When dask is missing we get the following error message:

python-traceback ValueError: unrecognized chunk manager dask - must be one of: []

this could be confusing - the error message seems geared towards a typo in the requested manager. However, I think it's much more likely that a chunk manager is just not installed. I tried to update the error message - happy to get feedback.

Supress some more internal warnings in the test suite.

• [ ] Adresses points 2 and 3 and closes #8612
• [ ] Tests added
• [ ] User visible changes (including notable bug fixes) are documented in whats-new.rst
• [ ] New functions/methods are listed in api.rst

Probably not ready for review yet.

• [x] Closes #7104
• [x] Tests added
• [x] User visible changes (including notable bug fixes) are documented in whats-new.rst
• [ ] New functions/methods are listed in api.rst

Unstacking non-unique MultiIndex can lead to silent data loss, so we raise an error.

Fix a warning from ruff concerning the config:

warning: The top-level linter settings are deprecated in favour of their counterparts in the lint section. Please update the following options in pyproject.toml: - 'extend-safe-fixes' -> 'lint.extend-safe-fixes' - 'per-file-ignores' -> 'lint.per-file-ignores'

• [ ] Closes #xxxx
• [x] Tests added
• [x] User visible changes (including notable bug fixes) are documented in whats-new.rst

This allows negative freq strings as discussed in https://github.com/pydata/xarray/pull/8627#issuecomment-1905981660 Deciding which tests to update was not easy.

The pandas _generate_range function was moved to https://github.com/pandas-dev/pandas/blob/3c96b8ff6d399fbec8d4d533e8e8618c592bb64b/pandas/core/arrays/datetimes.py#L2725 They no longer rollback the end. I had to remove this as well such that the following are eqivalent:

python xr.date_range("2001", "2000", freq="-1YE", calendar="noleap") pd.date_range("2001", "2000", freq="-1YE")

I am slightly nervous about this but all the tests still pass...

Once again cc @spencerkclark

use ruff.flake8-tidy-imports to enforce absolute imports

• https://github.com/MarcoGorelli/absolufy-imports has been archived (no reason given)
• removes a pre-commit hook which should make it faster locally

• [ ] Closes #xxxx
• [ ] Tests added
• [ ] User visible changes (including notable bug fixes) are documented in whats-new.rst
• [ ] New functions/methods are listed in api.rst

I think we should use extend-exclude instead of exclude for ruff. We can then also remove ".eggs" as this is in the default.

From https://docs.astral.sh/ruff/settings/#exclude:

Note that you'll typically want to use extend-exclude to modify the excluded paths.

Default value: [".bzr", ".direnv", ".eggs", ".git", ".git-rewrite", ".hg", ".mypy_cache", ".nox", ".pants.d", ".pytype", ".ruff_cache", ".svn", ".tox", ".venv", "__pypackages__", "_build", "buck-out", "build", "dist", "node_modules", "venv"]

(I really dislike how github formats toml files... What would be the correct syntax, then?)

I realized that the return value of infer_freq was not updated. #8627 will try to suppress all warnings in the test suite, so this is just the minimal PR.

Sorry for all the spam @spencerkclark

• [x] Adresses point 1 of #8612
• [x] Fixes one of the failures in #8623
• [x] Tests added
• [ ] User visible changes (including notable bug fixes) are documented in whats-new.rst

This renames the frequency string "Y" (formerly "A") to "YE" to achieve pandas parity. It could be better to wait for the conclusion of pandas-dev/pandas#56840 before doing this (but fixing the related failure in #8623 seemed a good reason as any to do it know).

Let me know what you think @spencerkclark @aulemahal

• [x] Closes #7298
• [x] Tests added
• [ ] User visible changes (including notable bug fixes) are documented in whats-new.rst
• [ ] New functions/methods are listed in api.rst

Otherwise da.indexes and the html repr raise a ValueError. I first had "<undefined>" but I think None is better. cc @spencerkclark @keewis

By chance I saw that in #4700 the same code block was added twice. I think this can be removed.

cc @andersy005

• [x] Closes #8367
• [x] Tests added
• [ ] User visible changes (including notable bug fixes) are documented in whats-new.rst
• [ ] New functions/methods are listed in api.rst

Uses max with a default, which work with empty iterators, in contrast to if col_items else 0.

