Importing humanize is slow

[ashwin153]

I have noticed that importing humanize noticeably increases the startup time of my CLI. After doing some importtime analysis, I think the underlying problem is that importing pkg_resources is slow.

• Importing pkg_resources is slow. Work is done at import time. pypa/setuptools#926
• pkg_resources causes a slow import; any way to avoid? pydata/numexpr#291

Is is it possible to replace the usage of pkg_resources.get_distribution in __init__.py with something that isn't as expensive to import?

[hugovk]

An alternative is using importlib.metadata:

https://github.com/pypa/setuptools_scm/#retrieving-package-version-at-runtime

Would you like to test to see if that improves things?

It would be nice to see some numbers comparing both, and against setting a dummy value for the version.

[ashwin153]

Here are the results of importtime analysis of the three options using Tuna. TLDR pkg_resources is about 3 times slower than importlib_metadata, and importlib_metadata is about 5 times slower than hard-coding the version number.

Constant

importlib_metadata

pkg_resources

[hugovk]

Thanks for this!

That's such a big difference, perhaps we could try out the write_to option of setuptools_scm to hardcode it.

https://github.com/pypa/setuptools_scm/#configuration-parameters

Would you like to put together a PR?

[ashwin153]

Sounds good, will do tomorrow.

[ashwin153]

Note to self and future: Since the version file that setuptools writes to doesn't exist during development, we'll need to guard the import just like in pypa/setuptools-scm#143 (comment).

[hugovk]

Does installing in editiable mode (pip install -e .) generate the version file?

[ashwin153]

Closed by #227.

[hugovk]

Is the new importlib_metadata option good enough compared to a constant?

[ashwin153]

It is not good enough to consider constant but it is 3 times faster than it was before, which is why I closed this. Happy to reopen if you think that there is more to be done.