numpy dependency incorrect

[DManowitz]

As seen in the arraykit requirements.txt (https://github.com/static-frame/arraykit/blob/master/requirements.txt), since arraykit v0.2.4, the numpy dependency is numpy>=1.18.5. However, this package has the numpy dependency pinned at 1.18.5.

[flexatone]

Thanks, @DManowitz , for posting this issue.

I was only trying to pin NumPy at 1.18.5 for Python 3.7 and 3.8, as shown here:

https://github.com/conda-forge/arraykit-feedstock/blob/main/recipe/meta.yaml

Are you using that version of Python, or you seeing this issue in a different context? If you are getting a specific error, please share it. In the mean time, even if you are using conda, you should be able to install arraykit wheels with pip.

I also notice that the Azure builds are failing; I will investigate that as well.

[DManowitz]

I am using it with Python 3.8. I noticed the issue with arraykit v0.2.3 and static-frame v1.3.1 installed. With that combination, I got the following error when trying to import static_frame:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "...\lib\site-packages\static_frame\__init__.py", line 9, in <module>
    from static_frame.core.archive_npy import NPY as NPY
  File "...\lib\site-packages\static_frame\core\archive_npy.py", line 16, in <module>
    from static_frame.core.container_util import ContainerMap
  File "...\lib\site-packages\static_frame\core\container_util.py", line 17, in <module>
    from static_frame.core.container import ContainerBase
  File "...\lib\site-packages\static_frame\core\container.py", line 7, in <module>
    from static_frame.core.display import Display
  File "...\lib\site-packages\static_frame\core\display.py", line 14, in <module>
    from static_frame.core.display_config import _DISPLAY_FORMAT_HTML
  File "...\lib\site-packages\static_frame\core\display_config.py", line 9, in <module>
    from static_frame.core.interface_meta import InterfaceMeta
  File "...\lib\site-packages\static_frame\core\interface_meta.py", line 3, in <module>
    from static_frame.core.doc_str import doc_inject
  File "...\lib\site-packages\static_frame\core\doc_str.py", line 8, in <module>
    from static_frame.core.util import AnyCallable
  File "...\lib\site-packages\static_frame\core\util.py", line 27, in <module>
    from arraykit import first_true_1d
ImportError: cannot import name 'first_true_1d' from 'arraykit' (C:\Users\manow\miniconda3\envs\py38_ml_TfGpuMain\lib\site-packages\arraykit\__init__.py)

At that point, I started to dig into why I got that error and found the issues with the dependencies for the static-frame and arraykit conda-forge packages vs their requirements as listed in their own repos.

[flexatone]

I have removed the pinned NumPy for Python 3.7 and 3.8 and the build seems to work.

Thanks again for pointing this out.

[DManowitz]

I see that the NumPy pin at 1.18.5 has been removed for this package for versions 3.4 and 4.0. However, there is an additional requirement of numpy >=1.24.3 now also. The requirements.txt for arraykit only shows numpy >=1.18.5, so why has this additional requirement been added?

[DManowitz]

Any updates on this? I see that the numpy >=1.24.3 requirement is still present for recent arraykit packages.

[flexatone]

Hi @DManowitz , thanks for bringing this up, and apologies for the delay.

My best understanding is that this is out of my control. As you can see, the NumPy requirement that I specify in the conda recipe is numpy >= 1.18:

arraykit-feedstock/recipe/meta.yaml

Line 27 in 896b297

- numpy >=1.18.5

However, the conda build process, in producing the compiled packages, seems to define the requirement you are seeing elsewhere, and I do not see how I can change this. This is exactly why I was previously trying to pin 1.18.5 for Python 3.7 and 3.8.

The best alternative is to use pip (which you can do along with installing conda) to install pre-compiled wheels, which I make available for all platforms across many versions of Python. StaticFrame is actively tested using ArrayKit with NumPy 1.18.5 as shown in the requirements file used for those tests:

https://github.com/static-frame/static-frame/blob/master/requirements-test-backward.txt

Apologies for the challenges with this, but again, I think this is out of my control.

[DManowitz]

@flexatone I forgot to mention this earlier, but you might want to try asking in the conda-forge discussion group about this issue and someone might know the fix for it.