This release brings a few important performance improvements, a wide range of usability upgrades, lots of bug fixes, and some new features. These include a plugin API to add backend engines, a new theme for the documentation, curve fitting methods, and several new plotting functions.

Many thanks to the 38 contributors to this release: Aaron Spring, Alessandro Amici, Alex Marandon, Alistair Miles, Ana Paula Krelling, Anderson Banihirwe, Aureliana Barghini, Baudouin Raoult, Benoit Bovy, Blair Bonnett, David Trémouilles, Deepak Cherian, Gabriel Medeiros Abrahão, Giacomo Caria, Hauke Schulz, Illviljan, Mathias Hauser, Matthias Bussonnier, Mattia Almansi, Maximilian Roos, Ray Bell, Richard Kleijn, Ryan Abernathey, Sam Levang, Spencer Clark, Spencer Jones, Tammas Loughran, Tobias Kölling, Todd, Tom Nicholas, Tom White, Victor Negîrneac, Xianxiang Li, Zeb Nicholls, crusaderky, dschwoerer, johnomotani, keewis

New Features

• apply combine_attrs on data variables and coordinate variables when concatenating and merging datasets and dataarrays (4902). By Justus Magin.

• Add Dataset.to_pandas (5247) By Giacomo Caria.

• Add DataArray.plot.surface which wraps matplotlib's plot_surface to make surface plots (2235 5084 5101). By John Omotani.

• Allow passing multiple arrays to Dataset.__setitem__ (5216). By Giacomo Caria.

• Add 'cumulative' option to Dataset.integrate and DataArray.integrate so that result is a cumulative integral, like scipy.integrate.cumulative_trapezoidal (5153). By John Omotani.

• Add safe_chunks option to Dataset.to_zarr which allows overriding checks made to ensure Dask and Zarr chunk compatibility (5056). By Ryan Abernathey

• Add Dataset.query and DataArray.query which enable indexing of datasets and data arrays by evaluating query expressions against the values of the data variables (4984). By Alistair Miles.

• Allow passing combine_attrs to Dataset.merge (4895). By Justus Magin.

• Support for dask.graph_manipulation (requires dask >=2021.3) By Guido Imperiale

• Add Dataset.plot.streamplot for streamplot plots with Dataset variables (5003). By John Omotani.

• Many of the arguments for the DataArray.str methods now support providing an array-like input. In this case, the array provided to the arguments is broadcast against the original array and applied elementwise.

• DataArray.str now supports +, *, and % operators. These behave the same as they do for str, except that they follow array broadcasting rules.

• A large number of new DataArray.str methods were implemented, DataArray.str.casefold, DataArray.str.cat, DataArray.str.extract, DataArray.str.extractall, DataArray.str.findall, DataArray.str.format, DataArray.str.get_dummies, DataArray.str.islower, DataArray.str.join, DataArray.str.normalize, DataArray.str.partition, DataArray.str.rpartition, DataArray.str.rsplit, and DataArray.str.split. A number of these methods allow for splitting or joining the strings in an array. (4622) By Todd Jennings

• Thanks to the new pluggable backend infrastructure external packages may now use the xarray.backends entry point to register additional engines to be used in open_dataset, see the documentation in add-a-backend (4309, 4803, 4989, 4810 and many others). The backend refactor has been sponsored with the "Essential Open Source Software for Science" grant from the Chan Zuckerberg Initiative and developed by B-Open. By Aureliana Barghini and Alessandro Amici.

• ~core.accessor_dt.DatetimeAccessor.date added (4983, 4994). By Hauke Schulz.

• Implement __getitem__ for both ~core.groupby.DatasetGroupBy and ~core.groupby.DataArrayGroupBy, inspired by pandas' ~pandas.core.groupby.GroupBy.get_group. By Deepak Cherian.

• Switch the tutorial functions to use pooch (which is now a optional dependency) and add tutorial.open_rasterio as a way to open example rasterio files (3986, 4102, 5074). By Justus Magin.

• Add typing information to unary and binary arithmetic operators operating on Dataset, DataArray, Variable, ~core.groupby.DatasetGroupBy or ~core.groupby.DataArrayGroupBy (4904). By Richard Kleijn.

• Add a combine_attrs parameter to open_mfdataset (4971). By Justus Magin.

• Enable passing arrays with a subset of dimensions to DataArray.clip & Dataset.clip; these methods now use xarray.apply_ufunc; (5184). By Maximilian Roos.

• Disable the cfgrib backend if the eccodes library is not installed (5083). By Baudouin Raoult.

• Added DataArray.curvefit and Dataset.curvefit for general curve fitting applications. (4300, 4849) By Sam Levang.

• Add options to control expand/collapse of sections in display of Dataset and DataArray. The function set_options now takes keyword arguments display_expand_attrs, display_expand_coords, display_expand_data, display_expand_data_vars, all of which can be one of True to always expand, False to always collapse, or default to expand unless over a pre-defined limit (5126). By Tom White.

• Significant speedups in Dataset.interp and DataArray.interp. (4739, 4740). By Deepak Cherian.

• Prevent passing concat_dim to xarray.open_mfdataset when combine='by_coords' is specified, which should never have been possible (as xarray.combine_by_coords has no concat_dim argument to pass to). Also removes unneeded internal reordering of datasets in xarray.open_mfdataset when combine='by_coords' is specified. Fixes (5230). By Tom Nicholas.

• Implement __setitem__ for xarray.core.indexing.DaskIndexingAdapter if dask version supports item assignment. (5171, 5174) By Tammas Loughran.

Breaking changes

• The minimum versions of some dependencies were changed:

  Package	Old	New
  boto3	1.12	1.13
  cftime	1.0	1.1
  dask	2.11	2.15
  distributed	2.11	2.15
  matplotlib	3.1	3.2
  numba	0.48	0.49

• open_dataset and open_dataarray now accept only the first argument as positional, all others need to be passed are keyword arguments. This is part of the refactor to support external backends (4309, 4989). By Alessandro Amici.

• Functions that are identities for 0d data return the unchanged data if axis is empty. This ensures that Datasets where some variables do not have the averaged dimensions are not accidentally changed (4885, 5207). By David Schwörer.

• DataArray.coarsen and Dataset.coarsen no longer support passing keep_attrs via its constructor. Pass keep_attrs via the applied function, i.e. use ds.coarsen(...).mean(keep_attrs=False) instead of ds.coarsen(..., keep_attrs=False).mean(). Further, coarsen now keeps attributes per default (5227). By Mathias Hauser.

• switch the default of the merge combine_attrs parameter to "override". This will keep the current behavior for merging the attrs of variables but stop dropping the attrs of the main objects (4902). By Justus Magin.

Deprecations

• Warn when passing concat_dim to xarray.open_mfdataset when combine='by_coords' is specified, which should never have been possible (as xarray.combine_by_coords has no concat_dim argument to pass to). Also removes unneeded internal reordering of datasets in xarray.open_mfdataset when combine='by_coords' is specified. Fixes (5230), via (5231, 5255). By Tom Nicholas.

• The lock keyword argument to open_dataset and open_dataarray is now a backend specific option. It will give a warning if passed to a backend that doesn't support it instead of being silently ignored. From the next version it will raise an error. This is part of the refactor to support external backends (5073). By Tom Nicholas and Alessandro Amici.

Bug fixes

• Properly support DataArray.ffill, DataArray.bfill, Dataset.ffill, Dataset.bfill along chunked dimensions. (2699). By Deepak Cherian.

• Fix 2d plot failure for certain combinations of dimensions when x is 1d and y is 2d (5097, 5099). By John Omotani.

• Ensure standard calendar times encoded with large values (i.e. greater than approximately 292 years), can be decoded correctly without silently overflowing (5050). This was a regression in xarray 0.17.0. By Zeb Nicholls.

• Added support for numpy.bool_ attributes in roundtrips using h5netcdf engine with invalid_netcdf=True [which casts bool s to numpy.bool_] (4981, 4986). By Victor Negîrneac.

• Don't allow passing axis to Dataset.reduce methods (3510, 4940). By Justus Magin.

• Decode values as signed if attribute _Unsigned = "false" (4954) By Tobias Kölling.

• Keep coords attributes when interpolating when the indexer is not a Variable. (4239, 4839 5031) By Jimmy Westling.

• Ensure standard calendar dates encoded with a calendar attribute with some or all uppercase letters can be decoded or encoded to or from np.datetime64[ns] dates with or without cftime installed (5093, 5180). By Spencer Clark.

• Warn on passing keep_attrs to resample and rolling_exp as they are ignored, pass keep_attrs to the applied function instead (5265). By Mathias Hauser.

Documentation

• New section on add-a-backend in the "Internals" chapter aimed to backend developers (4803, 4810). By Aureliana Barghini.

• Add Dataset.polyfit and DataArray.polyfit under "See also" in the docstrings of Dataset.polyfit and DataArray.polyfit (5016, 5020). By Aaron Spring.

• New sphinx theme & rearrangement of the docs (4835). By Anderson Banihirwe.

Internal Changes

• Enable displaying mypy error codes and ignore only specific error codes using # type: ignore[error-code] (5096). By Mathias Hauser.

• Replace uses of raises_regex with the more standard pytest.raises(Exception, match="foo"); (5188), (5191). By Maximilian Roos.

This release brings a few important performance improvements, a wide range of usability upgrades, lots of bug fixes, and some new features. These include better cftime support, a new quiver plot, better unstack performance, more efficient memory use in rolling operations, and some python packaging improvements. We also have a few documentation improvements (and more planned!).

Many thanks to the 36 contributors to this release: Alessandro Amici, Anderson Banihirwe, Aureliana Barghini, Ayrton Bourn, Benjamin Bean, Blair Bonnett, Chun Ho Chow, DWesl, Daniel Mesejo-León, Deepak Cherian, Eric Keenan, Illviljan, Jens Hedegaard Nielsen, Jody Klymak, Julien Seguinot, Julius Busecke, Kai Mühlbauer, Leif Denby, Martin Durant, Mathias Hauser, Maximilian Roos, Michael Mann, Ray Bell, RichardScottOZ, Spencer Clark, Tim Gates, Tom Nicholas, Yunus Sevinchan, alexamici, aurghs, crusaderky, dcherian, ghislainp, keewis, rhkleijn

Breaking changes

• xarray no longer supports python 3.6

  The minimum version policy was changed to also apply to projects with irregular releases. As a result, the minimum versions of some dependencies have changed:

  Package	Old	New
  Python	3.6	3.7
  setuptools	38.4	40.4
  numpy	1.15	1.17
  pandas	0.25	1.0
  dask	2.9	2.11
  distributed	2.9	2.11
  bottleneck	1.2	1.3
  h5netcdf	0.7	0.8
  iris	2.2	2.4
  netcdf4	1.4	1.5
  pseudonetcdf	3.0	3.1
  rasterio	1.0	1.1
  scipy	1.3	1.4
  seaborn	0.9	0.10
  zarr	2.3	2.4

  (4688, 4720, 4907, 4942)

• As a result of 4684 the default units encoding for datetime-like values (np.datetime64[ns] or cftime.datetime) will now always be set such that int64 values can be used. In the past, no units finer than "seconds" were chosen, which would sometimes mean that float64 values were required, which would lead to inaccurate I/O round-trips.

