What happened?

I ran into a situation where I was indexing with a 2D integer array into some chunked netCDF data. This indexing operation is extremely slow. Using a flat 1D index instead is as fast as expected.

What did you expect to happen?

I would expect indexing on dask data to be very quick since the work is delayed, and indeed it is so in the 1D case. However, the 2D case is very slow -- slower than actually doing the all the work with numpy arrays!

Minimal Complete Verifiable Example

```Python import dask.array import numpy as np import xarray as xr

%%

da = xr.DataArray( data=np.random.rand(100, 1_000_000), dims=("time", "x"), ) dask_da = xr.DataArray( data=dask.array.from_array(da.to_numpy(), chunks=(1, 1_000_000)), dims=("time", "x"), )

indexer = np.random.randint(0, 1_000_000, size=100_000) indexer2d = xr.DataArray( data=indexer.reshape((4, -1)), dims=("a", "b"), )

%%

%timeit da.isel(x=indexer) # 162 ms %timeit da.isel(x=indexer2d) # 164 ms %timeit dask_da.isel(x=indexer) # 5.3 ms %timeit dask_da.isel(x=indexer2d) # 860 ms according to timeit, but 6 to 14 (!) seconds in interactive use ```

MVCE confirmation

• [x] Minimal example — the example is as focused as reasonably possible to demonstrate the underlying issue in xarray.
• [x] Complete example — the example is self-contained, including all data and the text of any traceback.
• [x] Verifiable example — the example copy & pastes into an IPython prompt or Binder notebook, returning the result.
• [x] New issue — a search of GitHub Issues suggests this is not a duplicate.
• [x] Recent environment — the issue occurs with the latest version of xarray and its dependencies.

Relevant log output

No response

Anything else we need to know?

No response

Environment

Is your feature request related to a problem?

If I have a DataArray of values:

python da = xr.DataArray([0, 1, 2, 3, 4, 5]) And I'd like to replace to_replace=[1, 3, 5] by value=[10, 30, 50], there's no method da.replace(to_replace, value) to do this.

There's no easy way like pandas (https://pandas.pydata.org/docs/reference/api/pandas.DataFrame.replace.html) to do this.

(Apologies if I've missed related issues, searching for "replace" gives many hits as the word is obviously used quite often.)

Describe the solution you'd like

python da = xr.DataArray([0, 1, 2, 3, 4, 5]) replaced = da.replace([1, 3, 5], [10, 30, 50]) print(replaced)

<xarray.DataArray (dim_0: 6)> array([ 0, 10, 2, 30, 4, 50]) Dimensions without coordinates: dim_0

I've had a try at a relatively efficient implementation below. I'm wondering whether it's a worthwhile addition to xarray?

Describe alternatives you've considered

Ignoring issues such as dealing with NaNs, chunks, etc., a simple dict lookup:

python def dict_replace(da, to_replace, value): d = {k: v for k, v in zip(to_replace, value)} out = np.vectorize(lambda x: d.get(x, x))(da.values) return da.copy(data=out)

Alternatively, leveraging pandas:

python def pandas_replace(da, to_replace, value): df = pd.DataFrame() df["values"] = da.values.ravel() df["values"].replace(to_replace, value, inplace=True) return da.copy(data=df["values"].values.reshape(da.shape))

But I also tried my hand at a custom implementation, letting np.unique do the heavy lifting: ```python def custom_replace(da, to_replace, value): # Use np.unique to create an inverse index flat = da.values.ravel() uniques, index = np.unique(flat, return_inverse=True)
replaceable = np.isin(flat, to_replace)

# Create a replacement array in which there is a 1:1 relation between
# uniques and the replacement values, so that we can use the inverse index
# to select replacement values. 
valid = np.isin(to_replace, uniques, assume_unique=True)
# Remove to_replace values that are not present in da. If no overlap
# exists between to_replace and the values in da, just return a copy.
if not valid.any():
    return da.copy()
to_replace = to_replace[valid]
value = value[valid]

replacement = np.zeros_like(uniques)
replacement[np.searchsorted(uniques, to_replace)] = value

out = flat.copy()
out[replaceable] = replacement[index[replaceable]]
return da.copy(data=out.reshape(da.shape))

```

Such an approach seems like it's consistently the fastest:

```python da = xr.DataArray(np.random.randint(0, 100, 100_000)) to_replace = np.random.choice(np.arange(100), 10, replace=False) value = to_replace * 200

test1 = custom_replace(da, to_replace, value) test2 = pandas_replace(da, to_replace, value) test3 = dict_replace(da, to_replace, value)

assert test1.equals(test2) assert test1.equals(test3)

6.93 ms ± 295 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

%timeit custom_replace(da, to_replace, value)

9.37 ms ± 212 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

%timeit pandas_replace(da, to_replace, value)

26.8 ms ± 1.59 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

%timeit dict_replace(da, to_replace, value) ```

With the advantage growing the number of values involved:

