Fixes #126: generator objects leak on Python 3

[mdboom]

The issue here is that generator objects that don't run to completion are uncollectable. There are a few places where iter_errors is called simply to retrieve the first error, and thus iter_errors is left uncompleted and holding references to a bunch of things (including self, which prevents deletion of the iter_errors generator instance).

The fix here isn't terribly elegant: It adds a kwarg to iter_errors to yield only the first error. Then, this is used from the places that only want the first error (along with list to fully exhaust the generator). The alternative (to just use list) would mean the possibly expensive creation of many error results only to have them immediately ignored and thrown away -- less efficient, but a simpler/less hackish solution perhaps.

Lastly, I couldn't figure out how to write a test for this: except perhaps to do:

warnings.filterwarnings("error", ".*", ResourceWarning)

which would raise an exception at the end of the test run rather than just a warning.

[Julian]

First of all, thanks a ton for taking a shot at this. Not the type of issue that contributors typically are willing to try and tackle, so :).

That being said, I'm skeptical that partially-run generators are uncollectable. References that are being held by the generator shouldn't prevent the generator from being collected -- but I haven't produced a minimal reproducing example one way or the other yet

Either way, I agree that the fix isn't really that pretty, if this were true, I'd be inclined to "fix" it by making the suite always exhaust its generators to tear down, not by changing actual library code (beside the fact that "private" function attributes are unfortunately a lie :( ).

I'm gonna take a closer look at this, but I also notice that the place where I was noticing this happen doesn't appear to be complaining anymore -- I don't see ResourceWarnings when running the test suite under 3.4. Not sure if anything has changed or not. How did you re-notice this?

[mdboom]

I noticed this as a result of running the test suite in my own project: spacetelescope/pyfinf (I see it both with jsonschema 2.3.0 and git master). It turns out, however, that I only see this on Python 3.3.2, not Python 3.4.0. I think my Python 3 userbase tends to upgrade pretty quickly -- it's Python 2 where legacy support is more important -- so I'm fine with this being marked as "won't fix" maybe with some sort of note about Python 3 prior to 3.4.

I think the relevant change is this:

http://legacy.python.org/dev/peps/pep-0442/

given that jsonschema does have some generators with finally clauses in them, which were reportedly uncollectable when part of cycles prior to Python 3.4.

[Julian]

Ah got it, I actually have lost access to 3.3, but that was my suspicion, thanks for finding the relevant PEP.

If it in fact is something fixed in 3.4 I'd be inclined to just say yeah we should leave it as-is, but I'll do a bit more research here before closing.

Again, very much appreciated.