• [x] User visible changes (including notable bug fixes) are documented in whats-new.rst
• [ ] New functions/methods are listed in api.rst

This makes many arguments keyword-only, except for dim to avoid da.weighted(...).mean("lat", "lon") (i.e. da.weighted(...).mean(dim="lat", skipna="lon")) which silently does the wrong thing. I am sure I forgot some and for some I was unsure so I left them as is.

Question: do we want an deprecation cycle? Currently it just errors for da.weighted(...).mean("dim", True). Might be nice to do it, however, @dcherian if I am not mistaken you did this without a deprecation in #5950, e.g. for da.mean etc.?

python import numpy as np import xarray as xr air = xr.tutorial.open_dataset("air_temperature") wgt = np.cos(np.deg2rad(air.lat)) air.weighted(wgt).mean("lat", "lon")

• [ ] Towards #8239
• [ ] Tests added
• [ ] User visible changes (including notable bug fixes) are documented in whats-new.rst
• [ ] New functions/methods are listed in api.rst

This partly updates the pytest config as suggested in #8239 and un-xfails some tests (or xfails the tests more precisely).

See https://github.com/pydata/xarray/issues/8239#issuecomment-1739363809 for why we cannot exactly follow the suggestions given in #8239

• [x] Tests added
• [ ] User visible changes (including notable bug fixes) are documented in whats-new.rst
• [ ] New functions/methods are listed in api.rst

I found the error message in ds.to_stacked_array confusing, so I tried to make it clearer. Also renames some if the internal symbols (so should have no user facing change).

Suppress the pkg_resources.namespace_package DeprecationError to make the doctest pass again (similar to #7322). This is reported upstream: pydap/pydap#277 and matplotlib/matplotlib#25244

• [x] Closes #6092
• [x] Passes pre-commit run --all-files
• [x] User visible changes (including notable bug fixes) are documented in whats-new.rst

One change is that it is no longer possible to compare to a string, i.e. version.parse(xr.__version__) < "0.20.0" errors.

