false in python.

Maybe you can replace:

if autolim and collection._paths and len(collection._offsets):

by

if autolim and collection._paths and collection._offsets:

As I wrote in a comment earlier, collection._offsets is a numpy array, not a list. This means that bool(np.array([0])) == False, and bool(np.array([0, 0])) raises an error.

My mistake.

I'd then use the shape method instead of the length attribute, but that's too much nitpicking :)

Thanks for clarifying.

Behavior of len is well-defined and documented for numpy arrays, so I still believe it's the better way due to improved readability.

I agree with @akhmerov on this point. I personally don't like the idiom of the empty list evaluating to false partly because of this reason. I work with numpy arrays so much. My second reason is that an iterator to an empty list evaluates to True, and so could cause a lot of confusion when coding in py3k with iterators being so prevelent there.

The more and more I look at this PR, the more I can't bring myself to accept it. I would rather fix the underlying problem.

This PR doesn't introduce yet another special condition, it rather corrects the condition which already existed before (I even suspect the original condition was a typo, since the second clause was pointless). In this sense it doesn't hurt anything. Even if path.get_path_collection_extents would be working properly, this condition could be reasonable to keep.

The problem that can now occur is if collection._offsets is None. Previously, the check on collection._paths would be sufficient to prevent an exception from being thrown when doing a len() on None. Now that protection is gone.

Line 1440 of collections.py in my fork of matplotlib uses the same assumption:

if len(self._offsets):
    xs = self.convert_xunits(self._offsets[:0])

Additionally, searching for all occurences of _offsets, I cannot confirm that they can ever assume a None value, instead _offsets are always an array.

@WeatherGod / @akhmerov - I've submitted a PR which moves this line to the Collection class in v1.3.1: #2444

This needs to have the @cleanup decorator to the figures get cleared after the test.

I think this test needs a comment re what the purpose of the test is.