• Variables referred to in attributes like bounds and grid_mapping can be set as coordinate variables. These attributes are moved to DataArray.encoding from DataArray.attrs. This behaviour is controlled by the decode_coords kwarg to open_dataset and open_mfdataset. The full list of decoded attributes is in weather-climate (2844, 3689)

• As a result of 4911 the output from calling DataArray.sum or DataArray.prod on an integer array with skipna=True and a non-None value for min_count will now be a float array rather than an integer array.

Deprecations

• dim argument to DataArray.integrate is being deprecated in favour of a coord argument, for consistency with Dataset.integrate. For now using dim issues a FutureWarning. It will be removed in version 0.19.0 (3993). By Tom Nicholas.

• Deprecated autoclose kwargs from open_dataset are removed (4725). By Aureliana Barghini.

• the return value of Dataset.update is being deprecated to make it work more like dict.update. It will be removed in version 0.19.0 (4932). By Justus Magin.

New Features

• ~xarray.cftime_range and DataArray.resample now support millisecond ("L" or "ms") and microsecond ("U" or "us") frequencies for cftime.datetime coordinates (4097, 4758). By Spencer Clark.

• Significantly higher unstack performance on numpy-backed arrays which contain missing values; 8x faster than previous versions in our benchmark, and now 2x faster than pandas (4746). By Maximilian Roos.

• Add Dataset.plot.quiver for quiver plots with Dataset variables. By Deepak Cherian.

• Add "drop_conflicts" to the strategies supported by the combine_attrs kwarg (4749, 4827). By Justus Magin.

• Allow installing from git archives (4897). By Justus Magin.

• ~computation.rolling.DataArrayCoarsen and ~computation.rolling.DatasetCoarsen now implement a reduce method, enabling coarsening operations with custom reduction functions (3741, 4939). By Spencer Clark.

• Most rolling operations use significantly less memory. (4325). By Deepak Cherian.

• Add Dataset.drop_isel and DataArray.drop_isel (4658, 4819). By Daniel Mesejo.

• Xarray now leverages updates as of cftime version 1.4.1, which enable exact I/O roundtripping of cftime.datetime objects (4758). By Spencer Clark.

• open_dataset and open_mfdataset now accept fsspec URLs (including globs for the latter) for engine="zarr", and so allow reading from many remote and other file systems (4461) By Martin Durant

• DataArray.swap_dims & Dataset.swap_dims now accept dims in the form of kwargs as well as a dict, like most similar methods. By Maximilian Roos.

Bug fixes

• Use specific type checks in xarray.core.variable.as_compatible_data instead of blanket access to values attribute (2097) By Yunus Sevinchan.

• DataArray.resample and Dataset.resample do not trigger computations anymore if Dataset.weighted or DataArray.weighted are applied (4625, 4668). By Julius Busecke.

• merge with combine_attrs='override' makes a copy of the attrs (4627).

• By default, when possible, xarray will now always use values of type int64 when encoding and decoding numpy.datetime64[ns] datetimes. This ensures that maximum precision and accuracy are maintained in the round-tripping process (4045, 4684). It also enables encoding and decoding standard calendar dates with time units of nanoseconds (4400). By Spencer Clark and Mark Harfouche.

• DataArray.astype, Dataset.astype and Variable.astype support the order and subok parameters again. This fixes a regression introduced in version 0.16.1 (4644, 4683). By Richard Kleijn .

• Remove dictionary unpacking when using .loc to avoid collision with .sel parameters (4695). By Anderson Banihirwe.

• Fix the legend created by Dataset.plot.scatter (4641, 4723). By Justus Magin.

• Fix a crash in orthogonal indexing on geographic coordinates with engine='cfgrib' (4733 4737). By Alessandro Amici.

• Coordinates with dtype str or bytes now retain their dtype on many operations, e.g. reindex, align, concat, assign, previously they were cast to an object dtype (2658 and 4543). By Mathias Hauser.

• Limit number of data rows when printing large datasets. (4736, 4750). By Jimmy Westling.

• Add missing_dims parameter to transpose (4647, 4767). By Daniel Mesejo.

• Resolve intervals before appending other metadata to labels when plotting (4322, 4794). By Justus Magin.

• Fix regression when decoding a variable with a scale_factor and add_offset given as a list of length one (4631). By Mathias Hauser.

• Expand user directory paths (e.g. ~/) in open_mfdataset and Dataset.to_zarr (4783, 4795). By Julien Seguinot.

• Raise DeprecationWarning when trying to typecast a tuple containing a DataArray. User now prompted to first call .data on it (4483). By Chun Ho Chow.

• Ensure that Dataset.interp raises ValueError when interpolating outside coordinate range and bounds_error=True (4854, 4855). By Leif Denby.

• Fix time encoding bug associated with using cftime versions greater than 1.4.0 with xarray (4870, 4871). By Spencer Clark.

• Stop DataArray.sum and DataArray.prod computing lazy arrays when called with a min_count parameter (4898, 4911). By Blair Bonnett.

• Fix bug preventing the min_count parameter to DataArray.sum and DataArray.prod working correctly when calculating over all axes of a float64 array (4898, 4911). By Blair Bonnett.

• Fix decoding of vlen strings using h5py versions greater than 3.0.0 with h5netcdf backend (4570, 4893). By Kai Mühlbauer.

• Allow converting Dataset or DataArray objects with a MultiIndex and at least one other dimension to a pandas object (3008, 4442). By ghislainp.

Documentation

• Add information about requirements for accessor classes (2788, 4657). By Justus Magin.

• Start a list of external I/O integrating with xarray (683, 4566). By Justus Magin.

• Add concat examples and improve combining documentation (4620, 4645). By Ray Bell and Justus Magin.

• explicitly mention that Dataset.update updates inplace (2951, 4932). By Justus Magin.

• Added docs on vectorized indexing (4711). By Eric Keenan.

Internal Changes

• Speed up of the continuous integration tests on azure.

  • Switched to mamba and use matplotlib-base for a faster installation of all dependencies (4672).

  • Use pytest.mark.skip instead of pytest.mark.xfail for some tests that can currently not succeed (4685).

  • Run the tests in parallel using pytest-xdist (4694).

  By Justus Magin and Mathias Hauser.

• Use pyproject.toml instead of the setup_requires option for setuptools (4897). By Justus Magin.

• Replace all usages of assert x.identical(y) with assert_identical(x, y) for clearer error messages (4752). By Maximilian Roos.

• Speed up attribute style access (e.g. ds.somevar instead of ds["somevar"]) and tab completion in IPython (4741, 4742). By Richard Kleijn.

• Added the set_close method to Dataset and DataArray for backends to specify how to voluntary release all resources. (#4809) By Alessandro Amici.

• Update type hints to work with numpy v1.20 (4878). By Mathias Hauser.

• Ensure warnings cannot be turned into exceptions in testing.assert_equal and the other assert_* functions (4864). By Mathias Hauser.

• Performance improvement when constructing DataArrays. Significantly speeds up repr for Datasets with large number of variables. By Deepak Cherian.

This release brings the ability to write to limited regions of zarr files, open zarr files with open_dataset and open_mfdataset, increased support for propagating attrs using the keep_attrs flag, as well as numerous bugfixes and documentation improvements.

Many thanks to the 31 contributors who contributed to this release: Aaron Spring, Akio Taniguchi, Aleksandar Jelenak, alexamici, Alexandre Poux, Anderson Banihirwe, Andrew Pauling, Ashwin Vishnu, aurghs, Brian Ward, Caleb, crusaderky, Dan Nowacki, darikg, David Brochart, David Huard, Deepak Cherian, Dion Häfner, Gerardo Rivera, Gerrit Holl, Illviljan, inakleinbottle, Jacob Tomlinson, James A. Bednar, jenssss, Joe Hamman, johnomotani, Joris Van den Bossche, Julia Kent, Julius Busecke, Kai Mühlbauer, keewis, Keisuke Fujii, Kyle Cranmer, Luke Volpatti, Mathias Hauser, Maximilian Roos, Michaël Defferrard, Michal Baumgartner, Nick R. Papior, Pascal Bourgault, Peter Hausamann, PGijsbers, Ray Bell, Romain Martinez, rpgoldman, Russell Manser, Sahid Velji, Samnan Rahee, Sander, Spencer Clark, Stephan Hoyer, Thomas Zilio, Tobias Kölling, Tom Augspurger, Wei Ji, Yash Saboo, Zeb Nicholls,

Deprecations

• ~core.accessor_dt.DatetimeAccessor.weekofyear and ~core.accessor_dt.DatetimeAccessor.week have been deprecated. Use DataArray.dt.isocalendar().week instead (4534). By Mathias Hauser. Maximilian Roos, and Spencer Clark.

• DataArray.rolling and Dataset.rolling no longer support passing keep_attrs via its constructor. Pass keep_attrs via the applied function, i.e. use ds.rolling(...).mean(keep_attrs=False) instead of ds.rolling(..., keep_attrs=False).mean() Rolling operations now keep their attributes per default (4510). By Mathias Hauser.

New Features

• open_dataset and open_mfdataset now works with engine="zarr" (3668, 4003, 4187). By Miguel Jimenez and Wei Ji Leong.

• Unary & binary operations follow the keep_attrs flag (3490, 4065, 3433, 3595, 4195). By Deepak Cherian.

• Added ~core.accessor_dt.DatetimeAccessor.isocalendar() that returns a Dataset with year, week, and weekday calculated according to the ISO 8601 calendar. Requires pandas version 1.1.0 or greater (4534). By Mathias Hauser, Maximilian Roos, and Spencer Clark.