As mentioned in #6092 there are 3 options - if there is a preference I am happy to update this PR.

```python from distutils.version import LooseVersion from packaging import version

LooseVersion(xr.version) version.parse(xr.version) version.Version(xr.version)

currently:

if LooseVersion(mod.version) < LooseVersion(minversion): pass

options:

if version.parse(mod.version) < version.parse(minversion): pass

if version.Version(mod.version) < version.Version(minversion): pass

if Version(mod.version) < Version(minversion): pass ```

• [x] Closes #7322

This PR renames the da and ds fixtures (to da_fixture and ds_fixture) and moves them to conftest.py. This allows to remove the flake8 error suppression for the tests and seems more how the fixtures are intended to be used (from the pytest side). I think the name changes makes it a bit more obvious what happens but moving them to may make it a bit less obvious (if you don't know where to look).

Removing the flake8 error ignores also unearthed some unused imports:

https://github.com/pydata/xarray/blob/787a96c15161c9025182291b672b3d3c5548a6c7/setup.cfg#L155-L156

(What I actually wanted to do is move the tests for rolling to it's own file - but I think it makes sense to do this first.)

I forgot to remove some debug print statements in #6910 - thanks for noting @shoyer & @dcherian

flake8 v6 now errors on inline comments in the config file. I don't like it but oh well...

In #4864 I added a a decorator for the xarray.testing.assert_* functions to ensure warnings that were to errors (pytest.mark.filterwarnings("error")) do not error in assert_* (see https://github.com/pydata/xarray/pull/4760#issuecomment-774101639). As a solution I added

https://github.com/pydata/xarray/blob/5470d933452d88deb17cc9294a164c4a03f55dec/xarray/testing.py#L32

However, this is sub-optimal because this now removes all ignore filters! As dask stuff only gets evaluated in assert_* filters like warnings.filterwarnings("ignore", "Mean of empty slice") don't work for dask arrays!

I thought of setting

python warnings.simplefilter("ignore")

but this could suppress warnings we want to keep.

So now I remove all "error" warning filters and keep the rest. Note that the original filters get restored after with warnings.catch_warnings():. ().

I am not sure I expressed myself very clearly... let me know and I can try again. @keewis you had a look at #4864 maybe you can review this PR as well?

I tried to use the _deprecate_positional_args decorator from #6934 & it turns out that it still had some errors - passing on the arguments did not work properly in certain cases... I now added tests for this as well (which I should have done in the first place...).

5844 excluded pydap from our tests - but the new version has been released in the meantime (on conda not on pypi, though, pydap/pydap#268) - so let's see if this still works.

• [x] Closes #6960
• [ ] Tests added - tested manually

Get rid of some warnings in the tests.

• [x] Closes #6932

@zmoon - obviously it would be nice if we had a linter for this but this is for another time.

• [x] Supersedes #6403, see also #5531
• [x] Tests added
• [ ] User visible changes (including notable bug fixes) are documented in whats-new.rst
• [ ] New functions/methods are listed in api.rst

Adds a helper function to deprecate positional arguments. IMHO this offers a good trade-off between magic and complexity. (As mentioned this was adapted from scikit-learn).

edit: I suggest to actually deprecate positional arguments in another PR.

fix some typos (using codespell). Called using:

bash codespell --skip=".git,.mypy_cache,*.tex" --ignore-words-list coo,nd,inferrable,hist,ND,splitted,soler,slowy,ba,ser,nin,te,fo -w -i 3

• [x] Tests added
• [x] User visible changes (including notable bug fixes) are documented in whats-new.rst

skipna=None did not skip missing values for quantile, inconsistent with other methods. Discovered while testing #6059.

• [x] Closes #6136

Convert the code block in xr.DataArray.from_dict and xr.Dataset.from_dict to doctest/ examples.

pytest v7.0.0 no longer want's us to use pytest.warns(None) to test for no warning, so we can use warnings.catch_warnings(record=True) instead.

numpy/numpy#20327 introduces some changes to np.quantile (and related) for the upcoming numpy release (v1.22.0). It renames the interpolation keyword to method and offers some new interpolation methods. This PR does two things

1. it restores compatibility with numpy 1.22

2. it renames the interpolation keyword to method in xarray - this change is not strictly necessary but I thought better to be consistent with numpy

3. [x] Tests added

4. [x] Passes pre-commit run --all-files
5. [x] User visible changes (including notable bug fixes) are documented in whats-new.rst

(Side note in dask.array.percentile the method keyword is used differently from the interpolation keyword (docs). However, xarray does not use the dask function.)

TODO: need to import ArrayLike from npcompat.

• [x] Closes #6186

pydata/bottleneck#378 was merged - so this should work again.

Fixes or suppresses a number of warnings that turn up in our upstream CI.

pd.Index.is_monotonic is an alias for pd.Index.is_monotonic_increasing and does not stand for pd.Index.is_monotonic_increasing or pd.Index.is_monotonic_decreasing.

Reverts pydata/xarray#6200

And another tiny one... The somewhat ambiguous date string triggers a warning in pandas which makes our doc build fail.

• [x] Closes #5721
• [ ] Tests added
• [ ] User visible changes (including notable bug fixes) are documented in whats-new.rst

Fixed as per @shoyer & @spencerkclark suggestion from https://github.com/pydata/xarray/issues/5721#issuecomment-903095007

Now that pandas 1.4 is out it would be good to get this fixed (there are about 5000 warnings in our tests, mostly because of interp, though). Also leads to a warning in our docs which breaks them (although that can also be fixed with an :okwarning: directive).

Fixes the doctests for the newest version of cftime. @spencerkclark

This of course means that the doctests will fail for environments with older versions of cftime present. I don't think there is anything we can do.

Thanks for pytest-accept b.t.w @max-sixty

5740 changed PendingDeprecationWarning to FutureWarning - suppress the warnings again in the test suite.

https://github.com/pydata/xarray/blob/36f05d70c864ee7c61603c8a43ba721bf7f434b3/xarray/ufuncs.py#L47-L49

• [x] Closes #6016

cftime 1.5.2 renames "gregorian" to "standard". AFAIK this only changes the repr of cftime indices and does not seem to influence the creation of cftime indices.

In #6096 I realized that DuckArrayModule("dask") is called a lot in our tests - 145'835 times. Most of those are from is_dask_collection (is_duck_dask_array) This change avoids that the instance needs to be built every time.

