Basic example hangs in 0.4dev

The basic example in the README hangs when using UMAP from the master or 0.4dev branch.

import umap
from sklearn.datasets import load_digits

digits = load_digits()

embedding = umap.UMAP().fit_transform(digits.data)

If I do

pip install umap-learn

then the code works as expected (taking only 5-10 seconds). However, it hangs when using a version of UMAP built from the 0.4dev branch or master. It looks like it is doing something since I see the message

/home/gclenden/anaconda3/envs/umap_debug/lib/python3.7/site-packages/umap/spectral.py:243: UserWarning: Embedding a total of 2 separate connected components using meta-embedding (experimental)
  n_components

I think the commit where this occurs is https://github.com/lmcinnes/umap/commit/eb6774056aa5fc8a858d466eaba4a1cb35c978ae. It works as expected on the commit prior to that one https://github.com/lmcinnes/umap/commit/c26bc8dac3e8fc71a4221986cf3cf0926b573766. However, I don’t know enough about numba/tbb to explain why this might be happening.

[lmcinnes]

One thing to try would be rolling back to numba=0.46 and llvmlite=0.30. There were defintiely some issues there, but that may not be this problem. I’ll try to look in more detail tomorrow. Worst case you can probably remove the requirement and uninstall tbb to get things working.

[lmcinnes]

It turns out that it is an issue specific to numba 0.47 (there was bug; see numba/numba:#5087). The current master (which has 0.4dev merged in) updates the requirements files accordingly. I will update 0.4dev as well, but people should start working from master now.