• Dataset.to_zarr now supports a region keyword for writing to limited regions of existing Zarr stores (4035). See io.zarr.appending for full details. By Stephan Hoyer.

• Added typehints in align to reflect that the same type received in objects arg will be returned (4522). By Michal Baumgartner.

• Dataset.weighted and DataArray.weighted are now executing value checks lazily if weights are provided as dask arrays (4541, 4559). By Julius Busecke.

• Added the keep_attrs keyword to rolling_exp.mean(); it now keeps attributes per default. By Mathias Hauser (4592).

• Added freq as property to CFTimeIndex and into the CFTimeIndex.repr. (2416, 4597) By Aaron Spring.

Bug fixes

• Fix bug where reference times without padded years (e.g. since 1-1-1) would lose their units when being passed by encode_cf_datetime (4422, 4506). Such units are ambiguous about which digit represents the years (is it YMD or DMY?). Now, if such formatting is encountered, it is assumed that the first digit is the years, they are padded appropriately (to e.g. since 0001-1-1) and a warning that this assumption is being made is issued. Previously, without cftime, such times would be silently parsed incorrectly (at least based on the CF conventions) e.g. "since 1-1-1" would be parsed (via pandas and dateutil) to since 2001-1-1. By Zeb Nicholls.

• Fix DataArray.plot.step. By Deepak Cherian.

• Fix bug where reading a scalar value from a NetCDF file opened with the h5netcdf backend would raise a ValueError when decode_cf=True (4471, 4485). By Gerrit Holl.

• Fix bug where datetime64 times are silently changed to incorrect values if they are outside the valid date range for ns precision when provided in some other units (4427, 4454). By Andrew Pauling

• Fix silently overwriting the engine key when passing open_dataset a file object to an incompatible netCDF (4457). Now incompatible combinations of files and engines raise an exception instead. By Alessandro Amici.

• The min_count argument to DataArray.sum() and DataArray.prod() is now ignored when not applicable, i.e. when skipna=False or when skipna=None and the dtype does not have a missing value (4352). By Mathias Hauser.

• combine_by_coords now raises an informative error when passing coordinates with differing calendars (4495). By Mathias Hauser.

• DataArray.rolling and Dataset.rolling now also keep the attributes and names of of (wrapped) DataArray objects, previously only the global attributes were retained (4497, 4510). By Mathias Hauser.

• Improve performance where reading small slices from huge dimensions was slower than necessary (4560). By Dion Häfner.

• Fix bug where dask_gufunc_kwargs was silently changed in apply_ufunc (4576). By Kai Mühlbauer.

Documentation

• document the API not supported with duck arrays (4530). By Justus Magin.

• Mention the possibility to pass functions to Dataset.where or DataArray.where in the parameter documentation (4223, 4613). By Justus Magin.

• Update the docstring of DataArray and Dataset. (4532); By Jimmy Westling.

• Raise a more informative error when DataArray.to_dataframe is is called on a scalar, (4228); By Pieter Gijsbers.

• Fix grammar and typos in the contributing guide (4545). By Sahid Velji.

• Fix grammar and typos in the user-guide/io guide (4553). By Sahid Velji.

• Update link to NumPy docstring standard in the contributing guide (4558). By Sahid Velji.

• Add docstrings to isnull and notnull, and fix the displayed signature (2760, 4618). By Justus Magin.

Internal Changes

• Optional dependencies can be installed along with xarray by specifying extras as pip install "xarray[extra]" where extra can be one of io, accel, parallel, viz and complete. See docs for updated installation instructions. (2888, 4480). By Ashwin Vishnu, Justus Magin and Mathias Hauser.

• Removed stray spaces that stem from black removing new lines (4504). By Mathias Hauser.

• Ensure tests are not skipped in the py38-all-but-dask test environment (4509). By Mathias Hauser.

• Ignore select numpy warnings around missing values, where xarray handles the values appropriately, (4536); By Maximilian Roos.

• Replace the internal use of pd.Index.__or__ and pd.Index.__and__ with pd.Index.union and pd.Index.intersection as they will stop working as set operations in the future (4565). By Mathias Hauser.

• Add GitHub action for running nightly tests against upstream dependencies (4583). By Anderson Banihirwe.

• Ensure all figures are closed properly in plot tests (4600). By Yash Saboo, Nirupam K N and Mathias Hauser.

This patch release fixes an incompatibility with a recent pandas change, which was causing an issue indexing with a datetime64. It also includes improvements to rolling, to_dataframe, cov & corr methods and bug fixes. Our documentation has a number of improvements, including fixing all doctests and confirming their accuracy on every commit.

Many thanks to the 36 contributors who contributed to this release:

Aaron Spring, Akio Taniguchi, Aleksandar Jelenak, Alexandre Poux, Caleb, Dan Nowacki, Deepak Cherian, Gerardo Rivera, Jacob Tomlinson, James A. Bednar, Joe Hamman, Julia Kent, Kai Mühlbauer, Keisuke Fujii, Mathias Hauser, Maximilian Roos, Nick R. Papior, Pascal Bourgault, Peter Hausamann, Romain Martinez, Russell Manser, Samnan Rahee, Sander, Spencer Clark, Stephan Hoyer, Thomas Zilio, Tobias Kölling, Tom Augspurger, alexamici, crusaderky, darikg, inakleinbottle, jenssss, johnomotani, keewis, and rpgoldman.

Breaking changes

• DataArray.astype and Dataset.astype now preserve attributes. Keep the old behavior by passing keep_attrs=False (2049, 4314). By Dan Nowacki and Gabriel Joel Mitchell.

New Features

• ~xarray.DataArray.rolling and ~xarray.Dataset.rolling now accept more than 1 dimension. (4219) By Keisuke Fujii.

• ~xarray.DataArray.to_dataframe and ~xarray.Dataset.to_dataframe now accept a dim_order parameter allowing to specify the resulting dataframe's dimensions order (4331, 4333). By Thomas Zilio.

• Support multiple outputs in xarray.apply_ufunc when using dask='parallelized'. (1815, 4060). By Kai Mühlbauer.

• min_count can be supplied to reductions such as .sum when specifying multiple dimension to reduce over; (4356). By Maximilian Roos.

• xarray.cov and xarray.corr now handle missing values; (4351). By Maximilian Roos.

• Add support for parsing datetime strings formatted following the default string representation of cftime objects, i.e. YYYY-MM-DD hh:mm:ss, in partial datetime string indexing, as well as ~xarray.cftime_range (4337). By Spencer Clark.

• Build CFTimeIndex.__repr__ explicitly as pandas.Index. Add calendar as a new property for CFTimeIndex and show calendar and length in CFTimeIndex.__repr__ (2416, 4092) By Aaron Spring.

• Use a wrapped array's _repr_inline_ method to construct the collapsed repr of DataArray and Dataset objects and document the new method in internals/index. (4248). By Justus Magin.

• Allow per-variable fill values in most functions. (4237). By Justus Magin.

• Expose use_cftime option in ~xarray.open_zarr (2886, 3229) By Samnan Rahee and Anderson Banihirwe.

Bug fixes

• Fix indexing with datetime64 scalars with pandas 1.1 (4283). By Stephan Hoyer and Justus Magin.

• Variables which are chunked using dask only along some dimensions can be chunked while storing with zarr along previously unchunked dimensions (4312) By Tobias Kölling.

• Fixed a bug in backend caused by basic installation of Dask (4164, 4318) Sam Morley.

• Fixed a few bugs with Dataset.polyfit when encountering deficient matrix ranks (4190, 4193). By Pascal Bourgault.

• Fixed inconsistencies between docstring and functionality for DataArray.str.get and DataArray.str.wrap (4334). By Mathias Hauser.

• Fixed overflow issue causing incorrect results in computing means of cftime.datetime arrays (4341). By Spencer Clark.

• Fixed Dataset.coarsen, DataArray.coarsen dropping attributes on original object (4120, 4360). By Julia Kent.

• fix the signature of the plot methods. (4359) By Justus Magin.

• Fix xarray.apply_ufunc with vectorize=True and exclude_dims (3890). By Mathias Hauser.

• Fix KeyError when doing linear interpolation to an nd DataArray that contains NaNs (4233). By Jens Svensmark

• Fix incorrect legend labels for Dataset.plot.scatter (4126). By Peter Hausamann.

• Fix dask.optimize on DataArray producing an invalid Dask task graph (3698) By Tom Augspurger

• Fix pip install . when no .git directory exists; namely when the xarray source directory has been rsync'ed by PyCharm Professional for a remote deployment over SSH. By Guido Imperiale

• Preserve dimension and coordinate order during xarray.concat (2811, 4072, 4419). By Kai Mühlbauer.

• Avoid relying on set objects for the ordering of the coordinates (4409) By Justus Magin.

Documentation

• Update the docstring of DataArray.copy to remove incorrect mention of 'dataset' (3606) By Sander van Rijn.

• Removed skipna argument from DataArray.count, DataArray.any, DataArray.all. (755) By Sander van Rijn

• Update the contributing guide to use merges instead of rebasing and state that we squash-merge. (4355). By Justus Magin.

• Make sure the examples from the docstrings actually work (4408). By Justus Magin.

• Updated Vectorized Indexing to a clearer example. By Maximilian Roos

Internal Changes

• Fixed all doctests and enabled their running in CI. By Justus Magin.

• Relaxed the mindeps_policy to support:

  • all versions of setuptools released in the last 42 months (but no older than 38.4)

  • all versions of dask and dask.distributed released in the last 12 months (but no older than 2.9)

  • all versions of other packages released in the last 12 months

  All are up from 6 months (4295) Guido Imperiale.

• Use dask.array.apply_gufunc instead of dask.array.blockwise in xarray.apply_ufunc when using dask='parallelized'. (4060, 4391, 4392) By Kai Mühlbauer.

• Align mypy versions to 0.782 across requirements and .pre-commit-config.yml files. (4390) By Maximilian Roos

• Only load resource files when running inside a Jupyter Notebook (4294) By Guido Imperiale

• Silenced most numpy warnings such as Mean of empty slice. (4369) By Maximilian Roos

• Enable type checking for concat (4238) By Mathias Hauser.

• Updated plot functions for matplotlib version 3.3 and silenced warnings in the plot tests (4365). By Mathias Hauser.