```python import xarray as xr

%timeit xr.core.pycompat.DuckArrayModule("dask").available %timeit xr.core.pycompat.dsk.available ```

18.9 µs ± 97.7 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each) 77.1 ns ± 1.22 ns per loop (mean ± std. dev. of 7 runs, 10000000 loops each)

Which leads to an incredible speed up of our tests of about 2.7 s :grin: ((18.9 - 0.0771) * 145835 / 1e6).

• [x] Closes #4601
• [x] Passes isort . && black . && mypy . && flake8
• [x] User visible changes (including notable bug fixes) are documented in whats-new.rst

It's not pretty as I had to define a number of empty methods... I think this should wait for 0.17

• [x] Closes #4727
• [ ] Tests added
• [ ] Passes isort . && black . && mypy . && flake8
• [ ] User visible changes (including notable bug fixes) are documented in whats-new.rst
• [ ] New functions/methods are listed in api.rst

This adds a dtype check for equal, identical, broadcast_equal, and the xr.testing.assert_* functions. It is far from complete: tests and documentation are still missing, but I wanted to get it online for feedback.

When I set check_dtype=True there are around 600 failures. Fixing that is for another PR. #4759 should help a bit.

• [ ] I added the checks to lazy_array_equiv, however, sometimes dask can get the dtype wrong before the compute (see below). Do you think I need to put it in the non-lazy part?

```python import numpy as np import xarray as xr

da = xr.DataArray(np.array([0, np.nan], dtype=object)).chunk()

da.prod().dtype # -> dtype('O') da.prod().compute().dtype # -> dtype('int64')

```

• [ ] check_dtype is still missing from assert_duckarray_allclose & assert_duckarray_equal - do you think there are required?

• [ ] The dtypes of array elements are not tested (see below). I don't think I'll implement that here.

```python da0 = xr.DataArray(np.array([0], dtype=object)) da1 = xr.DataArray(np.array([0.], dtype=object))

xr.testting.assert_equal(da0, da1, check_dtype=True) ```

• dask 2022.01 is out, so we can remove the pin

I hunted down some version checks that should no longer be necessary as we have moved beyond the minimum versions.

• [x] see #6186

I think it would be good to re-enable the upstream CI, even if this means we have to stick to py3.9 for the moment. I just subscribed to pydata/bottleneck#378, so I should see when we can switch to 3.10.

typing fixes for mypy 0.931 and numpy 1.22. Also tested with numpy 1.20 which probably many still have installed.

Our CI fails with some pytest-xdist error. Let's see if we get a clearer picture when disabling parallel tests. (Maybe some interaction between dask and pytest-xdist?).

• [x] Closes #6101

Reverts pydata/xarray#6077 (after dask has been pinned in #6111)

h5netcdf no longer warns for invalid netCDF (unless passing save_kwargs = {"invalid_netcdf": True}. We need to adapt our tests.

@kmuehlbauer

edit: I added h5netcdf to the upstream tests - I can also revert this change if you prefer.

• [x] Closes #5872
• [x] Passes pre-commit run --all-files

pandas will disallow skipna=None (pandas-dev/pandas#44178) - this fixes a test which relies on this. I don't think we have any user facing use of this. AFAIK we don't use pandas for reductions anywhere)

Fixes one part of #5872

pseudoNETCDF adds two attrs to ict files, which breaks the following two tests:

Test 1: https://github.com/pydata/xarray/blob/07de257c5884df49335496ee6347fb633a7c302c/xarray/tests/test_backends.py#L3944 Test 2:

https://github.com/pydata/xarray/blob/07de257c5884df49335496ee6347fb633a7c302c/xarray/tests/test_backends.py#L4030

I reproduced the test file so that the tests pass again. To reproduce the file I used the following bit of code:

```python import xarray as xr from xarray.tests import test_backends

fN = "xarray/tests/data/example.ict" fmtkw = {"format": "ffi1001"}

ds = xr.open_dataset(fN, engine="pseudonetcdf", backend_kwargs={"format": "ffi1001"})

c = test_backends.TestPseudoNetCDFFormat() c.save(ds, fN, **fmtkw) ```

The save method is here:

https://github.com/pydata/xarray/blob/07de257c5884df49335496ee6347fb633a7c302c/xarray/tests/test_backends.py#L4124

@barronh I would appreciate your review here - I am not sure if this is the right approach.

• [x] Part of #4513
• [x] Tests added
• [x] Passes pre-commit run --all-files
• [x] User visible changes (including notable bug fixes) are documented in whats-new.rst

This PR warns when passing keep_attrs to resample(..., keep_attrs=True) and rolling_exp(..., keep_attrs=True) as they have no effect (rightfully). Also removes keep_attrs from the docstring of resample.

• [x] Part of #4513 (maybe the last one - need to double check)
• [x] Tests added
• [x] Passes pre-commit run --all-files
• [x] User visible changes (including notable bug fixes) are documented in whats-new.rst

As per https://github.com/pydata/xarray/issues/3891#issuecomment-612522628 I also changed the default to True.

Another enormous PR from my side ;) unpin pip again. numpy probably fixed the issue re the name of the nightly build. But I also need to doublecheck if scipy is ok.

• [x] Closes #4858
• [x] Passes pre-commit run --all-files

iris moved its documentation to https://scitools-iris.readthedocs.io/en/stable/

There is a compatibility issue with h5py v3. Pin h5py to version 2 for the moment. I can open an issue shortly.

• [x] Closes #4521
• [x] Passes isort . && black . && mypy . && flake8

That was fast - eccodes already fixed the issue.

and put the comment on the same line

• [x] Closes #3756
• [x] Fully documented, including whats-new.rst for all changes and api.rst for new API

• [x] Closes #2019
• [ ] Tests added
• [x] Passes isort -rc . && black . && mypy . && flake8
• [ ] Fully documented, including whats-new.rst for all changes and api.rst for new API
• [x] reverts #2021

Let's see if this passes....

• [x] Closes #3634
• [x] Tests added
• [x] Passes black . && mypy . && flake8
• [x] Fully documented, including whats-new.rst for all changes and api.rst for new API

np.nanquantile was added in numpy 1.15.0, the current minimum requirement for numpy is 1.14.0, therefore we have to test this ourselves.

• [x] Closes #5517
• [x] Passes pre-commit run --all-files

The minimal typing for numpy 1.21. As always I am by no means a typing specialist.

• do not change user-specified units
• always format infered units as 'YYYY-mmmm-ddTHH:MM:SS'

This is my naïve approach.

• [ ] Closes #1449
• [ ] Tests added / passed
• [ ] Passes git diff upstream/master | flake8 --diff
• [ ] Fully documented, including whats-new.rst for all changes and api.rst for new API

Silences a number of warnings that accumulated in our test suite (c.f. #3266). The changes are mostly unrelated but small.

matplotlib warns: Calling gca() with keyword arguments was deprecated in Matplotlib 3.4. Starting two minor releases later, gca() will take no keyword arguments. The gca() function should only be used to get the current axes, or if no axes exist, create new axes with default keyword arguments. To create a new axes with non-default arguments, use plt.axes() or plt.subplot().

This only uses plt.gca() if there are active axes, else it calls plt.axes(**kwargs). Note that this can silently ignore some arguments. However, that this is already the case.

• [x] Closes #5444

This was changed in dask/dask#7669. Looks like they did not deprecate this behavior (i.e. passing both meta and output_dtypes). I'd suggest to follow dask's example here and not add a deprecation cycle. Thoughts?