```python da = xr.DataArray(np.random.randint(0, 10_000, 100_000)) to_replace = np.random.choice(np.arange(10_000), 10_000, replace=False) value = to_replace * 200

test1 = custom_replace(da, to_replace, value) test2 = pandas_replace(da, to_replace, value) test3 = dict_replace(da, to_replace, value)

assert test1.equals(test2) assert test1.equals(test3)

21.6 ms ± 990 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

%timeit custom_replace(da, to_replace, value)

3.12 s ± 574 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

%timeit pandas_replace(da, to_replace, value)

42.7 ms ± 1.47 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

%timeit dict_replace(da, to_replace, value) ```

In my real-life example, with a DataArray of approx 110 000 elements, with 60 000 values to replace, the custom one takes 33 ms, the dict one takes 135 ms, while pandas takes 26 s (!).

Additional context

In all cases, we need dealing with NaNs, checking the input, etc.:

```python def replace(da: xr.DataArray, to_replace: Any, value: Any): from xarray.core.utils import is_scalar

if is_scalar(to_replace):
    if not is_scalar(value):
        raise TypeError("if to_replace is scalar, then value must be a scalar")
    if np.isnan(to_replace):
        return da.fillna(value) 
    else:
        return da.where(da != to_replace, other=value)
else:
    to_replace = np.asarray(to_replace)
    if to_replace.ndim != 1:
        raise ValueError("to_replace must be 1D or scalar")
    if is_scalar(value):
        value = np.full_like(to_replace, value)
    else:
        value = np.asarray(value)
        if to_replace.shape != value.shape:
            raise ValueError(
                f"Replacement arrays must match in shape. "
                f"Expecting {to_replace.shape} got {value.shape} "
            )

_, counts = np.unique(to_replace, return_counts=True)
if (counts > 1).any():
    raise ValueError("to_replace contains duplicates")

# Replace NaN values separately, as they will show up as separate values
# from numpy.unique.
isnan = np.isnan(to_replace)
if isnan.any():
    i = np.nonzero(isnan)[0]
    da = da.fillna(value[i])

# Use np.unique to create an inverse index
flat = da.values.ravel()
uniques, index = np.unique(flat, return_inverse=True)    
replaceable = np.isin(flat, to_replace)

# Create a replacement array in which there is a 1:1 relation between
# uniques and the replacement values, so that we can use the inverse index
# to select replacement values. 
valid = np.isin(to_replace, uniques, assume_unique=True)
# Remove to_replace values that are not present in da. If no overlap
# exists between to_replace and the values in da, just return a copy.
if not valid.any():
    return da.copy()
to_replace = to_replace[valid]
value = value[valid]

replacement = np.zeros_like(uniques)
replacement[np.searchsorted(uniques, to_replace)] = value

out = flat.copy()
out[replaceable] = replacement[index[replaceable]]
return da.copy(data=out.reshape(da.shape))

```

It think it should be easy to use e.g. let it operate on the numpy arrays so e.g. apply_ufunc will work. The primary issue is whether values can be sorted; in such a case the dict lookup might be an okay fallback? I've had a peek at the pandas implementation, but didn't become much wiser.

Anyway, for your consideration! I'd be happy to submit a PR.

This wasn't entirely surprising to me, but nbytes currently doesn't return the right value for sparse data -- at least, I think nbytes should show the actual size in memory?

Since it uses size here:

https://github.com/pydata/xarray/blob/a0c71c1508f34345ad7eef244cdbbe224e031c1b/xarray/core/variable.py#L349

Rather than something like data.nnz, which of course only exists for sparse arrays... I'm not sure if there's a sparse flag or something, or whether you'd have to do a typecheck?

Minimal Complete Verifiable Example:

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

df = pd.DataFrame() df["x"] = np.repeat(np.random.rand(10_000), 10) df["y"] = np.repeat(np.random.rand(10_000), 10) df["time"] = np.tile(pd.date_range("2000-01-01", "2000-03-10", freq="W"), 10_000) df["rate"] = 10.0 df = df.set_index(["time", "y", "x"])

sparse_ds = xr.Dataset.from_dataframe(df, sparse=True) print(sparse_ds["rate"].nbytes) ```

python 8000000000 Anything else we need to know?:

Environment:

I was saving my dataset into a ZipStore -- apparently succesfully -- but then I couldn't reopen them.

The issue appears to be that a regular DirectoryStore behaves a little differently: it doesn't need to be closed, while a ZipStore.

