Jérémie du Boisberranger

I changed it to hard because it requires some knowledge to figure out which test should use the fixture and which test should not.

I think the list should be updated to remove all estimators that do not preserve the dype yet. If an estimator doesn't doesn't preserve the dtype, it means that it...

The second option seems more like our usual practices :)

@thomasjpfan seems like you messed up you sync with main 😃

> was the IndexError you once observed on Windows by any chanced related to scikit-learn/sklearn/ensemble/_gradient_boosting.pyx? Or was it related to HGBT? No they are all related to the HGBT

> @jeremiedbb do you recall any issue with KMeans and some MacOS versions? No I don't. Maybe the output of the following could help ```python -m threadpoolctl -i sklearn```

I don't see anything weird @glemaitre another difference is that yours are packaged by conda-forge.

The issue is that minkowski with p< 1 is not a metric (no triangular inequality). Then the question is do we want to support similarity measures that are not metrics...

> Actually a pseudo metric would still satisfy the triangular inequality. Actually minkowski with p

I would allow it for brute force and raise for KD/ball-tree, but no strong opinion.