• Versions in pre-commit.yaml are now pinned, to reduce the chances of conflicting versions. (4388) By Maximilian Roos

This release adds xarray.cov & xarray.corr for covariance & correlation respectively; the idxmax & idxmin methods, the polyfit method & xarray.polyval for fitting polynomials, as well as a number of documentation improvements, other features, and bug fixes. Many thanks to all 44 contributors who contributed to this release:

Akio Taniguchi, Andrew Williams, Aurélien Ponte, Benoit Bovy, Dave Cole, David Brochart, Deepak Cherian, Elliott Sales de Andrade, Etienne Combrisson, Hossein Madadi, Huite, Joe Hamman, Kai Mühlbauer, Keisuke Fujii, Maik Riechert, Marek Jacob, Mathias Hauser, Matthieu Ancellin, Maximilian Roos, Noah D Brenowitz, Oriol Abril, Pascal Bourgault, Phillip Butcher, Prajjwal Nijhara, Ray Bell, Ryan Abernathey, Ryan May, Spencer Clark, Spencer Hill, Srijan Saurav, Stephan Hoyer, Taher Chegini, Todd, Tom Nicholas, Yohai Bar Sinai, Yunus Sevinchan, arabidopsis, aurghs, clausmichele, dmey, johnomotani, keewis, raphael dussin, risebell

Breaking changes

• Minimum supported versions for the following packages have changed: dask >=2.9, distributed>=2.9. By Deepak Cherian

• groupby operations will restore coord dimension order. Pass restore_coord_dims=False to revert to previous behavior.

• DataArray.transpose will now transpose coordinates by default. Pass transpose_coords=False to revert to previous behaviour. By Maximilian Roos

• Alternate draw styles for plot.step must be passed using the drawstyle (or ds) keyword argument, instead of the linestyle (or ls) keyword argument, in line with the upstream change in Matplotlib. (3274) By Elliott Sales de Andrade

• The old auto_combine function has now been removed in favour of the combine_by_coords and combine_nested functions. This also means that the default behaviour of open_mfdataset has changed to use combine='by_coords' as the default argument value. (2616, 3926) By Tom Nicholas.

• The DataArray and Variable HTML reprs now expand the data section by default (4176) By Stephan Hoyer.

New Features

• DataArray.argmin and DataArray.argmax now support sequences of 'dim' arguments, and if a sequence is passed return a dict (which can be passed to DataArray.isel to get the value of the minimum) of the indices for each dimension of the minimum or maximum of a DataArray. (3936) By John Omotani, thanks to Keisuke Fujii for work in 1469.

• Added xarray.cov and xarray.corr (3784, 3550, 4089). By Andrew Williams and Robin Beer.

• Implement DataArray.idxmax, DataArray.idxmin, Dataset.idxmax, Dataset.idxmin. (60, 3871) By Todd Jennings

• Added DataArray.polyfit and xarray.polyval for fitting polynomials. (3349, 3733, 4099) By Pascal Bourgault.

• Added xarray.infer_freq for extending frequency inferring to CFTime indexes and data (4033). By Pascal Bourgault.

• chunks='auto' is now supported in the chunks argument of Dataset.chunk. (4055) By Andrew Williams

• Control over attributes of result in merge, concat, combine_by_coords and combine_nested using combine_attrs keyword argument. (3865, 3877) By John Omotani

• missing_dims argument to Dataset.isel, DataArray.isel and Variable.isel to allow replacing the exception when a dimension passed to isel is not present with a warning, or just ignore the dimension. (3866, 3923) By John Omotani

• Support dask handling for DataArray.idxmax, DataArray.idxmin, Dataset.idxmax, Dataset.idxmin. (3922, 4135) By Kai Mühlbauer and Pascal Bourgault.

• More support for unit aware arrays with pint (3643, 3975, 4163) By Justus Magin.

• Support overriding existing variables in to_zarr() with mode='a' even without append_dim, as long as dimension sizes do not change. By Stephan Hoyer.

• Allow plotting of boolean arrays. (3766) By Marek Jacob

• Enable using MultiIndex levels as coordinates in 1D and 2D plots (3927). By Mathias Hauser.

• A days_in_month accessor for xarray.CFTimeIndex, analogous to the days_in_month accessor for a pandas.DatetimeIndex, which returns the days in the month each datetime in the index. Now days in month weights for both standard and non-standard calendars can be obtained using the ~core.accessor_dt.DatetimeAccessor (3935). This feature requires cftime version 1.1.0 or greater. By Spencer Clark.

• For the netCDF3 backend, added dtype coercions for unsigned integer types. (4014, 4018) By Yunus Sevinchan

• map_blocks now accepts a template kwarg. This allows use cases where the result of a computation could not be inferred automatically. By Deepak Cherian

• map_blocks can now handle dask-backed xarray objects in args. (3818) By Deepak Cherian

• Add keyword decode_timedelta to xarray.open_dataset, (xarray.open_dataarray, xarray.open_dataarray, xarray.decode_cf) that allows to disable/enable the decoding of timedeltas independently of time decoding (1621) Aureliana Barghini

Enhancements

• Performance improvement of DataArray.interp and Dataset.interp We performs independent interpolation sequentially rather than interpolating in one large multidimensional space. (2223) By Keisuke Fujii.

• DataArray.interp now support interpolations over chunked dimensions (4155). By Alexandre Poux.

• Major performance improvement for Dataset.from_dataframe when the dataframe has a MultiIndex (4184). By Stephan Hoyer. - DataArray.reset_index and Dataset.reset_index now keep coordinate attributes (4103). By Oriol Abril.

• Axes kwargs such as facecolor can now be passed to DataArray.plot in subplot_kws. This works for both single axes plots and FacetGrid plots. By Raphael Dussin.

• Array items with long string reprs are now limited to a reasonable width (3900) By Maximilian Roos

• Large arrays whose numpy reprs would have greater than 40 lines are now limited to a reasonable length. (3905) By Maximilian Roos

Bug fixes

• Fix errors combining attrs in open_mfdataset (4009, 4173) By John Omotani

• If groupby receives a DataArray with name=None, assign a default name (158) By Phil Butcher.

• Support dark mode in VS code (4024) By Keisuke Fujii.

• Fix bug when converting multiindexed pandas objects to sparse xarray objects. (4019) By Deepak Cherian.

• ValueError is raised when fill_value is not a scalar in full_like. (3977) By Huite Bootsma.

• Fix wrong order in converting a pd.Series with a MultiIndex to DataArray. (3951, 4186) By Keisuke Fujii and Stephan Hoyer.

• Fix renaming of coords when one or more stacked coords is not in sorted order during stack+groupby+apply operations. (3287, 3906) By Spencer Hill

• Fix a regression where deleting a coordinate from a copied DataArray can affect the original DataArray. (3899, 3871) By Todd Jennings

• Fix ~xarray.plot.FacetGrid plots with a single contour. (3569, 3915). By Deepak Cherian

• Use divergent colormap if levels spans 0. (3524) By Deepak Cherian

• Fix ~xarray.plot.FacetGrid when vmin == vmax. (3734) By Deepak Cherian

• Fix plotting when levels is a scalar and norm is provided. (3735) By Deepak Cherian

• Fix bug where plotting line plots with 2D coordinates depended on dimension order. (3933) By Tom Nicholas.

• Fix RasterioDeprecationWarning when using a vrt in open_rasterio. (3964) By Taher Chegini.

• Fix AttributeError on displaying a Variable in a notebook context. (3972, 3973) By Ian Castleden.

• Fix bug causing DataArray.interpolate_na to always drop attributes, and added keep_attrs argument. (3968) By Tom Nicholas.

• Fix bug in time parsing failing to fall back to cftime. This was causing time variables with a time unit of 'msecs' to fail to parse. (3998) By Ryan May.

• Fix weighted mean when passing boolean weights (4074). By Mathias Hauser.

• Fix html repr in untrusted notebooks: fallback to plain text repr. (4053) By Benoit Bovy.

• Fix DataArray.to_unstacked_dataset for single-dimension variables. (4049) By Deepak Cherian

• Fix open_rasterio for WarpedVRT with specified src_crs. (4104) By Dave Cole.

Documentation

• update the docstring of DataArray.assign_coords : clarify how to add a new coordinate to an existing dimension and illustrative example (3952, 3958) By Etienne Combrisson.

• update the docstring of Dataset.diff and DataArray.diff so it does document the dim parameter as required. (1040, 3909) By Justus Magin.

• Updated Calculating Seasonal Averages from Timeseries of Monthly Means example notebook to take advantage of the new days_in_month accessor for xarray.CFTimeIndex (3935). By Spencer Clark.

• Updated the list of current core developers. (3892) By Tom Nicholas.

• Add example for multi-dimensional extrapolation and note different behavior of kwargs in Dataset.interp and DataArray.interp for 1-d and n-d interpolation (3956). By Matthias Riße.

• Apply black to all the code in the documentation (4012) By Justus Magin.

• Narrative documentation now describes map_blocks: dask.automatic-parallelization. By Deepak Cherian.

• Document .plot, .dt, .str accessors the way they are called. (3625, 3988) By Justus Magin.

• Add documentation for the parameters and return values of DataArray.sel. By Justus Magin.

Internal Changes

• Raise more informative error messages for chunk size conflicts when writing to zarr files. By Deepak Cherian.

• Run the isort pre-commit hook only on python source files and update the flake8 version. (3750, 3711) By Justus Magin.

• Add blackdoc to the list of checkers for development. (4177) By Justus Magin.

• Add a CI job that runs the tests with every optional dependency except dask. (3794, 3919) By Justus Magin.

• Use async / await for the asynchronous distributed tests. (3987, 3989) By Justus Magin.

• Various internal code clean-ups (4026, 4038). By Prajjwal Nijhara.

This release brings many new features such as Dataset.weighted methods for weighted array reductions, a new jupyter repr by default, and the start of units integration with pint. There's also the usual batch of usability improvements, documentation additions, and bug fixes.

Breaking changes

• Raise an error when assigning to the .values or .data attribute of dimension coordinates i.e. IndexVariable objects. This has been broken since v0.12.0. Please use DataArray.assign_coords or Dataset.assign_coords instead. (3470, 3862) By Deepak Cherian

New Features

• Weighted array reductions are now supported via the new DataArray.weighted and Dataset.weighted methods. See compute.weighted. (422, 2922). By Mathias Hauser.

• The new jupyter notebook repr (Dataset._repr_html_ and DataArray._repr_html_) (introduced in 0.14.1) is now on by default. To disable, use xarray.set_options(display_style="text"). By Julia Signell.