(I'm not sure how this relates to #2586, the remarks there don't appear to be applicable anymore.)

MCVE Code Sample

This errors: ```python import xarray as xr import zarr

works as expected

ds = xr.Dataset({'foo': [2,3,4], 'bar': ('x', [1, 2]), 'baz': 3.14}) ds.to_zarr(zarr.DirectoryStore("test.zarr")) print(xr.open_zarr(zarr.DirectoryStore("test.zarr")))

error with ValueError "group not found at path ''

ds.to_zarr(zarr.ZipStore("test.zip")) print(xr.open_zarr(zarr.ZipStore("test.zip"))) ```

Calling close, or using with does the trick:

```python store = zarr.ZipStore("test2.zip") ds.to_zarr(store) store.close() print(xr.open_zarr(zarr.ZipStore("test2.zip")))

with zarr.ZipStore("test3.zip") as store: ds.to_zarr(store) print(xr.open_zarr(zarr.ZipStore("test3.zip"))) ```

Expected Output

I think it would be preferable to close the ZipStore in this case. But I might be missing something?

Problem Description

Because to_zarr works in this situation with a DirectoryStore, it's easy to assume a ZipStore will work similarly. However, I couldn't get it to read my data back in this case.

Versions

Disclaimer: I'm not a 100% sure this is necessarily the right place to discuss this. I think some of the things I propose could be broadly useful. At any rate, I appreciate feedback!

EDIT: to clarify, I don't think my original wording was entirely clear: I'm not suggesting adding this to xarray, but I'm thinking about a dataset accessor or a wrapping object in a separate package.

Introduction

Xarray is a great tool for structured data, but there isn't anything quite like it for unstructured meshes of convex cells. I'm coming from a hydrological background, and I believe many of my colleagues working with unstructured meshes could greatly benefit from xarray.

In terms of data models and formats, there are many unstructured mesh formats, I think UGRID is the most widely shared (outside of specific model formats) and it supports many aspects (e.g. data on faces, nodes, edges). Although Gridded exists and provides a number of features, I really need something that consumes and produces xarray Datasets (I don't want to go without e.g. dask, or deal with the netcdf4 library myself).

My most immediate need is some utilities for 2D unstructured meshes (y, x), and layered meshes (layer, y, x). This is also what UGRID is mostly used for now, as there don't seem to be many examples of the 3D unstructured mesh definitions in the wild. Sticking to 2D simplies things somewhat, so I'll ignore the third dimension for now -- but I don't see any reason why one couldn't generalize many ideas to 3D.

I currently use xarray a great deal with structured data, in dealing with "regular" netCDFs. In my thinking, it's possible to "mirror" a few methods which by and large behave similarly to xarray, using the UGRID data model for the mesh topology.

A simple UGRID example

To clarify what I'm talking about, this is what UGRID looks like in xarray, there's a Dataset with a dummy variable which describes in which variables the mesh topology is stored, in this case for a triangle and a quad.

python ds = xr.Dataset() ds["mesh2d"] = xr.DataArray( data=0, attrs={ "cf_role": "mesh_topology", "long_name": "Topology data of 2D mesh", "topology_dimension": 2, "node_coordinates": "node_x node_y", "face_node_connectivity": "face_nodes", } ) ds = ds.assign_coords( node_x=xr.DataArray( data=np.array([0.0, 10.0, 10.0, 0.0, 20.0, 20.0]), dims=["node"], ) ) ds = ds.assign_coords( node_y=xr.DataArray( data=np.array([0.0, 0.0, 10.0, 10.0, 0.0, 10.0]), dims=["node"], ) ) ds["face_nodes"] = xr.DataArray( data=np.array([ [0, 1, 2, 3], [1, 4, 5, -1] ]), coords={ "face_x": ("face", np.array([5.0, 15.0])), "face_y": ("face", np.array([5.0, 5.0])), }, dims=["face", "nmax_face"], attrs={ "cf_role": "face_node_connectivity", "long_name": "Vertex nodes of mesh faces (counterclockwise)", "start_index": 0, "_FillValue": -1, } )

It's only the topology that is special cased. A time dependent variable looks the same as it would otherwise, except it would have dimensions ("time", "face") rather than ("time", "y", "x").

UGRID netCDFs can in principle be read and written without issue. I would prefer to work with standard xarray Datasets like this if possible, because it means there's no need to touch any IO methods, as well as maintaining the full flexibility of xarray.

Some methods

I'm thinking of a few methods which behave similarly as the xarray methods except that they know about the mesh topology (in x and y) and the dependencies between the different variables. So, you could do:

```Python import xugrid # name of extension

selection_ds = ds.xugrid.sel(node_x=slice(-30.0, 30.0)) ```

And it would slice based on the x coordinate of the nodes, but update all the other variables to produce a consistent mesh (and renumber the face_nodes appriopriately). xugrid.isel would behave in the same way.

xugrid.plot would e.g. perform triangulation automatically (if the mesh isn't triangular already), and create a matplotlib.tri.Triangulation object, and so forth.

