Do you have an accidental binary file in the commit? https://github.com/pydata/xarray/pull/3126/files#diff-337006a573a2971594e829114137f91b

Okay, a couple changes to the original PR: we now perform a shallow copy of the returned index in DataWithCoords.get_index and Indexes.__getitem__. The latter is needed because both Dataset.indexes and DataArray.indexes returns an Indexes object. Finally, I added a bunch of tests for various index accessors.

This seems pretty robust:

```python

import pandas as pd import xarray as xr index = pd.Index(list('abcd'), name='foo') series = pd.Series(range(4), index) array = xr.DataArray.from_series(series) array.get_index('foo').name = 'bar' array.get_index('foo').name 'foo' array['foo'].to_index().name = 'bar' array['foo'].to_index().name 'foo' array.to_index().name = 'bar' array.to_index().name 'foo' array.indexes['foo'].name = 'bar' array.indexes['foo'].name 'foo' ```

Since a Dataset is just a collection of DataArray objects, the behavior propagates nicely, e.g.

```python

dataset = xr.Dataset({'myvar': array}) dataset['myvar'].to_index().name = 'bar' dataset['myvar'].to_index().name 'foo' dataset.indexes['foo'].name = 'bar' dataset.indexes['foo'].name 'foo' ```

Do we want all views of indexes to be immutable or do we trust the user to not do stuff like this:

Is there a reasonable way of making that immutable? Given that the index allows its name to be mutated, I'm not sure.

(we can also separate that issue from this change, which is already 👍 )

Hello @gwgundersen! 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! :beers:

Comment last updated at 2019-07-16 22:33:30 UTC

Codecov Report

Merging #3126 into master will decrease coverage by 0.91%. The diff coverage is 100%.

```diff @@ Coverage Diff @@

master #3126 +/-

========================================== - Coverage 95.99% 95.08% -0.92%
========================================== Files 63 63
Lines 12799 12801 +2
========================================== - Hits 12287 12172 -115
- Misses 512 629 +117 ```

Locally, I've moved my changes from DataArray.to_series to DataWithCoords.get_index, and my tests still pass. After work, I'll write some more tests, e.g. for for DataArray.to_index, and push.

@shoyer, it seems like this issue happens anywhere the user can modify a Pandas index. For example, the Dataset bug in my comment yesterday. Do we want all views of indexes to be immutable or do we trust the user to not do stuff like this:

```python

dataset.indexes['mutable?'].name = 'yes' dataset.indexes['mutable?'].name 'yes' ```

It looks like to_series calls to_index, which calls get_index,

Thanks for finding that. That sounds right - if we make these changes there then I think we've solved the problem at its root and we should fix all these cases?

@max-sixty, name is just a mutable property on a Pandas index:

```python

dates = pd.date_range('01-Jan-2019', '31-Jan-2019', name='mutable?') series = pd.Series(np.random.randn(dates.size), dates) series.index.name 'mutable?' series.index.name = 'yes' series.index.name 'yes' ```

But you're right that to_series is the wrong place. It looks like to_series calls to_index, which calls get_index, which is inherited by other classes such as Dataset. For example, the bug persists with Dataset despite my fix:

```python

dates = pd.date_range('01-Jan-2019', '31-Jan-2019', name='mutable?') series = pd.Series(np.random.randn(dates.size), dates) dataset = xr.Dataset({'foo': series}) dataset.indexes['mutable?'].name = 'yes' dataset.indexes['mutable?'].name 'yes' ```

Does anyone know if get_index is the best place for this? Maybe the tests should be somewhere more generic as well.

I don't quite understand why the previous constructor was returning a mutable view, but the tests look great.

@gwgundersen do you know if we should instead be changing the to_index method which the to_series calls?

Any other comments from anyone?