• Added support for pandas.DatetimeIndex-style rounding of cftime.datetime objects directly via a CFTimeIndex or via the ~core.accessor_dt.DatetimeAccessor. By Spencer Clark

• Support new h5netcdf backend keyword phony_dims (available from h5netcdf v0.8.0 for ~xarray.backends.H5NetCDFStore. By Kai Mühlbauer.

• Add partial support for unit aware arrays with pint. (3706, 3611) By Justus Magin.

• Dataset.groupby and DataArray.groupby now raise a TypeError on multiple string arguments. Receiving multiple string arguments often means a user is attempting to pass multiple dimensions as separate arguments and should instead pass a single list of dimensions. (3802) By Maximilian Roos

• map_blocks can now apply functions that add new unindexed dimensions. By Deepak Cherian

• An ellipsis (...) is now supported in the dims argument of Dataset.stack and DataArray.stack, meaning all unlisted dimensions, similar to its meaning in DataArray.transpose. (3826) By Maximilian Roos

• Dataset.where and DataArray.where accept a lambda as a first argument, which is then called on the input; replicating pandas' behavior. By Maximilian Roos.

• skipna is available in Dataset.quantile, DataArray.quantile, core.groupby.DatasetGroupBy.quantile, core.groupby.DataArrayGroupBy.quantile (3843, 3844) By Aaron Spring.

• Add a diff summary for testing.assert_allclose. (3617, 3847) By Justus Magin.

Bug fixes

• Fix Dataset.interp when indexing array shares coordinates with the indexed variable (3252). By David Huard.

• Fix recombination of groups in Dataset.groupby and DataArray.groupby when performing an operation that changes the size of the groups along the grouped dimension. By Eric Jansen.

• Fix use of multi-index with categorical values (3674). By Matthieu Ancellin.

• Fix alignment with join="override" when some dimensions are unindexed. (3681). By Deepak Cherian.

• Fix Dataset.swap_dims and DataArray.swap_dims producing index with name reflecting the previous dimension name instead of the new one (3748, 3752). By Joseph K Aicher.

• Use dask_array_type instead of dask_array.Array for type checking. (3779, 3787) By Justus Magin.

• concat can now handle coordinate variables only present in one of the objects to be concatenated when coords="different". By Deepak Cherian.

• xarray now respects the over, under and bad colors if set on a provided colormap. (3590, 3601) By johnomotani.

• coarsen and rolling now respect xr.set_options(keep_attrs=True) to preserve attributes. Dataset.coarsen accepts a keyword argument keep_attrs to change this setting. (3376, 3801) By Andrew Thomas.

• Delete associated indexes when deleting coordinate variables. (3746). By Deepak Cherian.

• Fix Dataset.to_zarr when using append_dim and group simultaneously. (3170). By Matthias Meyer.

• Fix html repr on Dataset with non-string keys (3807). By Maximilian Roos.

Documentation

• Fix documentation of DataArray removing the deprecated mention that when omitted, dims are inferred from a coords-dict. (3821) By Sander van Rijn.

• Improve the where docstring. By Maximilian Roos

• Update the installation instructions: only explicitly list recommended dependencies (3756). By Mathias Hauser.

Internal Changes

• Remove the internal import_seaborn function which handled the deprecation of the seaborn.apionly entry point (3747). By Mathias Hauser.

• Don't test pint integration in combination with datetime objects. (3778, 3788) By Justus Magin.

• Change test_open_mfdataset_list_attr to only run with dask installed (3777, 3780). By Bruno Pagani.

• Preserve the ability to index with method="nearest" with a CFTimeIndex with pandas versions greater than 1.0.1 (3751). By Spencer Clark.

• Greater flexibility and improved test coverage of subtracting various types of objects from a CFTimeIndex. By Spencer Clark.

• Update Azure CI MacOS image, given pending removal. By Maximilian Roos

• Remove xfails for scipy 1.0.1 for tests that append to netCDF files (3805). By Mathias Hauser.

• Remove conversion to pandas.Panel, given its removal in pandas in favor of xarray's objects. By Maximilian Roos

The minor release includes a number of bug-fixes and backwards compatible enhancements.

Enhancements

• ~xarray.DataArray.isin and ~xarray.Dataset.isin methods, which test each value in the array for whether it is contained in the supplied list, returning a bool array. See selecting-values-with-isin for full details. Similar to the np.isin function. By Maximilian Roos.

• Some speed improvement to construct ~xarray.computation.rolling.DataArrayRolling object (1993) By Keisuke Fujii.

• Handle variables with different values for missing_value and _FillValue by masking values for both attributes; previously this resulted in a ValueError. (2016) By Ryan May.

Bug fixes

• Fixed decode_cf function to operate lazily on dask arrays (1372). By Ryan Abernathey.

• Fixed labeled indexing with slice bounds given by xarray objects with datetime64 or timedelta64 dtypes (1240). By Stephan Hoyer.

• Attempting to convert an xarray.Dataset into a numpy array now raises an informative error message. By Stephan Hoyer.

• Fixed a bug in decode_cf_datetime where int32 arrays weren't parsed correctly (2002). By Fabien Maussion.

• When calling xr.auto_combine() or xr.open_mfdataset() with a concat_dim, the resulting dataset will have that one-element dimension (it was silently dropped, previously) (1988). By Ben Root.

The minor release includes a number of bug-fixes and enhancements, along with one possibly backwards incompatible change.

Backwards incompatible changes

• The addition of __array_ufunc__ for xarray objects (see below) means that NumPy ufunc methods (e.g., np.add.reduce) that previously worked on xarray.DataArray objects by converting them into NumPy arrays will now raise NotImplementedError instead. In all cases, the work-around is simple: convert your objects explicitly into NumPy arrays before calling the ufunc (e.g., with .values).

Enhancements

• Added ~xarray.dot, equivalent to numpy.einsum. Also, ~xarray.DataArray.dot now supports dims option, which specifies the dimensions to sum over. (1951) By Keisuke Fujii.

• Support for writing xarray datasets to netCDF files (netcdf4 backend only) when using the dask.distributed scheduler (1464). By Joe Hamman.

• Support lazy vectorized-indexing. After this change, flexible indexing such as orthogonal/vectorized indexing, becomes possible for all the backend arrays. Also, lazy transpose is now also supported. (1897) By Keisuke Fujii.

• Implemented NumPy's __array_ufunc__ protocol for all xarray objects (1617). This enables using NumPy ufuncs directly on xarray.Dataset objects with recent versions of NumPy (v1.13 and newer):

  ds = xr.Dataset({"a": 1})
  np.sin(ds)

  This obliviates the need for the xarray.ufuncs module, which will be deprecated in the future when xarray drops support for older versions of NumPy. By Stephan Hoyer.

• Improve ~xarray.DataArray.rolling logic. ~xarray.computation.rolling.DataArrayRolling object now supports ~xarray.computation.rolling.DataArrayRolling.construct method that returns a view of the DataArray / Dataset object with the rolling-window dimension added to the last axis. This enables more flexible operation, such as strided rolling, windowed rolling, ND-rolling, short-time FFT and convolution. (1831, 1142, 819) By Keisuke Fujii.

• ~plot.line() learned to make plots with data on x-axis if so specified. (575) By Deepak Cherian.

Bug fixes

• Raise an informative error message when using apply_ufunc with numpy v1.11 (1956). By Stephan Hoyer.

• Fix the precision drop after indexing datetime64 arrays (1932). By Keisuke Fujii.

• Silenced irrelevant warnings issued by open_rasterio (1964). By Stephan Hoyer.

• Fix kwarg colors clashing with auto-inferred cmap (1461) By Deepak Cherian.

• Fix ~xarray.plot.imshow error when passed an RGB array with size one in a spatial dimension. By Zac Hatfield-Dodds.

The minor release includes a number of bug-fixes and backwards compatible enhancements.

Documentation

• Added a new guide on contributing (640) By Joe Hamman.

• Added apply_ufunc example to /examples/weather-data (1844). By Liam Brannigan.

• New entry Why don’t aggregations return Python scalars? in the faq (1726). By 0x0L.

Enhancements

New functions and methods:

• Added DataArray.to_iris and DataArray.from_iris for converting data arrays to and from Iris Cubes with the same data and coordinates (621 and 37). By Neil Parley and Duncan Watson-Parris.

• Experimental support for using Zarr as storage layer for xarray (1223). By Ryan Abernathey and Joe Hamman.

• New ~xarray.DataArray.rank on arrays and datasets. Requires bottleneck (1731). By 0x0L.

• .dt accessor can now ceil, floor and round timestamps to specified frequency. By Deepak Cherian.

Plotting enhancements:

• xarray.plot.imshow now handles RGB and RGBA images. Saturation can be adjusted with vmin and vmax, or with robust=True. By Zac Hatfield-Dodds.

• ~plot.contourf() learned to contour 2D variables that have both a 1D coordinate (e.g. time) and a 2D coordinate (e.g. depth as a function of time) (1737). By Deepak Cherian.

• ~plot.plot() rotates x-axis ticks if x-axis is time. By Deepak Cherian.

• ~plot.line() can draw multiple lines if provided with a 2D variable. By Deepak Cherian.

Other enhancements:

• Reduce methods such as DataArray.sum() now handles object-type array.

  da = xr.DataArray(np.array([True, False, np.nan], dtype=object), dims="x")
  da.sum()

  (1866) By Keisuke Fujii.

• Reduce methods such as DataArray.sum() now accepts dtype arguments. (1838) By Keisuke Fujii.

• Added nodatavals attribute to DataArray when using ~xarray.open_rasterio. (1736). By Alan Snow.

• Use pandas.Grouper class in xarray resample methods rather than the deprecated pandas.TimeGrouper class (1766). By Joe Hamman.

• Experimental support for parsing ENVI metadata to coordinates and attributes in xarray.open_rasterio. By Matti Eskelinen.

• Reduce memory usage when decoding a variable with a scale_factor, by converting 8-bit and 16-bit integers to float32 instead of float64 (1840), and keeping float16 and float32 as float32 (1842). Correspondingly, encoded variables may also be saved with a smaller dtype. By Zac Hatfield-Dodds.

• Speed of reindexing/alignment with dask array is orders of magnitude faster when inserting missing values (1847). By Stephan Hoyer.

• Fix axis keyword ignored when applying np.squeeze to DataArray (1487). By Florian Pinault.

• netcdf4-python has moved the its time handling in the netcdftime module to a standalone package (netcdftime). As such, xarray now considers netcdftime an optional dependency. One benefit of this change is that it allows for encoding/decoding of datetimes with non-standard calendars without the netcdf4-python dependency (1084). By Joe Hamman.