Regridding

I think one of the major needs for dealing with unstructured grids is being able to map data from one mesh to another, go from (GIS) raster data to unstructured, or go from unstructured model output to GIS raster data. For my use case, I often require area weighted methods, which requires computing overlap between source and destination meshes. I haven't really found a suitable package, e.g. xemsf looks very useful, but doesn't have quite the methods I'd want. Also troubling is that I can't really get it installed on Windows.

Existing GIS routines for vector data are far too slow (they have to take too many special cases into account, I think), see some benchmarks here.

I think numba provides an interesting opportunity here, not just for ease of development and distribution, but also because of JIT-compilation, the user can provide their own regridding methods. E.g.

```python import numba as nb import numpy as np

def median(values, weights): return np.nanpercentile(values, 50)

class Regridder: def init(self, source, destination): # compute and store weights # ...

# Use a closure to avoid numba's function overhead:
# https://numba.pydata.org/numba-doc/latest/user/faq.html#can-i-pass-a-function-as-an-argument-to-a-jitted-function
def _make_regrid(self, method):
    jitted_func = nb.njit(method)
    def regrid(src_indices, dst_indices, data, weights):
        out = np.full(dst_indices.size, np.nan)
        for src_index, dst_index in zip(src_indices, dst_indices):  # or something
            # accumulate values          
            out[dst_index] = jitted_func(values)
        return out
    return regrid

def regrid(source_data, method):
    self._regrid = self._make_regrid(method)
    # Cache the method in the object or something to avoid repeat compilation
    # provide default methods (e.g. min or max) as string or enumerators
    return self._regrid(self.src_indices, self.dst_indices, source_data, self.weights)

```

The best way to find overlapping cells seems via a cell tree, as implemented here, and requires searching for bounding boxes rather than points. I've ported the CellTree2D to numba, and added a bounding box search.

The serial version is a little bit slower than C++ for point searches, but with numba's prange (and enough processors) searches are a few times faster. (In retrospect, I could've just re-used the C++ tree rather than porting it, but numba does have some benefits...)

Interpolation on an unstructured grids can probably be handled by pyinterp, and can be made available via the regridder. (Scipy interpolate takes too much memory in my experience for large meshes.)