• [x] Closes #4693

• [x] Passes pre-commit run --all-files
• [x] User visible changes (including notable bug fixes) are documented in whats-new.rst

Adds error codes to all # type: ignore comments. mypy should raise if a different type of error is encountered. Also enable showing the error code for typing errors. See: mypy: displaying-error-codes. Also remove some # type: ignore that are no longer necessary.

• [x] Part of #5077
• [x] Tests added
• [x] Passes pre-commit run --all-files

• [x] Closes #2007
• [ ] Tests added (for all bug fixes or enhancements)
• [ ] Tests passed (for all non-documentation changes)
• [ ] Fully documented, including whats-new.rst for all changes and api.rst for new API

Ok, this was easier to do than initially thought, we can use np.pad(a, pads, mode='wrap') in nputils.rolling_window. However, I'm not sure if that is enough already<sup>*</sup>.

I added an initial test, but could use some pointers where else you want this to be tested.

Questions: * is fill_value='periodic' a good api? * should the fill_value keyvalue be ported to rolling? * should this also be mentioned in the docs for rolling (I only learned about rolling.construct yesterday)

<sup>*</sup>rolling is present incore/dataset.py, core/dataarray.py, core/variable.py, core/rolling.py, core/dask_array_ops.py, core/nputils.py, core/ops.py, core/common.py, core/missing.py, and core/duck_array_ops.py that can be a bit daunting...

• [x] Closes #4970

The clean thing here would be to update min_deps_check.py so it works properly for this case. I am not sure it's really worth it.

• xref #4945

Try to limit the time of "CI Upstream" to avoid a silent failure.

numpy v1.20.0 introduced type hints which leads to some mypy errors in xarray. This is the minimum set of changes to make mypy happy again. I tried to avoid #type: ignore and Any but in some instances I think it was not worth it.

I am sure there is much more fun to be had with numpy typing ;-)

• [x] Closes #4881
• [x] Tests added
• [x] Passes pre-commit run --all-files
• [ ] User visible changes (including notable bug fixes) are documented in whats-new.rst

I only added one run for py3.8 latest. To be entirely sure we could also check the typing py37-min-all-deps but it feels like overkill...

Ok, looks good - the failure is expected - see #4878.

Small doc fix, see http://xarray.pydata.org/en/stable/io.html#formats-supported-by-pynio (the ..note:: did not get picked up)

• [ ] see https://github.com/pydata/xarray/pull/4760#issuecomment-774101639
• [x] Tests added
• [x] Passes pre-commit run --all-files
• [ ] User visible changes (including notable bug fixes) are documented in whats-new.rst

mypy v0.800 was ignoring our config (https://github.com/pydata/xarray/pull/4874#issuecomment-774771523)... Adding an (empty) [mypy] section to setup.cfg seems to do the trick, this is in contrast to the documentation and may get changed again (c.f. python/mypy#9940), but it does not hurt either.

@keewis

• [X] Closes #422
• [X] Tests added
• [X] Fully documented, including whats-new.rst for all changes and api.rst for new API

I took a shot at the weighted function - I added a DataArrayWeighted class, that currently only implements mean. So, there is still quite a bit missing (e.g. DatasetWeighted), but let me know what you think.

``` python import numpy as np import xarray as xr da = xr.DataArray([1, 2]) weights = xr.DataArray([4, 6]) da.weighted(weights).mean()

<xarray.DataArray ()>

array(1.6)

```

There are quite a number of difficult edge cases with invalid data, that can be discussed.

• I decided to replace all NaN in the weights with 0.
• if weights sum to 0 it returns NaN (and not inf) python weights = xr.DataArray([0, 0]) da.weighted(weights).mean()
• The following returns NaN (could be 1) python da = xr.DataArray([1, np.nan]) weights = xr.DataArray([1, 0]) da.weighted(weights).mean(skipna=False)

It could be good to add all edge-case logic to a separate function but I am not sure if this is possible...

Small improvement to weighted - more accurate typing and using _to_temp_dataset & map. No user facing changes.