New functions/methods

• New ~xarray.DataArray.rank on arrays and datasets. Requires bottleneck (1731). By 0x0L.

Bug fixes

• Rolling aggregation with center=True option now gives the same result with pandas including the last element (1046). By Keisuke Fujii.

• Support indexing with a 0d-np.ndarray (1921). By Keisuke Fujii.

• Added warning in api.py of a netCDF4 bug that occurs when the filepath has 88 characters (1745). By Liam Brannigan.

• Fixed encoding of multi-dimensional coordinates in ~Dataset.to_netcdf (1763). By Mike Neish.

• Fixed chunking with non-file-based rasterio datasets (1816) and refactored rasterio test suite. By Ryan Abernathey

• Bug fix in open_dataset(engine='pydap') (1775) By Keisuke Fujii.

• Bug fix in vectorized assignment (1743, 1744). Now item assignment to ~DataArray.__setitem__ checks

• Bug fix in vectorized assignment (1743, 1744). Now item assignment to DataArray.__setitem__ checks coordinates of target, destination and keys. If there are any conflict among these coordinates, IndexError will be raised. By Keisuke Fujii.

• Properly point DataArray.__dask_scheduler__ to dask.threaded.get. By Matthew Rocklin.

• Bug fixes in DataArray.plot.imshow: all-NaN arrays and arrays with size one in some dimension can now be plotted, which is good for exploring satellite imagery (1780). By Zac Hatfield-Dodds.

• Fixed UnboundLocalError when opening netCDF file (1781). By Stephan Hoyer.

• The variables, attrs, and dimensions properties have been deprecated as part of a bug fix addressing an issue where backends were unintentionally loading the datastores data and attributes repeatedly during writes (1798). By Joe Hamman.

• Compatibility fixes to plotting module for NumPy 1.14 and pandas 0.22 (1813). By Joe Hamman.

• Bug fix in encoding coordinates with {'_FillValue': None} in netCDF metadata (1865). By Chris Roth.

• Fix indexing with lists for arrays loaded from netCDF files with engine='h5netcdf (1864). By Stephan Hoyer.

• Corrected a bug with incorrect coordinates for non-georeferenced geotiff files (1686). Internally, we now use the rasterio coordinate transform tool instead of doing the computations ourselves. A parse_coordinates kwarg has been added to ~open_rasterio (set to True per default). By Fabien Maussion.

• The colors of discrete colormaps are now the same regardless if seaborn is installed or not (1896). By Fabien Maussion.

• Fixed dtype promotion rules in where and concat to match pandas (1847). A combination of strings/numbers or unicode/bytes now promote to object dtype, instead of strings or unicode. By Stephan Hoyer.

• Fixed bug where ~xarray.DataArray.isnull was loading data stored as dask arrays (1937). By Joe Hamman.

This is a major release that includes bug fixes, new features and a few backwards incompatible changes. Highlights include:

• Indexing now supports broadcasting over dimensions, similar to NumPy's vectorized indexing (but better!).

• ~DataArray.resample has a new groupby-like API like pandas.

• ~xarray.apply_ufunc facilitates wrapping and parallelizing functions written for NumPy arrays.

• Performance improvements, particularly for dask and open_mfdataset.

Breaking changes

• xarray now supports a form of vectorized indexing with broadcasting, where the result of indexing depends on dimensions of indexers, e.g., array.sel(x=ind) with ind.dims == ('y',). Alignment between coordinates on indexed and indexing objects is also now enforced. Due to these changes, existing uses of xarray objects to index other xarray objects will break in some cases.

  The new indexing API is much more powerful, supporting outer, diagonal and vectorized indexing in a single interface. The isel_points and sel_points methods are deprecated, since they are now redundant with the isel / sel methods. See vectorized-indexing for the details (1444, 1436). By Keisuke Fujii and Stephan Hoyer.

• A new resampling interface to match pandas' groupby-like API was added to Dataset.resample and DataArray.resample (1272). Timeseries resampling is fully supported for data with arbitrary dimensions as is both downsampling and upsampling (including linear, quadratic, cubic, and spline interpolation).

  Old syntax:

  ds.resample("24H", dim="time", how="max")

  New syntax:

  ds.resample(time="24H").max()

  Note that both versions are currently supported, but using the old syntax will produce a warning encouraging users to adopt the new syntax. By Daniel Rothenberg.

• Calling repr() or printing xarray objects at the command line or in a Jupyter Notebook will not longer automatically compute dask variables or load data on arrays lazily loaded from disk (1522). By Guido Imperiale.

• Supplying coords as a dictionary to the DataArray constructor without also supplying an explicit dims argument is no longer supported. This behavior was deprecated in version 0.9 but will now raise an error (727).

• Several existing features have been deprecated and will change to new behavior in xarray v0.11. If you use any of them with xarray v0.10, you should see a FutureWarning that describes how to update your code:

  • Dataset.T has been deprecated an alias for Dataset.transpose() (1232). In the next major version of xarray, it will provide short- cut lookup for variables or attributes with name 'T'.

  • DataArray.__contains__ (e.g., key in data_array) currently checks for membership in DataArray.coords. In the next major version of xarray, it will check membership in the array data found in DataArray.values instead (1267).

  • Direct iteration over and counting a Dataset (e.g., [k for k in ds], ds.keys(), ds.values(), len(ds) and if ds) currently includes all variables, both data and coordinates. For improved usability and consistency with pandas, in the next major version of xarray these will change to only include data variables (884). Use ds.variables, ds.data_vars or ds.coords as alternatives.

• Changes to minimum versions of dependencies:

  • Old numpy < 1.11 and pandas < 0.18 are no longer supported (1512). By Keisuke Fujii.

  • The minimum supported version bottleneck has increased to 1.1 (1279). By Joe Hamman.

Enhancements

New functions/methods

• New helper function ~xarray.apply_ufunc for wrapping functions written to work on NumPy arrays to support labels on xarray objects (770). apply_ufunc also support automatic parallelization for many functions with dask. See compute.wrapping-custom and dask.automatic-parallelization for details. By Stephan Hoyer.

• Added new method Dataset.to_dask_dataframe, convert a dataset into a dask dataframe. This allows lazy loading of data from a dataset containing dask arrays (1462). By James Munroe.

• New function ~xarray.where for conditionally switching between values in xarray objects, like numpy.where:

  import xarray as xr
  
  arr = xr.DataArray([[1, 2, 3], [4, 5, 6]], dims=("x", "y"))
  
  xr.where(arr % 2, "even", "odd")

  <xarray.DataArray (x: 2, y: 3)>
  array([['even', 'odd', 'even'],
         ['odd', 'even', 'odd']],
        dtype='<U4')
  Dimensions without coordinates: x, y

  Equivalently, the ~xarray.Dataset.where method also now supports the other argument, for filling with a value other than NaN (576). By Stephan Hoyer.

• Added ~xarray.show_versions function to aid in debugging (1485). By Joe Hamman.

Performance improvements

• ~xarray.concat was computing variables that aren't in memory (e.g. dask-based) multiple times; ~xarray.open_mfdataset was loading them multiple times from disk. Now, both functions will instead load them at most once and, if they do, store them in memory in the concatenated array/dataset (1521). By Guido Imperiale.

• Speed-up (x 100) of xarray.conventions.decode_cf_datetime. By Christian Chwala.

IO related improvements

• Unicode strings (str on Python 3) are now round-tripped successfully even when written as character arrays (e.g., as netCDF3 files or when using engine='scipy') (1638). This is controlled by the _Encoding attribute convention, which is also understood directly by the netCDF4-Python interface. See io.string-encoding for full details. By Stephan Hoyer.

• Support for data_vars and coords keywords from ~xarray.concat added to ~xarray.open_mfdataset (438). Using these keyword arguments can significantly reduce memory usage and increase speed. By Oleksandr Huziy.

• Support for pathlib.Path objects added to ~xarray.open_dataset, ~xarray.open_mfdataset, xarray.to_netcdf, and ~xarray.save_mfdataset (799):

  from pathlib import Path  # In Python 2, use pathlib2!
  
  data_dir = Path("data/")
  
  one_file = data_dir / "dta_for_month_01.nc"
  
  xr.open_dataset(one_file)

  By Willi Rath.

• You can now explicitly disable any default _FillValue (NaN for floating point values) by passing the encoding {'_FillValue': None} (1598). By Stephan Hoyer.

• More attributes available in ~xarray.Dataset.attrs dictionary when raster files are opened with ~xarray.open_rasterio. By Greg Brener.

• Support for NetCDF files using an _Unsigned attribute to indicate that a a signed integer data type should be interpreted as unsigned bytes (1444). By Eric Bruning.

• Support using an existing, opened netCDF4 Dataset with ~xarray.backends.NetCDF4DataStore. This permits creating an ~xarray.Dataset from a netCDF4 Dataset that has been opened using other means (1459). By Ryan May.

• Changed ~xarray.backends.PydapDataStore to take a Pydap dataset. This permits opening Opendap datasets that require authentication, by instantiating a Pydap dataset with a session object. Also added xarray.backends.PydapDataStore.open which takes a url and session object (1068). By Philip Graae.

• Support reading and writing unlimited dimensions with h5netcdf (1636). By Joe Hamman.

Other improvements

• Added _ipython_key_completions_ to xarray objects, to enable autocompletion for dictionary-like access in IPython, e.g., ds['tem + tab -> ds['temperature'] (1628). By Keisuke Fujii.

• Support passing keyword arguments to load, compute, and persist methods. Any keyword arguments supplied to these methods are passed on to the corresponding dask function (1523). By Joe Hamman.

• Encoding attributes are now preserved when xarray objects are concatenated. The encoding is copied from the first object (1297). By Joe Hamman and Gerrit Holl.

• Support applying rolling window operations using bottleneck's moving window functions on data stored as dask arrays (1279). By Joe Hamman.

• Experimental support for the Dask collection interface (1674). By Matthew Rocklin.

Bug fixes

• Suppress RuntimeWarning issued by numpy for "invalid value comparisons" (e.g. NaN). Xarray now behaves similarly to pandas in its treatment of binary and unary operations on objects with NaNs (1657). By Joe Hamman.

• Unsigned int support for reduce methods with skipna=True (1562). By Keisuke Fujii.

• Fixes to ensure xarray works properly with pandas 0.21:

  • Fix ~xarray.DataArray.isnull method (1549).