API proposal

```python class Regridder: # Basically same API as xemsf def init(self, source: xr.Dataset, destination: xr.Dataset) -> None: pass

def regrid(self, source_data: xr.DataArray, method: Union[str, Callable]) -> xr.Dataset:
    pass

xugrid.triangulate(self) -> xr.Dataset: xugrid.rasterize(self, resolution: Mapping) -> xr.Dataset: xugrid.reproject(self, dst_crs, src_crs) -> xr.Dataset: # just offload all coordinates to pyproj?

xugrid.sel(self, indexers: Mapping) -> xr.Dataset: xugrid.isel(self, indexers: Mapping) -> xr.Dataset: xugrid.plot(self) -> None: # calls triangulate if needed xugrid.plot.imshow(self) -> None: # maybe calls rasterize first, provides a quicker peek

xugrid.slice_vertical(self, linestring) -> xr.Dataset: # traces one or more rays through the mesh

xugrid.check_mesh(self) -> None: # check for convexity, all indices present, etc. xugrid.connectivity_matrix(self) -> scipy.sparse? xugrid.reverse_cuthill_mckee(self) -> xr.Dataset: # use scipy.sparse.csgraph xugrid.binary_erosion(self, iterations: int) -> xr.DataArray: xugrid.binary_dilation(self, iterations: int) -> xr.DataArray:

xugrid.meshfix(self) # call pymeshfix xugrid.to_pyvista(self) -> pyvista.UnstructuredGrid: # create a pyvista (vtk) mesh ```

Most of these methods require dealing with numpy arrays (or numpy arrays representing sparse arrays), so I think numba is very suitable for the algorithms (if they don't already exist).

For most methods, 3D meshes would be possible too. It requires some changes in e.g. the CellTree, and plotting would first require a slice_horizontal or something to get a plan view image.

Some methods might not be very needed, I'm basing it off of my impression what I currently use xarray for in combination with some other tools that assume structured data.

Resources

• Cell tree https://github.com/NOAA-ORR-ERD/cell_tree2d
• Numba KD tree: https://github.com/jackd/numba-neighbors
• Interpolation: https://github.com/CNES/pangeo-pyinterp
• UGRID conventions: https://ugrid-conventions.github.io/ugrid-conventions/
• Gridded: https://github.com/NOAA-ORR-ERD/gridded
• Triangulating polygons (could be useful for vector data -> mesh): https://pypi.org/project/mapbox-earcut/
• Pymeshfix: https://github.com/pyvista/pymeshfix

(Have to check for non-permissive / copyleft licenses...)

Thoughts?

Does this sound useful? Does it belong somewhere else / does it already exist?

(I'm aware a number e.g. the tree structures are available in vtk as well, but the Python API unfortunately doesn't expose them; having a numba version also makes changes a lot easier.)

Something that popped up: Some of the algorithms benefit from some persistent data structures (e.g. a cell tree or connectivity matrix), can these be maintained via an xarray-extension?

I'm assuming mesh topologies that fit within memory, that seems challenging enough for now...

Code Sample

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

time = [np.datetime64(datetime.strptime("10000101", "%Y%m%d"))] print(time[0]) print(np.dtype(time[0]))

da = xr.DataArray(time, ("time",), {"time":time}) print(da) Results in: 1000-01-01T00:00:00.000000 datetime64[us]

<xarray.DataArray (time: 1)> array(['2169-02-08T23:09:07.419103232'], dtype='datetime64[ns]') Coordinates: * time (time) datetime64[ns] 2169-02-08T23:09:07.419103232 ```

Problem description

I was happy to see cftime as default in the release notes for 0.11.0:

Xarray will now always use cftime.datetime objects, rather than by default trying to coerce them into np.datetime64[ns] objects. A CFTimeIndex will be used for indexing along time coordinates in these cases.

However, it seems that the DataArray constructor does not use cftime (yet?), and coerces to np.datetime64[ns] here: https://github.com/pydata/xarray/blob/0d6056e8816e3d367a64f36c7f1a5c4e1ce4ed4e/xarray/core/variable.py#L183-L189

Expected Output

I think I'd expect cftime.datetime in this case as well. Some coercion happens anyway as pandas timestamps are turned into np.datetime64[ns].

(But perhaps this was already on your radar, and am I just a little too eager!)

Output of xr.show_versions()

I've been running into some issues when using xr.full_like, when my other.data is a chunked dask array, and the fill_value is a numpy array.

Now, I just checked, full_like mentions only scalar in the signature. However, this is a very convenient way to get all the coordinates and dimensions attached to an array like this, so it feels like desirable functionality. And as I mention below, both numpy and dask function similary, taking much more than just scalars. https://xarray.pydata.org/en/stable/generated/xarray.full_like.html

MCVE Code Sample

python x = [1, 2, 3, 4] y = [1, 2, 3] da1 = xr.DataArray(dask.array.ones((3, 4), chunks=(1, 4)), {"y": y, "x": x}, ("y", "x")) da2 = xr.full_like(da1, np.ones((3, 4))) print(da2.values)

This results in an error: ValueError: could not broadcast input array from shape (1,3) into shape (1,4)

Expected Output

Expected is a DataArray with the dimensions and coords of other, and the numpy array of fill_value as its data.

Problem Description

The issue lies here: https://github.com/pydata/xarray/blob/2c77eb531b6689f9f1d2adbde0d8bf852f1f7362/xarray/core/common.py#L1420-L1436

Calling dask.array.full with the given number of chunks results in it trying to to apply the fill_value for every individual chunk.

As one would expect, if I set fill_value to the size of a single chunk it doesn't error: python da2 = xr.full_like(da1, np.ones((1, 4))) print(da2.values)

It does fail on a similarly chunked dask array (since it's applying it for every chunk): python da2 = xr.full_like(da1, dask.array.ones((3, 4))) print(da2.values)

The most obvious solution would be to force it down the np.full_like route, since all the values already exist in memory anyway. So maybe another type check does the trick. However, full() accepts quite a variety of arguments for the fill value (scalars, numpy arrays, lists, tuples, ranges). The dask docs mention only a scalar in the signature for dask.array.full: https://docs.dask.org/en/latest/array-api.html#dask.array.full As does numpy.full: https://docs.scipy.org/doc/numpy/reference/generated/numpy.full.html

However, in all cases, they still broadcast automatically... ```python a = np.full((2, 2), [1, 2]

array([[1, 2], [1, 2]]) ```

So kind of undefined behavior of a blocked full?

Versions

MCVE Code Sample

This came up in the following (simplified) case: ```python import numpy as np import xarray as xr

x = [1, 2, 3, 4] data = [1., 2., 3. , 4.] dims = ["x"] coords = {"x": x} da = xr.DataArray(data, coords, dims) da1 = da.sel(x=[1, 3]) da2 = da.sel(x=[2, 4]) da2["x"].values = da1["x"].values da3 = da2 - da1 print(da3) ```

In 0.14.1, this results in an empty DataArray da3: python <xarray.DataArray (x: 0)> array([], dtype=float64) Coordinates: * x (x) int64

0.14.0 (and before), no issues: python <xarray.DataArray (x: 2)> array([1., 1.]) Coordinates: * x (x) int32 1 3

Interestingly, the values of da["x"] are okay, print(da2["x"]") gives: array([1, 3])

However, the index has not been updated, print(da2.indexes["x"]): Int64Index([2, 4], dtype='int64', name='x')

Expected Output

I'm expecting output as in 0.14.0 and before. I've briefly stepped through the code for da2["x"].values = da1["x"].values; I don't see any changes there between 0.14.0 and 0.14.1, so I'm guessing it's due to some change in the indexes. Apparently the index no longer uses the same array?

Output of xr.show_versions()

Code Sample

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

def standardize(x): return (x - x.mean()) / x.std()

ds = xr.Dataset() ds["variable"] = xr.DataArray(np.random.rand(4,3,5), {"lat":np.arange(4), "lon":np.arange(3), "time":np.arange(5)}, ("lat", "lon", "time"), )

ds["id"] = xr.DataArray(np.arange(12.0).reshape((4,3)), {"lat": np.arange(4), "lon":np.arange(3)}, ("lat", "lon"), )

ds["id"].values[0,0] = np.nan

ds.groupby("id").apply(standardize) ```

Problem description

This results in an IndexError. This is mildly confusing, it took me a little while to figure out the NaN's were to blame. I'm guessing the NaN doesn't get filtered out everywhere.

The traceback: ```

IndexError Traceback (most recent call last) <ipython-input-2-267ba57bc264> in <module>() 15 ds["id"].values[0,0] = np.nan 16 ---> 17 ds.groupby("id").apply(standardize)

C:\Miniconda3\envs\main\lib\site-packages\xarray\core\groupby.py in apply(self, func, kwargs) 607 kwargs.pop('shortcut', None) # ignore shortcut if set (for now) 608 applied = (func(ds, kwargs) for ds in self._iter_grouped()) --> 609 return self._combine(applied) 610 611 def _combine(self, applied):

C:\Miniconda3\envs\main\lib\site-packages\xarray\core\groupby.py in _combine(self, applied) 614 coord, dim, positions = self._infer_concat_args(applied_example) 615 combined = concat(applied, dim) --> 616 combined = _maybe_reorder(combined, dim, positions) 617 if coord is not None: 618 combined[coord.name] = coord

C:\Miniconda3\envs\main\lib\site-packages\xarray\core\groupby.py in _maybe_reorder(xarray_obj, dim, positions) 428 429 def _maybe_reorder(xarray_obj, dim, positions): --> 430 order = _inverse_permutation_indices(positions) 431 432 if order is None:

C:\Miniconda3\envs\main\lib\site-packages\xarray\core\groupby.py in _inverse_permutation_indices(positions) 109 positions = [np.arange(sl.start, sl.stop, sl.step) for sl in positions] 110 --> 111 indices = nputils.inverse_permutation(np.concatenate(positions)) 112 return indices 113

C:\Miniconda3\envs\main\lib\site-packages\xarray\core\nputils.py in inverse_permutation(indices) 52 # use intp instead of int64 because of windows :( 53 inverse_permutation = np.empty(len(indices), dtype=np.intp) ---> 54 inverse_permutation[indices] = np.arange(len(indices), dtype=np.intp) 55 return inverse_permutation 56

IndexError: index 11 is out of bounds for axis 0 with size 11

