What is your issue?

I'm trying to perform 2D interpolation on a large 3D array that is heavily chunked along the interpolation dimensions and not the third dimension. The application could be extracting a timeseries from a reanalysis dataset chunked in space but not time, to compare to observed station data with more precise coordinates.

I use the advanced interpolation method as described in the documentation, with the interpolation coordinates specified by DataArray's with a shared dimension like so:

```python %load_ext memory_profiler import numpy as np import dask.array as da import xarray as xr

Synthetic dataset chunked in the two interpolation dimensions

nt = 40000 nx = 200 ny = 200 ds = xr.Dataset( data_vars = { 'foo':( ('t', 'x', 'y'), da.random.random(size=(nt, nx, ny), chunks=(-1, 10, 10))), }, coords = { 't': np.linspace(0, 1, nt), 'x': np.linspace(0, 1, nx), 'y': np.linspace(0, 1, ny), } )

Interpolate to some random 2D locations

ni = 10 xx = xr.DataArray(np.random.random(ni), dims='z', name='x') yy = xr.DataArray(np.random.random(ni), dims='z', name='y') interpolated = ds.foo.interp(x=xx, y=yy) %memit interpolated.compute() ```

With just 10 interpolation points, this example calculation uses about 1.5 * ds.nbytes of memory, and saturates around 2 * ds.nbytes by about 100 interpolation points.

This could definitely work better, as each interpolated point usually only requires a single chunk of the input dataset, and at most 4 if it is right on the corner of a chunk. For example we can instead do it in a loop and get very reasonable memory usage, but this isn't very scalable:

python interpolated = [] for n in range(ni): interpolated.append(ds.foo.interp(x=xx.isel(z=n), y=yy.isel(z=n))) interpolated = xr.concat(interpolated, dim='z') %memit interpolated.compute()

I tried adding a .chunk({'z':1}) to the interpolation coordinates but this doesn't help. We can also do .sel(x=xx, y=yy, method='nearest') with very good performance.

Any tips to make this calculation work better with existing options, or otherwise ways we might improve the interp method to handle this case? Given the performance behavior I'm guessing we may be doing sequntial interpolation for the dimensions, basically an interp1d call for all the xx points and from there another to the yy points, which for even a small number of points would require nearly all chunks to be loaded in. But I haven't explored the code enough yet to understand the details.

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

Small follow up to #8877. If we're going to drop the indices anyways for region writes, we may as well not raise if they are still in the dataset. This makes the user experience of region writes simpler:

```python ds = xr.tutorial.open_dataset("air_temperature") ds.to_zarr("test.zarr") region = {"time": slice(0, 10)}

This fails unless we remember to ds.drop_vars(["lat", "lon"])

ds.isel(**region).to_zarr("test.zarr", region=region) ```

I find this annoying because I often have a dataset with a bunch of unrelated indexes and have to remember which ones to drop, or use some verbose set logic. I thought #8877 might have already done this, but not quite. By just reordering the point at which we drop indices, we can now skip this. We still raise if data vars are passed that don't overlap with the region.

cc @dcherian

What is your issue?

```python import dask.array import xarray as xr

da1 = xr.DataArray(dask.array.ones((3,3), chunks=(1, 1)), dims=["x", "y"]) da2 = xr.DataArray(dask.array.ones((3,), chunks=(1,)), dims=["x"])

da2.broadcast_like(da1).chunksizes

Frozen({'x': (1, 1, 1), 'y': (3,)}) ```

Was surprised to not find any other issues around this. Feels like a major limitation of the method for a lot of use cases. Is there an easy hack around this?

Is your feature request related to a problem?

As far as I know, there is no straightforward API in xarray to do what np.sort or pandas.sort_values does. We have DataArray.sortby("x"), which will sort the array according to the coordinate itself. But if instead you want to sort the values of the array to be monotonic, you're on your own. There are probably a lot of ways we could do this, but I ended up with the couple line solution below after a little trial and error.

Describe the solution you'd like

Would there be interest in implementing a Dataset/DataArray.sort_values(dim="x") method?

Note: this 1D example is not really relevant, see the 2D version and more obvious implementation in comments below for what I really want.

python def sort_values(self, dim: str): sort_idx = self.argsort(axis=self.get_axis_num(dim)).drop_vars(dim) return self.isel({dim: sort_idx}).drop_vars(dim).assign_coords({dim: self[dim]})

The goal is to handle arrays that we want to monotize like so:

```python da = xr.DataArray([1, 3, 2, 4], coords={"x": [1, 2, 3, 4]}) da.sort_values("x")

<xarray.DataArray (x: 4)> array([1, 2, 3, 4]) Coordinates: * x (x) int64 1 2 3 4 ```

