As an aside, it would be kind of nice to get the conda-forge bot to run on the rc so we can fix things before we tag.

isn't it only failing on ppc? Not sure how you can reproduce them locally.

Was the error only in test_splitting.py:446?

That's surprising because the grower also passes the same type of array, so all tests should fail.

I'm also just confused by the error message: expected 'const char' but got 'signed char': const and signed aren't mutually exclusive properties

It was everywhere that created a splitter. Here is the link to the one of the tests https://travis-ci.com/github/conda-forge/scikit-learn-feedstock/jobs/333125503

A cleaner error message:

python3.7/site-packages/sklearn/ensemble/_hist_gradient_boosting/grower.py:234: in __init__
    min_samples_leaf, min_gain_to_split, hessians_are_constant)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

>   ???
E   ValueError: Buffer dtype mismatch, expected 'const char' but got 'signed char'

I'm very confused as well, isn't signed char simply auto casted into const signed char?

what's the status on this, do we still need to fix something?

BTW @thomasjpfan there are no "changes" in this PR anymore

I think this still fails on ppc, but we're keeping releasing w/o ppc builds I think. It'd be nice to fix it.

whoops I reverted everything.

I am guessing removing the const (sadly) should fix this. But the only way to test is with conda-forge?

I guess so. You can temporarily push the change to a non release branch on the main repo and have it tested on conda-forge

I am actively trying different patches out in conda-forge/scikit-learn-feedstock#125

I will mark this PR to be merged when I find something that works.

(We can go from PR to draft now and back again!)

Did this one not work?

If it works and a quick benchmark shows no performance drop, maybe our best bet is to go with it so @thomasjpfan doesn't waste too much time...

The patch does not work. The monotonic constraints tests are failing https://travis-ci.com/github/conda-forge/scikit-learn-feedstock/jobs/335600845

cpdef enum MonotonicConstraint:
    NO_CST = 0
    POS = 1
    NEG = -1

When I was coding in C (in a previous life) I remember having some edge cases regarding some enums with values that were not 0, 1, 2... Good old memories.

May I suggest to try:

• stop using the enum and declare global vars: MNT_CST_NONE = 0, MNT_CST_POST = 1, MNT_CST_NEG = -1

• if it doesn't work, use an enum with values 0, 1, 2 (would need a remapping :/)

• if it doesn't work, use global vars 0, 1, 2 and declare the array as const unsigned char. I really hope this would work since has_missing_values is a const unsigned char and there is no issue with it

• if it doesn't work, lie down and cry

Fun stuff. Turns out explicitly saying that the char is signed works.

@NicolasHug creating the 0.23.1 PR now, should we merge this?

Thanks @thomasjpfan

I guess there's something fishy going on with the compiler used of the ppc CI...