I added #16232 which skips a failing doctest

and #16040 which avoids a deprecation warning for pandas sparse array

Actually I merged #16586 too quickly: doc/index.rst also needs to be updated to add a link to 0.22.2 from the news section of the home page.

I'm not sure about that. This file no longer exists, it's now in template/index.html and the version seems automatically detected. It's correctly updated in the artifacts

you must have been looking to the old maintainers guide :)

indeed it's now in template/index.html but we still need a new line such as as:

        <li><strong>February 2020.</strong> scikit-learn 0.22.2 is available for download (<a href="whats_new/v0.22.html#version-0-22-2">Changelog</a>).

Right but that should be done after the release right ?

Right but that should be done after the release right ?

Yes but why not make it part of the release PR?

Actually maybe it's better no to update the homepage of scikit-learn before the wheels are on pypi.org.

As you wish. But building the website twice is kind of a waste :)

Right but that should be done after the release right ?

I kinda tend to put it in the release PR.

Actually maybe it's better no to update the homepage of scikit-learn before the wheels are on pypi.org.

The docs are updated when you merge this PR anyway. It's just a matter of what we include before, what after. I find it nicer if the tag is to the point where the wheels are made from.

I'm a bit lost here. updating the doc here will only impact the stable version of the web site when merged, right ?
we'll then need to also update the dev version in a PR pointing to master ?

Yes it only changes the stable one. For certain things you don't need to change anything with a bug-fix release. For instance, with this release the 0.23 doesn't change in the dev/master. But You need to add the bug-fix entry also to the main page in master if you add it here.

So I checked and for 0.22.1 it was done in 2 PRs, one for the dev branch and one backported to 0.22.X after the release.

I'm ok to include it here and then make a PR to master after the release. Is it ok for you ?

Sounds good to me

I made a sdist and built it in a fresh env and observed weird behavior. pip first gather numpy scipy and joblib then complains that numpy is not installed but keep working and build and install sklearn successfully. Can someone try locally and tell me if he reproduces ?

Then I ran the test suite and there's one test failing: test_ard_accuracy_on_easy_problem

def test_ard_accuracy_on_easy_problem():
        # Check that ARD converges with reasonable accuracy on an easy problem
        # (Github issue #14055)
        # This particular seed seems to converge poorly in the failure-case
        # (scipy==1.3.0, sklearn==0.21.2)
        seed = 45
        X = np.random.RandomState(seed=seed).normal(size=(250, 3))
        y = X[:, 1]
    
        regressor = ARDRegression(n_iter=600)
        regressor.fit(X, y)
    
        abs_coef_error = np.abs(1 - regressor.coef_[1])
        # Expect an accuracy of better than 1E-4 in most cases -
        # Failure-case produces 0.16!
>       assert abs_coef_error < 0.01
E       assert 0.018021599217421413 < 0.01

That seems to be a "known" issue. Although we don't seem to have worked on it: #15420, #16097, #15186

I still don't get in which order the setups are run. But I expect numpy to be installed by the time setup.py runs, which needs numpy. I vaguely remember having to install numpy before installing certain packages cause it wouldn't understand that it's the setup itself which needs numpy.

To get the wheels to build on the MacPython CI it might safer to skip the test_ard_accuracy_on_easy_problem test in the 0.22.X branch with a link to the issue tracker.

pip first gather numpy scipy and joblib then complains that numpy is not installed but keep working and build and install sklearn successfully.

This will probably be fixed in master with pyproject.toml that explicitly gives build dependencies. But I think it's not a problem to leave it the way it is for 0.22.2.

It's green, should we merge and tag ?

@ogrisel wanna double check and merge?

fwi the PR on MacPython /scikit-learn-wheels (MacPython/scikit-learn-wheels#47) is green when tested against this branch

Go!