In addition to sortby which can deal with an array that is just unordered according to the coordinate:

```python da = xr.DataArray([1, 3, 2, 4], coords={"x": [1, 3, 2, 4]}) da.sortby("x")

<xarray.DataArray (x: 4)> array([1, 2, 3, 4]) Coordinates: * x (x) int64 1 2 3 4 ```

Describe alternatives you've considered

I don't know if argsort is dask-enabled (the docs just point to the numpy function). Is there a more intelligent way to implement this with apply_ufunc and something else? I assume chunking in the sort dimension would be problematic.

Additional context

Some past related threads on this topic: https://github.com/pydata/xarray/issues/3957 https://stackoverflow.com/questions/64518239/sorting-dataset-along-axis-with-dask

What is your issue?

Follow on from #6891 and #6978 to discuss how we could homogenize the passing of keyword arguments to wrapped external functions across xarray methods.

There are quite a few methods like this where we are ultimately passing data to numpy, scipy, or some other library and want the option to send variable length kwargs to that underlying function. There are two different ways of doing this today:

1. xarray method accepts flexible **kwargs so these can be written directly in the xarray call
2. xarray method accepts a single dict kwargs (sometimes named differently) and passes these along in expanded form via **kwargs

I could only find a few examples of the latter:

• Dataset.interp, which takes kwargs
• Dataset.curvefit, which takes kwargs (although the docstring is wrong here)
• xr.apply_ufunc, which takes kwargs passed to func and dask_gufunc_kwargs passed to dask.array.apply_gufunc
• xr.open_dataset, which takes either **kwargs or backend_kwargs and merges the two

Allowing direct passage with **kwargs seems nice from a user perspective. But, this could occasionally be problematic, for example in the Dataset.interp case where this method also accepts the kwarg form of coords with **coords_kwargs. There are many methods like this that use **indexers_kwargs or **chunks_kwargs with either_dict_or_kwargs but don't happen to wrap external functions.

What is your issue?

I make use of fsspec to quickly open netcdf files in the cloud and pull out slices of data without needing to read the entire file. Quick and dirty is just ds = xr.open_dataset(fs.open("gs://...")).

This works great, in that a many GB file can be lazy-loaded as a dataset in a few hundred milliseconds, by only parsing the netcdf headers with under-the-hood byte range requests. But, only if the netcdf is written from dask-backed arrays. Somehow, writing from numpy-backed arrays produces a different netcdf that requires reading deeper into the file to parse as a dataset.

I spent some time digging into the backends and see xarray is ultimately passing off the store write to dask.array here. A look at ncdump and Dataset.encoding didn't reveal any obvious differences between these files, but there is clearly something. Anyone know why the straight xarray store methods would produce a different netcdf structure, despite the underlying data and encoding being identical?

This should work as an MCVE: ```python import os import string import fsspec import numpy as np import xarray as xr

fs = fsspec.filesystem("gs") bucket = "gs://<your-bucket>"

create a ~160MB dataset with 20 variables

variables = {v: (["x", "y"], np.random.random(size=(1000, 1000))) for v in string.ascii_letters[:20]} ds = xr.Dataset(variables)

Save one version from numpy backed arrays and one from dask backed arrays

ds.compute().to_netcdf("numpy.nc") ds.chunk().to_netcdf("dask.nc")

Copy these to a bucket of your choice

fs.put("numpy.nc", bucket) fs.put("dask.nc", bucket) ```

Then time reading in these files as datasets with fsspec: ```python %timeit xr.open_dataset(fs.open(os.path.join(bucket, "numpy.nc")))

2.15 s ± 40.2 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

```

```python %timeit xr.open_dataset(fs.open(os.path.join(bucket, "dask.nc")))

187 ms ± 26.1 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

```

• [x] Closes #6891
• [x] Tests added
• [x] User visible changes (including notable bug fixes) are documented in whats-new.rst

• [x] Closes #4554
• [x] Closes #5629
• [x] Closes #5644
• [ ] Tests added
• [x] Passes pre-commit run --all-files
• [ ] User visible changes (including notable bug fixes) are documented in whats-new.rst

Reimplement polyfit using apply_ufunc rather than dask.array.linalg.lstsq. This should solve a number of issues with memory usage and chunking that were reported on the current version of polyfit. The main downside is that variables chunked along the fitting dimension cannot be handled with this approach.

There is a bunch of fiddly code here for handling the differing outputs from np.polyfit depending on the values of the full and cov args. Depending on the performance implications, we could simplify some by keeping these in apply_ufunc and dropping later. Much of this parsing would still be required though, because the only way to get the covariances is to set cov=True, full=False.

A few minor departures from the previous implementation: 1. The rank and singular_values diagnostic variables returned by np.polyfit are now returned on a pointwise basis, since these can change depending on skipped nans. np.polyfit also returns the rcond used for each fit which I've included here. 2. As mentioned above, this breaks fitting done along a chunked dimension. To avoid regression, we could set allow_rechunk=True and warn about memory implications. 3. Changed default skipna=True, since the previous behavior seemed to be a limitation of the computational method. 4. For consistency with the previous version, I included a transpose operation to put degree as the first dimension. This is arbitrary though, and actually the opposite of how curvefit returns ordering. So we could match up with curvefit but it would be breaking for polyfit.

No new tests have been added since the previous suite was fairly comprehensive. Would be great to get some performance reports on real-world data such as the climate model detrending application in #5629.