```

Expected Output

My assumption was that it would throw out the values that fall within the NaN group, likepandas:

```python import pandas as pd import numpy as np

df = pd.DataFrame() df["var"] = np.random.rand(10) df["id"] = np.arange(10) df["id"].iloc[0:2] = np.nan df.groupby("id").mean() ```

Out: python var id 2.0 0.565366 3.0 0.744443 4.0 0.190983 5.0 0.196922 6.0 0.377282 7.0 0.141419 8.0 0.957526 9.0 0.207360

Output of xr.show_versions()

Code Sample, a copy-pastable example if possible

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

nrow, ncol = (4, 5) dx, dy = (1.0, -1.0) xmins = (0.0, 3.0, 3.0, 0.0) xmaxs = (5.0, 8.0, 8.0, 5.0) ymins = (0.0, 2.0, 0.0, 2.0) ymaxs = (4.0, 6.0, 4.0, 6.0) data = np.ones((nrow, ncol), dtype=np.float64)

das = [] for xmin, xmax, ymin, ymax in zip(xmins, xmaxs, ymins, ymaxs): kwargs = dict( name="example", dims=("y", "x"), coords={"y": np.arange(ymax, ymin, dy), "x": np.arange(xmin, xmax, dx)}, ) das.append(xr.DataArray(data, **kwargs))

xr.merge(das)

This won't flip the coordinate:

xr.merge([das[0]))