• [x] Closes #4631
• [x] Tests added
• [x] Passes pre-commit run --all-files
• [x] User visible changes (including notable bug fixes) are documented in whats-new.rst

cc @gerritholl, @rabernat

<sub>

Overriding CI behaviors

By default, the upstream dev CI is disabled on pull request and push events. You can override this behavior per commit by adding a [test-upstream] tag to the first line of the commit message. For documentation-only commits, you can skip the CI per commit by adding a [skip-ci] tag to the first line of the commit message </sub>

• [x] Closes #2658, closes #4543
• [x] Tests added
• [x] Passes isort . && black . && mypy . && flake8
• [x] User visible changes (including notable bug fixes) are documented in whats-new.rst

pd.Index("a") has dtype object. Therefore string coords change their dtype on certain operations - e.g. align, __setitem__ (& assign), IndexVariable.concat. This can be avoided by using the coords instead of the index in some cases but in two instances it was unavoidable to cast a pd.Index back to a np.array. I probably did not catch all of these conversions.

What I am not sure: does this somehow contradict the index refactor?

• [x] Part of #3266
• [x] Passes isort . && black . && mypy . && flake8

Supresses 3 warnings we get in the test suite. Two are trivial. The third is PytestUnraisableExceptionWarning: Exception ignored in: CachingFileManager.__del__:

• The test triggers an error is in CachingFileManager.__init__:

https://github.com/pydata/xarray/blob/31d540f9d668fc5f8c1c92165f950c568778db01/xarray/tests/test_backends_file_manager.py#L205-L207

• It then tries to run __del__ but there is an error because an attribute is not available (as __init__ failed)

https://github.com/pydata/xarray/blob/31d540f9d668fc5f8c1c92165f950c568778db01/xarray/backends/file_manager.py#L238

• However, errors in __del__ are ignored, therefore we get a PytestUnraisableExceptionWarning

I decided to suppress the error and not fix CachingFileManager as I think that's nothing user-facing...

• [x] Closes #3263, potentially part of #4671

It's worth a try.

4694 was just merged a bit too soon. Unfortunately using mamba to uninstall the packages for upstream dev did not work - it uninstalled too much, thus switch back to conda. I'll just merge right away.

@dcherian

Closes #4671

ok let's start with the simplest one

• [x] switch to the newest windows image (does not help)
• [x] use mamba to resolve dependencies
• [x] use matplotlib-base
• [x] find the slowest tests using py.test xarray/tests/ --durations=100
• [ ] de-parametrize tests with a slow setup (if possible)
• [x] reduce the number of xfails (#4685)
• [ ] other ideas?

Follow up to #4506

Suppress some "Ambiguous reference date string" warnings. See e.g.: https://dev.azure.com/xarray/xarray/_build/results?buildId=4233&view=logs&j=2280efed-fda1-53bd-9213-1fa8ec9b4fa8&t=175181ee-1928-5a6b-f537-168f7a8b7c2d&l=361

rasterio slightly changed an error message, so it's no longer ignored. Also switched to pytest.mark.filterwarnings. This makes the diff larger but saves a level of indendation.

https://github.com/pydata/xarray/runs/1508035460#step:6:322

sparse is removed as conda package but not re-installed (https://github.com/pydata/xarray/blob/9802411b35291a6149d850e8e573cde71a93bfbf/ci/install-upstream-wheels.sh)

I assume this is not intentional, @andersy005?

Our upstream-dev is failing again. See numpy/numpy#17885 and pypa/pip#9186 for the probable cause. So this will need to be fixed upstream first.

I feel I overdid it a bit with the parameterizations of the weighted tests. This brings the number of tests of test_weighted.py from about 1000 down to less than 200 without changing the coverage.

• [x] Closes #4487

Thanks for the pointer @keewis. Let's see if this works.

• [x] Towards #4513
• [x] Tests added
• [x] Passes isort . && black . && mypy . && flake8
• [x] User visible changes (including notable bug fixes) are documented in whats-new.rst

• [x] Closes #4571

• [x] Closes #4565
• [x] Passes isort . && black . && mypy . && flake8
• [x] User visible changes (including notable bug fixes) are documented in whats-new.rst

Pandas will change pd.Index.__or__ and pd.Index.__and__. Use pd.Index.union and pd.Index.intersection inseated. There should be no change in functionality.