Add glob to create_ensembles and mention flox in doc.

[aulemahal]

Pull Request Checklist:

• This PR addresses an already opened issue (for bug fixes / features)
  • This PR fixes run test against xarray main #1080
• Tests for the changes have been added (for bug fixes / features)
  • (If applicable) Documentation has been added / updated (for bug fixes / features)
• HISTORY.rst has been updated (with summary of main changes)
  • Link to issue (:issue:number) and pull request (:pull:number) has been added
• bumpversion patch has been called on this branch
• The relevant author information has been added to .zenodo.json

What kind of change does this PR introduce?

• This adds the possibility to pass a glob string to create_ensemble instead of a list of path.
• Added flox to the upstream test suite.
• Added mentions of flox in the installation instructions.
• Removed deprecated extra deps example in installation instructions and move clisops to the extra section.

Does this PR introduce a breaking change?

No.

Other information:

This is a very minor change, more like a reason to trigger the github actions and run all tests, including the one against xarray's main branch.

[aulemahal]

Oh! Well @dcherian, it looks like flox and the main of xarray breaks xclim! It seems the error are related to a different treatment of nans and errors in the results precision. Could the new algorithm modify the results slightly?

I'll test in more details on my side.

[coveralls]

Coverage decreased (-0.005%) to 91.656% when pulling daa1c8c on ensemble_glob into b08fa84 on master.

[dcherian]

Yeah I had to make some tests a little looser w.r.t precision.

The other failures look real :( NaNs in the wrong place; grouping by cftime;

[dcherian]

I noticed in some places you're using np.testing to compare xarray objects. Should be nicer to use xarray.testing instead (it will compare coordinate locations too)

    def test_season(self, tasmin_series, calendar):
        ts = tasmin_series(np.zeros(360))
        ts = convert_calendar(ts, calendar, missing=0, align_on="date")
    
        miss = missing.missing_any(ts, freq="YS", season="MAM")
>       np.testing.assert_equal(miss, [False])
E       AssertionError: 
E       Items are not equal:
E        ACTUAL: <xarray.DataArray (time: 1)>
E       array([ True])
E       Coordinates:
E         * time     (time) datetime64[ns] 2001-01-01
E        DESIRED: [False]

[aulemahal]

Some errors are caused by the resampling of cftime arrays, I opened an issue : pydata/xarray#6613, But for now, we can simply call those with a xr.set_options(use_flox=False) context.

[aulemahal]

Most other errors seem to be because of pydata/xarray#6615:
Flox uses np.add for the sumoperation. But np.add acts as a OR on boolean values, while np.sum implicitly converts booleans to integers.

Xclim extensively resamples and sums over boolean values. The older xclim code usually explicitly casted those boolean arrays to int by multiplying by 1. But the newer code assumes the implicit behaviour from numpy/xarray and these parts are now broken.

[aulemahal]

@tlogan2000 This is ready for review.

[tlogan2000]

This looks good to me! @dcherian thank-you for this addition to xarray I think xclim peformance will benefit a lot from this change.

@aulemahal @dcherian just so it's clear in my mind : what is expected behavior going forward on something like (xr.DataArray() > thresh).resample(time='YS').sum(dim='time') ? Will the booleans be cast to ints in the end with flox?

[dcherian]

Will the booleans be cast to ints in the end with flox?

Yes it will promote bool to int before aggregating. (xarray-contrib/flox@01b1afe). You could require flox >=0.5.3 to be extra sure.

[Zeitsperre]

@dcherian Thanks again for your work and considerations!