  • ~xarray.DataArray.to_series and ~xarray.Dataset.to_dataframe should not return a pandas.MultiIndex for 1D data (1548).

  • Fix plotting with datetime64 axis labels (1661).

  By Stephan Hoyer.

• ~xarray.open_rasterio method now shifts the rasterio coordinates so that they are centered in each pixel (1468). By Greg Brener.

• ~xarray.Dataset.rename method now doesn't throw errors if some Variable is renamed to the same name as another Variable as long as that other Variable is also renamed (1477). This method now does throw when two Variables would end up with the same name after the rename (since one of them would get overwritten in this case). By Prakhar Goel.

• Fix xarray.testing.assert_allclose to actually use atol and rtol arguments when called on DataArray objects (1488). By Stephan Hoyer.

• xarray quantile methods now properly raise a TypeError when applied to objects with data stored as dask arrays (1529). By Joe Hamman.

• Fix positional indexing to allow the use of unsigned integers (1405). By Joe Hamman and Gerrit Holl.

• Creating a Dataset now raises MergeError if a coordinate shares a name with a dimension but is comprised of arbitrary dimensions (1120). By Joe Hamman.

• ~xarray.open_rasterio method now skips rasterio's crs attribute if its value is None (1520). By Leevi Annala.

• Fix xarray.DataArray.to_netcdf to return bytes when no path is provided (1410). By Joe Hamman.

• Fix xarray.save_mfdataset to properly raise an informative error when objects other than Dataset are provided (1555). By Joe Hamman.

• xarray.Dataset.copy would not preserve the encoding property (1586). By Guido Imperiale.

• xarray.concat would eagerly load dask variables into memory if the first argument was a numpy variable (1588). By Guido Imperiale.

• Fix bug in ~xarray.Dataset.to_netcdf when writing in append mode (1215). By Joe Hamman.

• Fix netCDF4 backend to properly roundtrip the shuffle encoding option (1606). By Joe Hamman.

• Fix bug when using pytest class decorators to skipping certain unittests. The previous behavior unintentionally causing additional tests to be skipped (1531). By Joe Hamman.

• Fix pynio backend for upcoming release of pynio with Python 3 support (1611). By Ben Hillman.

• Fix seaborn import warning for Seaborn versions 0.8 and newer when the apionly module was deprecated. (1633). By Joe Hamman.

• Fix COMPAT: MultiIndex checking is fragile (1833). By Florian Pinault.

• Fix rasterio backend for Rasterio versions 1.0alpha10 and newer. (1641). By Chris Holden.

Bug fixes after rc1

• Suppress warning in IPython autocompletion, related to the deprecation of .T attributes (1675). By Keisuke Fujii.

• Fix a bug in lazily-indexing netCDF array. (1688) By Keisuke Fujii.

• (Internal bug) MemoryCachedArray now supports the orthogonal indexing. Also made some internal cleanups around array wrappers (1429). By Keisuke Fujii.

• (Internal bug) MemoryCachedArray now always wraps np.ndarray by NumpyIndexingAdapter. (1694) By Keisuke Fujii.

• Fix importing xarray when running Python with -OO (1706). By Stephan Hoyer.

• Saving a netCDF file with a coordinates with a spaces in its names now raises an appropriate warning (1689). By Stephan Hoyer.

• Fix two bugs that were preventing dask arrays from being specified as coordinates in the DataArray constructor (1684). By Joe Hamman.

• Fixed apply_ufunc with dask='parallelized' for scalar arguments (1697). By Stephan Hoyer.

• Fix "Chunksize cannot exceed dimension size" error when writing netCDF4 files loaded from disk (1225). By Stephan Hoyer.

• Validate the shape of coordinates with names matching dimensions in the DataArray constructor (1709). By Stephan Hoyer.

• Raise NotImplementedError when attempting to save a MultiIndex to a netCDF file (1547). By Stephan Hoyer.

• Remove netCDF dependency from rasterio backend tests. By Matti Eskelinen

Bug fixes after rc2

• Fixed unexpected behavior in Dataset.set_index() and DataArray.set_index() introduced by pandas 0.21.0. Setting a new index with a single variable resulted in 1-level pandas.MultiIndex instead of a simple pandas.Index (1722). By Benoit Bovy.

• Fixed unexpected memory loading of backend arrays after print. (1720). By Keisuke Fujii.

This release includes a number of backwards compatible enhancements and bug fixes.

Enhancements

• New ~xarray.Dataset.sortby method to Dataset and DataArray that enable sorting along dimensions (967). See the docs for examples. By Chun-Wei Yuan and Kyle Heuton.

• Add .dt accessor to DataArrays for computing datetime-like properties for the values they contain, similar to pandas.Series (358). By Daniel Rothenberg.

• Renamed internal dask arrays created by open_dataset to match new dask conventions (1343). By Ryan Abernathey.

• ~xarray.as_variable is now part of the public API (1303). By Benoit Bovy.

• ~xarray.align now supports join='exact', which raises an error instead of aligning when indexes to be aligned are not equal. By Stephan Hoyer.

• New function ~xarray.open_rasterio for opening raster files with the rasterio library. See the docs for details. By Joe Hamman, Nic Wayand and Fabien Maussion

Bug fixes

• Fix error from repeated indexing of datasets loaded from disk (1374). By Stephan Hoyer.

• Fix a bug where .isel_points wrongly assigns unselected coordinate to data_vars. By Keisuke Fujii.

• Tutorial datasets are now checked against a reference MD5 sum to confirm successful download (1392). By Matthew Gidden.

• DataArray.chunk() now accepts dask specific kwargs like Dataset.chunk() does. By Fabien Maussion.

• Support for engine='pydap' with recent releases of Pydap (3.2.2+), including on Python 3 (1174).

Documentation

• A new gallery allows to add interactive examples to the documentation. By Fabien Maussion.

Testing

• Fix test suite failure caused by changes to pandas.cut function (1386). By Ryan Abernathey.

• Enhanced tests suite by use of @network decorator, which is controlled via --run-network-tests command line argument to py.test (1393). By Matthew Gidden.

Remove an inadvertently introduced print statement.

This minor release includes bug-fixes and backwards compatible enhancements.

Enhancements

• New ~xarray.DataArray.persist method to Datasets and DataArrays to enable persisting data in distributed memory when using Dask (1344). By Matthew Rocklin.

• New ~xarray.DataArray.expand_dims method for DataArray and Dataset (1326). By Keisuke Fujii.

Bug fixes

• Fix .where() with drop=True when arguments do not have indexes (1350). This bug, introduced in v0.9, resulted in xarray producing incorrect results in some cases. By Stephan Hoyer.

• Fixed writing to file-like objects with ~xarray.Dataset.to_netcdf (1320). Stephan Hoyer.

• Fixed explicitly setting engine='scipy' with to_netcdf when not providing a path (1321). Stephan Hoyer.

• Fixed open_dataarray does not pass properly its parameters to open_dataset (1359). Stephan Hoyer.

• Ensure test suite works when runs from an installed version of xarray (1336). Use @pytest.mark.slow instead of a custom flag to mark slow tests. By Stephan Hoyer

The minor release includes bug-fixes and backwards compatible enhancements.

Enhancements

• rolling on Dataset is now supported (859).

• .rolling() on Dataset is now supported (859). By Keisuke Fujii.

• When bottleneck version 1.1 or later is installed, use bottleneck for rolling var, argmin, argmax, and rank computations. Also, rolling median now accepts a min_periods argument (1276). By Joe Hamman.

• When .plot() is called on a 2D DataArray and only one dimension is specified with x= or y=, the other dimension is now guessed (1291). By Vincent Noel.

• Added new method ~Dataset.assign_attrs to DataArray and Dataset, a chained-method compatible implementation of the dict.update method on attrs (1281). By Henry S. Harrison.

• Added new autoclose=True argument to ~xarray.open_mfdataset to explicitly close opened files when not in use to prevent occurrence of an OS Error related to too many open files (1198). Note, the default is autoclose=False, which is consistent with previous xarray behavior. By Phillip J. Wolfram.

• The repr() of Dataset and DataArray attributes uses a similar format to coordinates and variables, with vertically aligned entries truncated to fit on a single line (1319). Hopefully this will stop people writing data.attrs = {} and discarding metadata in notebooks for the sake of cleaner output. The full metadata is still available as data.attrs. By Zac Hatfield-Dodds.

• Enhanced tests suite by use of @slow and @flaky decorators, which are controlled via --run-flaky and --skip-slow command line arguments to py.test (1336). By Stephan Hoyer and Phillip J. Wolfram.

• New aggregation on rolling objects ~computation.rolling.DataArrayRolling.count which providing a rolling count of valid values (1138).

Bug fixes

• Rolling operations now keep preserve original dimension order (1125). By Keisuke Fujii.

• Fixed sel with method='nearest' on Python 2.7 and 64-bit Windows (1140). Stephan Hoyer.

• Fixed where with drop='True' for empty masks (1341). By Stephan Hoyer and Phillip J. Wolfram.

Renamed the "Unindexed dimensions" section in the Dataset and DataArray repr (added in v0.9.0) to "Dimensions without coordinates" (1199).

This major release includes five months worth of enhancements and bug fixes from 24 contributors, including some significant changes that are not fully backwards compatible. Highlights include:

• Coordinates are now optional in the xarray data model, even for dimensions.

• Changes to caching, lazy loading and pickling to improve xarray's experience for parallel computing.

• Improvements for accessing and manipulating pandas.MultiIndex levels.

• Many new methods and functions, including ~DataArray.quantile, ~DataArray.cumsum, ~DataArray.cumprod ~DataArray.combine_first ~DataArray.set_index, ~DataArray.reset_index, ~DataArray.reorder_levels, ~xarray.full_like, ~xarray.zeros_like, ~xarray.ones_like ~xarray.open_dataarray, ~DataArray.compute, Dataset.info, testing.assert_equal, testing.assert_identical, and testing.assert_allclose.

Breaking changes

• Index coordinates for each dimensions are now optional, and no longer created by default 1017. You can identify such dimensions without coordinates by their appearance in list of "Dimensions without coordinates" in the Dataset or DataArray repr:

  xr.Dataset({"foo": (("x", "y"), [[1, 2]])})

  <xarray.Dataset>
  Dimensions:  (x: 1, y: 2)
  Dimensions without coordinates: x, y
  Data variables:
      foo      (x, y) int64 1 2

  This has a number of implications:

  • ~align and ~Dataset.reindex can now error, if dimensions labels are missing and dimensions have different sizes.

