ux.open_dataset and other functions failing with xarray==2026.4.0

[VeckoTheGecko]

Version

main

How did you install UXarray?

Source

What happened?

Uxarray constructors are now failing with TypeError('Passing a Dataset as data_varsto the Dataset constructor is not supported. Useds.copy() to create a copy of a Dataset.') . Due to pydata/xarray#11139 in https://github.com/pydata/xarray/releases/tag/v2026.04.0

What did you expect to happen?

No failures

Can you provide a MCVE to repoduce the bug?

You can use the test suite in this project

[erogluorhan]

Thanks for reporting, @VeckoTheGecko ! We'll be looking into this right away (we'd already found this in the upstream CI of another tool of ours but missed that xarray would release that soon)

[VeckoTheGecko]

Haven't tested it, but I was thinking you can probably call ds.to_dict() in UxDataset.init before passing it through to xr.Dataset.init.

That's probably the most straightforward fix

[rajeeja]

Haven't tested it, but I was thinking you can probably call ds.to_dict() in UxDataset.init before passing it through to xr.Dataset.init.

That's probably the most straightforward fix

Hi Nick, ds.to_dict() might work, but it would be heavy conversion and could lose some fidelity compared to staying in xarray objects directly. We might want to try handling xr.Dataset inputs explicitly in UxDataset.init

[erogluorhan]

Haven't tested it, but I was thinking you can probably call ds.to_dict() in UxDataset.init before passing it through to xr.Dataset.init.
That's probably the most straightforward fix

Hi Nick, ds.to_dict() might work, but it would be heavy conversion and could lose some fidelity compared to staying in xarray objects directly. We might want to try handling xr.Dataset inputs explicitly in UxDataset.init

I agree with this and am currently looking into it

[erogluorhan]

I agree with this and am currently looking into it

Actually, now I don't think this maybe heavy conversions. I believe we will need to check in our constructor if the first arg is a xr.Dataset, and if so, we use its data_vars instead, and they will need to be used as a dict, right? Also, while doing that, handle attrs and coords properly is a must I believe.

Also, this whole conversion shouldn't introduce any issues with scalability I think (i.e. if the ds is chunked etc.)

Something like this:

if args and first arg is xr.Dataset
 convert first arg to xr.Dataset.data_vars
 Retain `coords` and `attrs`

super().__init__(*args, **kwargs)

Thoughts @rajeeja ?

[rajeeja]

I agree with this and am currently looking into it

Actually, now I don't think this maybe heavy conversions. I believe we will need to check in our constructor if the first arg is a xr.Dataset, and if so, we use its data_vars instead, and they will need to be used as a dict, right? Also, while doing that, handle attrs and coords properly is a must I believe.

Also, this whole conversion shouldn't introduce any issues with scalability I think (i.e. if the ds is chunked etc.)

Something like this:

if args and first arg is xr.Dataset
 convert first arg to xr.Dataset.data_vars
 Retain `coords` and `attrs`

super().__init__(*args, **kwargs)

Thoughts @rajeeja ?

Right direction. Yes, I think this is the cleaner fix: detect xr.Dataset as the first arg in UxDataset.init, pass its data_vars plus preserved coords/attrs explicitly into xr.Dataset.init, and avoid to_dict() so we stay in xarray-native objects and preserve chunked/lazy behavior.

[erogluorhan]

@rajeeja @VeckoTheGecko please see PR #1492