Allow .attrs to support any dict-likes

[Illviljan]

• ClosesAllow .attrs to use dict-likes #5655
• Tests added
• Passes pre-commit run --all-files
• User visible changes (including notable bug fixes) are documented in whats-new.rst
• New functions/methods are listed in api.rst

[pep8speaks]

Hello @Illviljan! Thanks for updating this PR. We checked the lines you've touched for PEP 8 issues, and found:

There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻

Comment last updated at 2021-10-31 09:57:53 UTC

[github-actions]

Unit Test Results

6 files 6 suites 56m 18s ⏱️
16 290 tests 14 550 ✔️ 1 739 💤 1 ❌
90 936 runs 82 737 ✔️ 8 198 💤 1 ❌

For more details on these failures, see this check.

Results for commit 650ce22.

♻️ This comment has been updated with latest results.

[Illviljan]

Both dict(x) and x.copy() are shallow copies.
Is it a good idea for it to be shallow though? Seems a little scary to me if the attrs happens to include mutable objects.

[max-sixty]

Yeah, I hear that, but IIUC that's what python generally does (I don't have a personally confident view though)

[Illviljan]

Merge had a nice test that was relevant to this:

xarray/xarray/tests/test_merge.py

Line 203 in 8f5b4a1

deftest_merge_attrs_override_copy(self):

This seems to imply that you would want to be able to tweak the attrs without worrying about other attrs changing. But the test is only made with immutable objects however. Lets do a test with:

ds1=xr.Dataset(attrs={"x": [0, 1]})
ds2=xr.Dataset(attrs={"x": 1})
ds3=xr.merge([ds1, ds2], combine_attrs="override")
ds3.attrs["x"][0] =2# assert ds1.x == [0, 1]print(ds1.x) # [2, 1]

That seems like surprising behaviour to me, but it's indeed also how python does this.

[max-sixty]

Yes, I agree. Copy-on-Write would be ideal but that's not python's strength...

[Illviljan]

Some fun performance comparisons related to copying and initializing dicts:

a=dict(a=2, b=3)
%timeitdict(a)
207ns ± 3.41nsperloop (mean ± std. dev. of7runs, 1000000loopseach)
%timeita.copy()
82.6ns ± 0.425nsperloop (mean ± std. dev. of7runs, 10000000loopseach)
importcopy%timeitcopy.copy(a)
313ns ± 3.59nsperloop (mean ± std. dev. of7runs, 1000000loopseach)
fromcopyimportcopy%timeitcopy(a)
290ns ± 3.63nsperloop (mean ± std. dev. of7runs, 1000000loopseach)
fromcopyimportdeepcopy%timeitdeepcopy(a)
3.39µs ± 55.6nsperloop (mean ± std. dev. of7runs, 100000loopseach)

Using a.copy() seems to be the way to go if you want to do a shallow copy of a dict.

[max-sixty]

Using a.copy() seems to be the way to go if you want to do a shallow copy of a dict.

That is an interesting result (even aside from the main result here, I'm not sure what python is doing such that copy.copy(a) has different performance from copy(a)!)

But python is slow, and nanos are short — unless there's a noticeable impact on the overall performance, then prioritizing flexibility and compatibility are more consistent with the goals of the library. Does that make sense?

[Illviljan]

I'm surprised about the deepcopy being so slow too, I thought it would be similar in speed in this case and just increase if dealing with mutable objects.

But using .copy is 100% compatible with how attrs has behaved before. So we come back to the question what type attrs should be?
The options I see is going with dict or MutableMapping.

I'm starting to lean towards mutablemapping because subclassing dict has been rather difficult compared to mutablemapping.

And if we go with mutablemapping then we should use copy.copy.

[max-sixty]

Is python/mypy#3004 still an issue?

pre-commit suggests it's here: https://github.com/pydata/xarray/pull/5667/files#diff-3c0ce7941684cbac55c00ab890684f86acc1de1908ee2afa915dbcb7c944105aR100 — but I guess there's some reason we can't only accept a MutableMapping in that function?

[Illviljan]

Is python/mypy#3004 still an issue?

pre-commit suggests it's here: https://github.com/pydata/xarray/pull/5667/files#diff-3c0ce7941684cbac55c00ab890684f86acc1de1908ee2afa915dbcb7c944105aR100 — but I guess there's some reason we can't only accept a MutableMapping in that function?

Yes, it is still an issue. I've cheated though and used type: ignore on a few places, that's why its been passing the checks.

Mappings (in the form of FrozenDict) are used to initialize .attrs in xr.open_dataset for example. So we can't unfortunately accept MutableMappings only.

Does pyright handle properties with setters and getters?

[Illviljan]

Here's further tests to check how fast different class checkers are:

fromtypingimportMutableMappingclassTest2(MutableMapping):
def__init__(self, *args, **kwargs):
self.data=dict(*args, **kwargs)
def__getitem__(self, key):
passdef__setitem__(self, key, value):
passdef__delitem__(self, key):
passdef__iter__(self):
passdef__len__(self):
passb=Test2()
%timeitissubclass(type(b), MutableMapping)
711ns ± 5.33nsperloop (mean ± std. dev. of7runs, 1000000loopseach)
%timeitisinstance(b, MutableMapping)
853ns ± 6.29nsperloop (mean ± std. dev. of7runs, 1000000loopseach)
# If you want to get really fast you can check for one of the required attributes MutableMapping has %timeithasattr(b, "update")
82.6ns ± 0.181nsperloop (mean ± std. dev. of7runs, 10000000loopseach)

isinstance is rather slow as can be seen. Considering just doing dict(b) takes about 200ns which is basically what the original implementation was it doesn't feel that good to add a check that adds 800ns of wait time.

[headtr1ck]

If that is still an open issue we could merge current main, try to fix the resulting typing problems.