Incompatibility with numpy 1.17 about pynndescent

[slinnarsson]

In case it helps anyone, I had the same issue on a brand new install of Anaconda (version 2019.07 for macOS), which has numba version 0.44.1. Doing conda update numba gave me numba version 0.45.1, which resolved the issue.

[lmcinnes]

That doesn't look good. I'm not sure exactly what would have changed to cause something like this -- it is definitely numba related. I'm on vacation right now, but I'll try to look at this when I get back.

[tomwhite]

I wasn't able to reproduce this with the following:

python3 -m venv venv
. venv/bin/activate
pip install pynndescent
python -c 'import pynndescent' # succeeds

This is with Python 3.7.2 on Mac. The package versions are:

Package      Version
------------ -------
joblib       0.13.2 
llvmlite     0.29.0 
numba        0.45.1 
numpy        1.17.0 
pip          18.1   
pynndescent  0.3.0  
scikit-learn 0.21.3 
scipy        1.3.1  
setuptools   40.6.2 

[pavlin-policar]

Sorry, I forgot to include the environment in the initial issue. I'm getting this consistently in a conda environment with

# Name                    Version                   Build  Channel
_libgcc_mutex             0.1                        main  
ca-certificates           2019.5.15                     0  
certifi                   2019.6.16                py37_0  
joblib                    0.13.2                   pypi_0    pypi
libedit                   3.1.20181209         hc058e9b_0  
libffi                    3.2.1                hd88cf55_4  
libgcc-ng                 9.1.0                hdf63c60_0  
libstdcxx-ng              9.1.0                hdf63c60_0  
llvmlite                  0.29.0                   pypi_0    pypi
ncurses                   6.1                  he6710b0_1  
numba                     0.44.1                   pypi_0    pypi
numpy                     1.17.0                   pypi_0    pypi
openssl                   1.1.1c               h7b6447c_1  
opentsne                  0.3.8                    pypi_0    pypi
pip                       19.1.1                   py37_0  
pynndescent               0.3.1                     dev_0    <develop>
python                    3.7.3                h0371630_0  
readline                  7.0                  h7b6447c_5  
scikit-learn              0.21.2                   pypi_0    pypi
scipy                     1.3.1                    pypi_0    pypi
setuptools                41.0.1                   py37_0  
sqlite                    3.28.0               h7b6447c_0  
tk                        8.6.8                hbc83047_0  
wheel                     0.33.4                   py37_0  
xz                        5.2.4                h14c3975_4  
zlib                      1.2.11               h7b6447c_3  

Channel pypi indicates that it was installed by running pip install <package>, the others were installed using conda. I'm running Arch Linux x86_64 with kernel 5.1.15-arch1-1-ARCH.

[pavlin-policar]

I noticed my environment had some openTSNE things in it, so I set up a new minimal environment like this

conda create --name tmp python=3.7.3
conda activate tmp
pip install pynndescent
python -c 'import pynndescent'           

which now results in a segfault

zsh: segmentation fault (core dumped) python -c 'import pynndescent'

The full environment ends up being

# Name                    Version                   Build  Channel
_libgcc_mutex             0.1                        main  
ca-certificates           2019.5.15                     1  
certifi                   2019.6.16                py37_1  
joblib                    0.13.2                   pypi_0    pypi
libedit                   3.1.20181209         hc058e9b_0  
libffi                    3.2.1                hd88cf55_4  
libgcc-ng                 9.1.0                hdf63c60_0  
libstdcxx-ng              9.1.0                hdf63c60_0  
llvmlite                  0.29.0                   pypi_0    pypi
ncurses                   6.1                  he6710b0_1  
numba                     0.45.1                   pypi_0    pypi
numpy                     1.17.0                   pypi_0    pypi
openssl                   1.1.1c               h7b6447c_1  
pip                       19.1.1                   py37_0  
pynndescent               0.3.0                    pypi_0    pypi
python                    3.7.3                h0371630_0  
readline                  7.0                  h7b6447c_5  
scikit-learn              0.21.3                   pypi_0    pypi
scipy                     1.3.1                    pypi_0    pypi
setuptools                41.0.1                   py37_0  
sqlite                    3.29.0               h7b6447c_0  
tk                        8.6.8                hbc83047_0  
wheel                     0.33.4                   py37_0  
xz                        5.2.4                h14c3975_4  
zlib                      1.2.11               h7b6447c_3  

[pavlin-policar]

After a bit more hunting, I found the issue is not specific to pynndescent, but to numba in general, and it only occurs in a conda environment.