Different behavior of groupby with / without flox about xarray

[dcherian]

The error message mentions: Try passing .groupby(..., squeeze=False) which ends up working. This will be fixed once we finally make that the default. It's been warning for 7 months now.

[dcherian]

Or we can fix the bug here:

xarray/xarray/core/groupby.py

Lines 812 to 818 in d0048ef

	# weird backcompat
	# reducing along a unique indexed dimension with squeeze=True
	# should raise an error
	if (dim is None or dim == name) and name in obj.xindexes:
	index = obj.indexes[name]
	if index.is_unique and self._squeeze:
	raise ValueError(f"cannot reduce over dimensions {name!r}")

We need to raise the error for virtual variables too.

[max-sixty]

OK great!

(so I understand — when we change the default, would da.groupby(dim).mean() work — i.e. with no dim passed to mean? Or we'll still mandate a dim passed to mean? Currently da.groupby('time').mean() fails passes but da.groupby('lat').mean() fails, seemingly because lat is a float index... )

[dcherian]

Turns out it has to do with sortedness of lat. We only check.is_monotonic_increasing but yes if I pass squeeze=False it all works.

xarray/xarray/core/groupers.py

Lines 122 to 126 in d0048ef

	def is_unique_and_monotonic(self) -> bool:
	if isinstance(self.group, _DummyGroup):
	return True
	index = self.group_as_index
	return index.is_unique and index.is_monotonic_increasing

Could switch it to add is_monotonic_decreasing too.

#6220
#7427

[max-sixty]

OK great!

So it sounds like after we make that change, it won't be required to pass any dimensions to the aggregation func — da.groupby('lat').mean(), (with nothing passed to mean()), will aggregate over groups along lat without reducing any other dims. (lmk if I'm mistaken)

That'll be much better than the current state, where not passing anything to mean() works depending on the dimension's sortedness + whether flox is installed. Notably, if we only add is_monotonically_decreasing without making squeeze=False the default, it'll still depend on the array sortedness.

TY!

[dcherian]

Setting squeeze=False makes it always work.

#7427 removes the dependence on sortedness. That is, once merged, it will dependably error on main. Perhaps we should just delete the squeeze kwarg, it's been a pain for quite long.

[max-sixty]

OK v nice — on #7427 this dependably works on both flox & no flox, both ordered & unordered.

with xr.set_options(use_flox=True):
    print(da.assign_coords(lat=lambda x: x.lat % 2).groupby('lat', squeeze=False).sum())

[max-sixty]

Closed by #9280, thanks !