```

Problem description

Let's say I have a number of geospatial grids that I'd like to merge (for example, loaded with xr.open_rasterio). To quote https://www.perrygeo.com/python-affine-transforms.html

The typical geospatial coordinate reference system is defined on a cartesian plane with the 0,0 origin in the bottom left and X and Y increasing as you go up and to the right. But raster data, coming from its image processing origins, uses a different referencing system to access pixels. We refer to rows and columns with the 0,0 origin in the upper left and rows increase and you move down while the columns increase as you go right. Still a cartesian plane but not the same one.

xr.merge will alway return the result with ascending coordinates, which creates some issues / confusion later on if you try to write it back to a GDAL format, for example (I've been scratching my head for some time looking at upside-down .tifs).

Expected Output

I think the expected output for these geospatial grids is that; if you provide only DataArrays with positive dx, negative dy; that the merged result comes out with a positive dx and a negative dy as well.

When the DataArrays to merge are mixed in coordinate direction (some with ascending, some with descending coordinate values), defaulting to an ascending sort seems sensible.

A suggestion

I saw that the sort is occurring here, in pandas; and that there's a is_monotonic_decreasing property in pandas.core.indexes.base.Index

I think this could work (it solves my issue at least), in xarray.core.alignment python index = joiner(matching_indexes) if all( (matching_index.is_monotonic_decreasing for matching_index in matching_indexes) ): index = index[::-1] joined_indexes[dim] = index But I lack the knowledge to say whether this plays nice in all cases. And does index[::-1] return a view or a copy? (And does it matter?)

For reference this is what it looks like now: python if (any(not matching_indexes[0].equals(other) for other in matching_indexes[1:]) or dim in unlabeled_dim_sizes): if join == 'exact': raise ValueError( 'indexes along dimension {!r} are not equal' .format(dim)) index = joiner(matching_indexes) joined_indexes[dim] = index else: index = matching_indexes[0] It's also worth highlighting that the else branch causes, arguably, some inconsistency. If the indexes are equal, no reversion occurs.

Code Sample

```python import xarray as xr import numpy as np import rasterio from affine import Affine

def write_tif(path, epsg): nrow = 5 ncol = 8 values = 10.0 * np.random.rand(nrow, ncol)

profile = dict()
profile["crs"] = rasterio.crs.CRS.from_epsg(epsg)
profile["transform"] = Affine.translation(0.0, 0.0)
profile["driver"] = "GTiff"
profile["height"] = nrow
profile["width"] = ncol
profile["count"] = 1
profile["dtype"] = np.float64

with rasterio.Env():
    with rasterio.open(path, "w", **profile) as dst:
        dst.write(values, 1)

write_tif("basic.tif", epsg=28992) da = xr.open_rasterio("basic.tif").squeeze("band", drop=True) likevalues = np.empty((10, 16)) likecoords = {"y": np.arange(0.25, 5.0, 0.5), "x": np.arange(0.25, 8.0, 0.5)} likedims = ("y", "x") like = xr.DataArray(likevalues, likecoords, likedims)

newda = da.reindex_like(like) ```

Problem description

The code above executes without issues in xarray version 0.10.8. However, it results in a "shape-mismatch" error in 0.10.9:

``` Traceback (most recent call last):

File "<ipython-input-57-f099072e9750>", line 8, in <module> newda = da.reindex_like(like)

File "d:\bootsma\miniconda3\envs\main\lib\site-packages\xarray\core\dataarray.py", line 919, in reindex_like **indexers)

File "d:\bootsma\miniconda3\envs\main\lib\site-packages\xarray\core\dataarray.py", line 969, in reindex indexers=indexers, method=method, tolerance=tolerance, copy=copy)

File "d:\bootsma\miniconda3\envs\main\lib\site-packages\xarray\core\dataset.py", line 1924, in reindex tolerance, copy=copy)

File "d:\bootsma\miniconda3\envs\main\lib\site-packages\xarray\core\alignment.py", line 377, in reindex_variables new_var = var._getitem_with_mask(key)

File "d:\bootsma\miniconda3\envs\main\lib\site-packages\xarray\core\variable.py", line 655, in _getitem_with_mask data = duck_array_ops.where(mask, fill_value, data)

File "d:\bootsma\miniconda3\envs\main\lib\site-packages\xarray\core\duck_array_ops.py", line 183, in where return _where(condition, *as_shared_dtype([x, y]))