  • Because pandas does not support missing indexes, methods such as to_dataframe/from_dataframe and stack/unstack no longer roundtrip faithfully on all inputs. Use ~Dataset.reset_index to remove undesired indexes.

  • Dataset.__delitem__ and ~Dataset.drop no longer delete/drop variables that have dimensions matching a deleted/dropped variable.

  • DataArray.coords.__delitem__ is now allowed on variables matching dimension names.

  • .sel and .loc now handle indexing along a dimension without coordinate labels by doing integer based indexing. See indexing.missing_coordinates for an example.

  • ~Dataset.indexes is no longer guaranteed to include all dimensions names as keys. The new method ~Dataset.get_index has been added to get an index for a dimension guaranteed, falling back to produce a default RangeIndex if necessary.

• The default behavior of merge is now compat='no_conflicts', so some merges will now succeed in cases that previously raised xarray.MergeError. Set compat='broadcast_equals' to restore the previous default. See combining.no_conflicts for more details.

• Reading ~DataArray.values no longer always caches values in a NumPy array 1128. Caching of .values on variables read from netCDF files on disk is still the default when open_dataset is called with cache=True. By Guido Imperiale and Stephan Hoyer.

• Pickling a Dataset or DataArray linked to a file on disk no longer caches its values into memory before pickling (1128). Instead, pickle stores file paths and restores objects by reopening file references. This enables preliminary, experimental use of xarray for opening files with dask.distributed. By Stephan Hoyer.

• Coordinates used to index a dimension are now loaded eagerly into pandas.Index objects, instead of loading the values lazily. By Guido Imperiale.

• Automatic levels for 2d plots are now guaranteed to land on vmin and vmax when these kwargs are explicitly provided (1191). The automated level selection logic also slightly changed. By Fabien Maussion.

• DataArray.rename() behavior changed to strictly change the DataArray.name if called with string argument, or strictly change coordinate names if called with dict-like argument. By Markus Gonser.

• By default to_netcdf() add a _FillValue = NaN attributes to float types. By Frederic Laliberte.

• repr on DataArray objects uses an shortened display for NumPy array data that is less likely to overflow onto multiple pages (1207). By Stephan Hoyer.

• xarray no longer supports python 3.3, versions of dask prior to v0.9.0, or versions of bottleneck prior to v1.0.

Deprecations

• Renamed the Coordinate class from xarray's low level API to ~xarray.IndexVariable. Variable.to_variable and Variable.to_coord have been renamed to ~xarray.Variable.to_base_variable and ~xarray.Variable.to_index_variable.

• Deprecated supplying coords as a dictionary to the DataArray constructor without also supplying an explicit dims argument. The old behavior encouraged relying on the iteration order of dictionaries, which is a bad practice (727).

• Removed a number of methods deprecated since v0.7.0 or earlier: load_data, vars, drop_vars, dump, dumps and the variables keyword argument to Dataset.

• Removed the dummy module that enabled import xray.

Enhancements

• Added new method ~DataArray.combine_first to DataArray and Dataset, based on the pandas method of the same name (see combine). By Chun-Wei Yuan.

• Added the ability to change default automatic alignment (arithmetic_join="inner") for binary operations via ~xarray.set_options() (see math-automatic-alignment). By Chun-Wei Yuan.

• Add checking of attr names and values when saving to netCDF, raising useful error messages if they are invalid. (911). By Robin Wilson.

• Added ability to save DataArray objects directly to netCDF files using ~xarray.DataArray.to_netcdf, and to load directly from netCDF files using ~xarray.open_dataarray (915). These remove the need to convert a DataArray to a Dataset before saving as a netCDF file, and deals with names to ensure a perfect 'roundtrip' capability. By Robin Wilson.

• Multi-index levels are now accessible as "virtual" coordinate variables, e.g., ds['time'] can pull out the 'time' level of a multi-index (see coordinates). sel also accepts providing multi-index levels as keyword arguments, e.g., ds.sel(time='2000-01') (see multi-level-indexing). By Benoit Bovy.

• Added set_index, reset_index and reorder_levels methods to easily create and manipulate (multi-)indexes (see reshape.set_index). By Benoit Bovy.

• Added the compat option 'no_conflicts' to merge, allowing the combination of xarray objects with disjoint (742) or overlapping (835) coordinates as long as all present data agrees. By Johnnie Gray. See combining.no_conflicts for more details.

• It is now possible to set concat_dim=None explicitly in ~xarray.open_mfdataset to disable inferring a dimension along which to concatenate. By Stephan Hoyer.

• Added methods DataArray.compute, Dataset.compute, and Variable.compute as a non-mutating alternative to ~DataArray.load. By Guido Imperiale.

• Adds DataArray and Dataset methods ~xarray.DataArray.cumsum and ~xarray.DataArray.cumprod. By Phillip J. Wolfram.

• New properties Dataset.sizes and DataArray.sizes for providing consistent access to dimension length on both Dataset and DataArray (921). By Stephan Hoyer.

• New keyword argument drop=True for ~DataArray.sel, ~DataArray.isel and ~DataArray.squeeze for dropping scalar coordinates that arise from indexing. DataArray (242). By Stephan Hoyer.

• New top-level functions ~xarray.full_like, ~xarray.zeros_like, and ~xarray.ones_like By Guido Imperiale.

• Overriding a preexisting attribute with ~xarray.register_dataset_accessor or ~xarray.register_dataarray_accessor now issues a warning instead of raising an error (1082). By Stephan Hoyer.

• Options for axes sharing between subplots are exposed to ~xarray.plot.FacetGrid and ~xarray.plot.plot, so axes sharing can be disabled for polar plots. By Bas Hoonhout.

• New utility functions ~xarray.testing.assert_equal, ~xarray.testing.assert_identical, and ~xarray.testing.assert_allclose for asserting relationships between xarray objects, designed for use in a pytest test suite.

• figsize, size and aspect plot arguments are now supported for all plots (897). See plotting-figsize for more details. By Stephan Hoyer and Fabien Maussion.

• New ~Dataset.info method to summarize Dataset variables and attributes. The method prints to a buffer (e.g. stdout) with output similar to what the command line utility ncdump -h produces (1150). By Joe Hamman.

• Added the ability write unlimited netCDF dimensions with the scipy and netcdf4 backends via the new xray.Dataset.encoding attribute or via the unlimited_dims argument to xray.Dataset.to_netcdf. By Joe Hamman.

• New ~DataArray.quantile method to calculate quantiles from DataArray objects (1187). By Joe Hamman.

Bug fixes

• groupby_bins now restores empty bins by default (1019). By Ryan Abernathey.

• Fix issues for dates outside the valid range of pandas timestamps (975). By Mathias Hauser.

• Unstacking produced flipped array after stacking decreasing coordinate values (980). By Stephan Hoyer.

• Setting dtype via the encoding parameter of to_netcdf failed if the encoded dtype was the same as the dtype of the original array (873). By Stephan Hoyer.

• Fix issues with variables where both attributes _FillValue and missing_value are set to NaN (997). By Marco Zühlke.

• .where() and .fillna() now preserve attributes (1009). By Fabien Maussion.

• Applying broadcast() to an xarray object based on the dask backend won't accidentally convert the array from dask to numpy anymore (978). By Guido Imperiale.

• Dataset.concat() now preserves variables order (1027). By Fabien Maussion.

• Fixed an issue with pcolormesh (781). A new infer_intervals keyword gives control on whether the cell intervals should be computed or not. By Fabien Maussion.

• Grouping over an dimension with non-unique values with groupby gives correct groups. By Stephan Hoyer.

• Fixed accessing coordinate variables with non-string names from .coords. By Stephan Hoyer.

• ~xarray.DataArray.rename now simultaneously renames the array and any coordinate with the same name, when supplied via a dict (1116). By Yves Delley.

• Fixed sub-optimal performance in certain operations with object arrays (1121). By Yves Delley.

• Fix .groupby(group) when group has datetime dtype (1132). By Jonas Sølvsteen.

• Fixed a bug with facetgrid (the norm keyword was ignored, 1159). By Fabien Maussion.

• Resolved a concurrency bug that could cause Python to crash when simultaneously reading and writing netCDF4 files with dask (1172). By Stephan Hoyer.

• Fix to make .copy() actually copy dask arrays, which will be relevant for future releases of dask in which dask arrays will be mutable (1180). By Stephan Hoyer.

• Fix opening NetCDF files with multi-dimensional time variables (1229). By Stephan Hoyer.

Performance improvements

• xarray.Dataset.isel_points and xarray.Dataset.sel_points now use vectorised indexing in numpy and dask (1161), which can result in several orders of magnitude speedup. By Jonathan Chambers.

This release includes a number of bug fixes and minor enhancements.

Breaking changes

• ~xarray.broadcast and ~xarray.concat now auto-align inputs, using join=outer. Previously, these functions raised ValueError for non-aligned inputs. By Guido Imperiale.

Enhancements

• New documentation on panel-transition. By Maximilian Roos.

• New Dataset and DataArray methods ~xarray.Dataset.to_dict and ~xarray.Dataset.from_dict to allow easy conversion between dictionaries and xarray objects (432). See dictionary IO for more details. By Julia Signell.

• Added exclude and indexes optional parameters to ~xarray.align, and exclude optional parameter to ~xarray.broadcast. By Guido Imperiale.

• Better error message when assigning variables without dimensions (971). By Stephan Hoyer.

• Better error message when reindex/align fails due to duplicate index values (956). By Stephan Hoyer.

Bug fixes

• Ensure xarray works with h5netcdf v0.3.0 for arrays with dtype=str (953). By Stephan Hoyer.

• Dataset.__dir__() (i.e. the method python calls to get autocomplete options) failed if one of the dataset's keys was not a string (852). By Maximilian Roos.

• Dataset constructor can now take arbitrary objects as values (647). By Maximilian Roos.

• Clarified copy argument for ~xarray.DataArray.reindex and ~xarray.align, which now consistently always return new xarray objects (927).

• Fix open_mfdataset with engine='pynio' (936). By Stephan Hoyer.

• groupby_bins sorted bin labels as strings (952). By Stephan Hoyer.

• Fix bug introduced by v0.8.0 that broke assignment to datasets when both the left and right side have the same non-unique index values (956).

Bug fixes

• Fix bug in v0.8.0 that broke assignment to Datasets with non-unique indexes (943). By Stephan Hoyer.