Optionally enforce exact distances

[jondo]

Since the threshold of 4096 for deciding the distance algorithm is a bit arbitrary and depends on machine speed, this pull request introduces another argument force_exact_distancesas an opposite to the existing force_approximation_algorithm.

[pep8speaks]

Hello @jondo! Thanks for updating this PR. We checked the lines you've touched for PEP 8 issues, and found:

There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻

Comment last updated at 2021-05-10 13:33:14 UTC

[coveralls]

Coverage decreased (-0.01%) to 88.372% when pulling 431787f on jondo:force_exact_distances into 47b585c on lmcinnes:master.

[lmcinnes]

This could certainly be useful, but it will be brutally slow (and incredibly memory hungry) for a number of use cases. It might be worth at least using sklearn's NearestNeighbor or KNeighborTransformer classes for the computation rather than computing all pairwise distances (which is fine for small data, but perilously memory expensive for large data).

[jondo]

@lmcinnes , this sounds plausible, but I won't do it now. Feel free to close this request.