File "d:\bootsma\miniconda3\envs\main\lib\site-packages\xarray\core\duck_array_ops.py", line 115, in as_shared_dtype arrays = [asarray(x) for x in scalars_or_arrays]

File "d:\bootsma\miniconda3\envs\main\lib\site-packages\xarray\core\duck_array_ops.py", line 115, in <listcomp> arrays = [asarray(x) for x in scalars_or_arrays]

File "d:\bootsma\miniconda3\envs\main\lib\site-packages\xarray\core\duck_array_ops.py", line 110, in asarray return data if isinstance(data, dask_array_type) else np.asarray(data)

File "d:\bootsma\miniconda3\envs\main\lib\site-packages\numpy\core\numeric.py", line 492, in asarray return array(a, dtype, copy=False, order=order)

File "d:\bootsma\miniconda3\envs\main\lib\site-packages\xarray\core\indexing.py", line 624, in array self._ensure_cached()

File "d:\bootsma\miniconda3\envs\main\lib\site-packages\xarray\core\indexing.py", line 621, in _ensure_cached self.array = NumpyIndexingAdapter(np.asarray(self.array))

File "d:\bootsma\miniconda3\envs\main\lib\site-packages\numpy\core\numeric.py", line 492, in asarray return array(a, dtype, copy=False, order=order)

File "d:\bootsma\miniconda3\envs\main\lib\site-packages\xarray\core\indexing.py", line 602, in array return np.asarray(self.array, dtype=dtype)

File "d:\bootsma\miniconda3\envs\main\lib\site-packages\numpy\core\numeric.py", line 492, in asarray return array(a, dtype, copy=False, order=order)

File "d:\bootsma\miniconda3\envs\main\lib\site-packages\xarray\core\indexing.py", line 508, in array return np.asarray(array[self.key], dtype=None)

File "d:\bootsma\miniconda3\envs\main\lib\site-packages\xarray\backends\rasterio_.py", line 119, in getitem key, self.shape, indexing.IndexingSupport.OUTER, self._getitem)

File "d:\bootsma\miniconda3\envs\main\lib\site-packages\xarray\core\indexing.py", line 776, in explicit_indexing_adapter result = raw_indexing_method(raw_key.tuple)

File "d:\bootsma\miniconda3\envs\main\lib\site-packages\xarray\backends\rasterio_.py", line 115, in _getitem return out[np_inds]

IndexError: shape mismatch: indexing arrays could not be broadcast together with shapes (10,) (16,)

```

I only get it to occur when loading a raster using open_rasterio(). Defining da directly does not lead to issues, using either numpy or dask.array.

Python values = 10.0 * np.random.rand(1, 5, 8) coords = {"band": [1.0], "y": np.arange(0.5, 5.0, 1.0), "x": np.arange(0.5, 8.0, 1.0)} dims = ("band", "y", "x") da = xr.DataArray(values, coords, dims) da = da.squeeze("band", drop=True) Python import dask.array values = 10.0 * dask.array.random.random((1, 5, 8), (1, 5, 8)) coords = {"band": [1.0], "y": np.arange(0.5, 5.0, 1.0), "x": np.arange(0.5, 8.0, 1.0)} dims = ("band", "y", "x") da = xr.DataArray(values, coords, dims) da = da.squeeze("band", drop=True)

The "shape mismatch" makes me think that maybe the squeeze() isn't executed properly or in time. This works: ```Python write_tif("basic.tif", epsg=28992) da = xr.open_rasterio("basic.tif") likevalues = np.empty((1, 10, 16)) likecoords = {"band":[1], "y": np.arange(0.25, 5.0, 0.5), "x": np.arange(0.25, 8.0, 0.5)} likedims = ("band", "y", "x") like = xr.DataArray(likevalues, likecoords, likedims)

newda = da.reindex_like(like) ```

So it looks to me like some lazy evaluation is going awry when open_rasterio is involved, in this case. Anything which forces a compute() before reindex_like in this case seems to make the problem go away, like calling da in the console, print(da) after opening the .tif, etc.

```Python write_tif("basic.tif", epsg=28992) da = xr.open_rasterio("basic.tif").compute().squeeze("band", drop=True) likevalues = np.empty((10, 16)) likecoords = {"y": np.arange(0.25, 5.0, 0.5), "x": np.arange(0.25, 8.0, 0.5)} likedims = ("y", "x") like = xr.DataArray(likevalues, likecoords, likedims)

newda = da.reindex_like(like) ```

Interestingly, I can't easily inspect it, because inspecting makes the problem go away!

Expected Output

A succesfully reindexed DataArray.

Output of xr.show_versions()