Load schemas via importlib.resources instead of pkgutil

[sevein]

Producing binaries that embed Python applications that import jsonschema is challenging because pkgutil.get_data relies on __file__ which may not be always defined, e.g. see these issues in PyOxidizer.

https://github.com/Julian/jsonschema/blob/272b4f24437d9cba1127058dc3479f9bc6d8c4b9/jsonschema/_utils.py#L49-L55

I was wondering if this project would be interested in having a fallback implemented that is capable to load the schema files using importlib.resources instead, which was added in Python 3.7. See https://stackoverflow.com/a/20885799 for more details. The creator of PyOxidizer, Gregory Szorc, shares some good details on loading resource files using PyOxidizer here.

[Julian]

Hi. importlib.resources is going through some instability now (having lots of APIs deprecated), which is something I haven't paid a ton of attention to.

In theory now that we're 3.7+ I'm indeed open to the idea of using it if we can do so without raising deprecation warnings on other versions, so a PR would be welcome --

I'll also just point out though that support for pyoxidizer (or pyinstaller, or cx_freeze, or etc.) is unofficial because jsonschema isn't run under it in CI, so whilst I'm OK with this proposal for independent reasons (that importlib.resources is the new way to do this), I can't really promise in the future that things will continue to work.

[sevein]

Thank you, @Julian. I woud love to submit a pull request!

If we want to use importlib.resources.read_text then I believe we need to convert schemas into a sub-package (adding an empty __init__.py file). I think this is one of the functions that may be eventually deprecated but available in the standard library since Python 3.7.

The alternative seems to be the much nicer importlib.resources.files, added in Python 3.9 but also in older releases of Python if we used importlib_resources.

Let me know what's your preference. Thanks again.

[Julian]

Cool!

I'd rather not use a function that's already been deprecated, unless it shakes out that it's going to be undeprecated which last I saw seemed unclear that that'd happen.

So I'd prefer the second route with the backport for pre-3.9, which I hear has its own perils but seems at least like the supported route by the upstream folks.