Skip mean over empty axis

[dschwoerer]

Avoids changing the datatype if the data does not have the requested
axis.

• Closes Dataset.mean changes variables without specified dimension #4885
• Tests added
• Passes pre-commit run --all-files
• User visible changes (including notable bug fixes) are documented in whats-new.rst

[max-sixty]

Thanks for the PR @dschwoerer !

I think the result is right.

It's implemented as a special-case, which isn't ideal. But I'm not sure how to abstract the rule that applies to mean but not to std — something like "does f([x]) == x". I guess it should apply to .sum too.

For other reviewers — here's an example of the std case (which this PR doesn't change, correctly):

In [5]: ds
Out[5]:
<xarray.Dataset>
Dimensions:  (pos: 3, time: 2)
Coordinates:
  * pos      (pos) int64 1 2 3
Dimensions without coordinates: time
Data variables:
    data     (pos, time) float64 1.0 2.0 3.0 4.0 5.0 6.0
    var      (pos) int64 2 3 4

In [6]: ds.std('time')
Out[6]:
<xarray.Dataset>
Dimensions:  (pos: 3)
Coordinates:
  * pos      (pos) int64 1 2 3
Data variables:
    data     (pos) float64 0.5 0.5 0.5
    var      (pos) float64 0.0 0.0 0.0    # this should be 0.0, unlike `.mean`'s result

[max-sixty]

:py:func:mean is actually a method (there should be other references in the whats-new which can be copied)

[dschwoerer]

I haven't found an example, so I have removed the explicit reference

[max-sixty]

Convention:

Suggested change

	ds2 = ds.mean(dim="time")
	result = ds.mean(dim="time")

[dschwoerer]

I do not store the result but use it directly. Is that ok as well?

[max-sixty]

FYI you can construct the dataset in a single expression, but it's also fine as-is (Dataset(dict(pos=..., data=...)))

[dschwoerer]

I have now changed so that several wrapped functions preserve the data.

It is more generic, and hopefully still readable. The flag might also be called identity_0d because the function is the identity function for 0d data, while invariant_0d means that any 0d data is invariant under this function.

[max-sixty]

Excellent! Thanks @dschwoerer .

Any thoughts from others before we merge, particularly on the identity_0d approach?

[max-sixty]

Merging — as ever, any additional feedback we can address in another PR.

Thank you very much @dschwoerer , we're happy to have you as a contributor!