Using Dataset Constructor to create a copy drops attrs

[Reshief]

What happened?

I encountered an unexpected behavior while using xarray, which I would consider to be a bug.
Similar to other types/classes, I attempted to create a copy of a dataset by using the constructor of the xarray.Dataset class.
For some time, I did not notice any issues, but just today I realized that this creates a new dataset that holds all relevant information except for the .attrs of the dataset.
Specifically, I checked that the new dataset copies over the data vars from the provided dataset and reconstructs the coordinates, but completely loses the attributes.

I am using xarray version "2025.12.0" for my tests, but the relevant code in the __init__() function of xarray.Dataset at [datataset.py:lines 378-410](https://github.com/pydata/xarray/blob/main/xarray/core/dataset.py) has not changed.

Looking at the code in the constructor, the first argument to the constructor is used as a source for variables in the new dataset and then the list of coordinates is reconstructed.
There is a special treatment for the coord parameter being a Dataset, but not for the data_vars parameter.

What did you expect to happen?

Given that the constructor for all my intents and purposes creates a shallow copy of the Dataset, I expected it to create a copy of the attributes as well.
I am aware of the .copy() function, but I would expect the call to xr.Dataset(ds) to effectively work as a copy instructor thus copying the attributes from the input as well, to fail the initialization or to output a warning that the result may be unexpected.

Minimal Complete Verifiable Example

# /// script
# requires-python = ">=3.11"
# dependencies = [
#   "xarray[complete]@git+https://github.com/pydata/xarray.git@main",
# ]
# ///
#
# This script automatically imports the development branch of xarray to check for issues.
# Please delete this header if you have _not_ tested this script with `uv run`!

import xarray as xr
xr.show_versions()

# your reproducer code ...import xarray as xr

ds = xr.Dataset()
ds.attrs['magic_key'] = True

assert 'magic_key' in ds.attrs # This is successful

# Create a copy:
ds2 = xr.Dataset(ds)
assert 'magic_key' in ds2.attrs # This is unsuccessful, the attrs are completely empty

Steps to reproduce

1. Create a dataset ds = xr.Dataset()
2. Set an attribute on the dataset ds.attrs['magic_key']=True
3. Create a new dataset from the old dataset ds2 = xr.Dataset(ds)
4. The new dataset has lost all attributes.

MVCE confirmation

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

Relevant log output

Anything else we need to know?

I would suggest a change to the constructor that it can effectively be used as a shallow copy of the dataset, e.g. such that it can used in copy operations of the form type(ds)(ds) similar to other datatypes.

For this, a simple change to line 398 in my version or line 404 in the current (at time of writing) commit to the main branch of the constructor would suffice:
self._attrs = dict(attrs) if attrs else None
could be replaced by
self._attrs = dict(attrs) if attrs else dict(data_vars.attrs) if isinstance(data_vars, Dataset) else None

If there is a reason for why the constructor should not be able to operate similar to a copy constructor, I would suggest adding a warning if a dataset is supplied with some attributes in ds.attrs set and no attrs parameter being provided because the current behavior may be unexpected.

Environment

Details

INSTALLED VERSIONS ------------------ commit: None python: 3.12.2 (tags/v3.12.2:6abddd9, Feb 6 2024, 21:26:36) [MSC v.1937 64 bit (AMD64)] python-bits: 64 OS: Windows OS-release: 10 machine: AMD64 processor: AMD64 Family 25 Model 33 Stepping 2, AuthenticAMD byteorder: little LC_ALL: None LANG: None LOCALE: ('de_DE', 'cp1252') libhdf5: 1.14.6 libnetcdf: None

xarray: 2025.12.0
pandas: 2.3.3
numpy: 2.3.5
scipy: 1.16.3
netCDF4: None
pydap: None
h5netcdf: 1.7.3
h5py: 3.15.1
zarr: None
cftime: None
nc_time_axis: None
iris: None
bottleneck: None
dask: None
distributed: None
matplotlib: 3.10.8
cartopy: None
seaborn: 0.13.2
numbagg: None
fsspec: 2025.12.0
cupy: None
pint: None
sparse: None
flox: None
numpy_groupies: None
setuptools: 80.9.0
pip: 25.3
conda: None
pytest: 9.0.2
mypy: 1.18.2
IPython: 9.8.0
sphinx: 8.2.3

[keewis]

I don't think we ever designed Dataset to support copying that way. The reason this sort of works is accidental: Dataset inherits from Mapping (i.e. a dict), and the data_vars parameter is supposed to be a dict.

I suppose we should error in this case? If unpacking the data variables was intentional, it is not too unreasonable to ask a user to pass ds.data_vars instead.

For reference, the intended API for copying is ds.copy(deep=False) for a shallow copy and ds.copy(deep=True) for a deep copy.

[Reshief]

I don't mind forcing the use of ds.copy(), I just think it is a bit too easy to shoot oneself in the foot with the current way of operation without a warning.

Then probably a

if isinstance(data_vars, Dataset) and coords is None and attrs is None:
    logging.warning("You have provided a dataset to the data_vars argument of the Dataset constructor. If you intended to create a copy of the provided dataset, please use ds.copy()")

or something along those lines should suffice?

[keewis]

I actually think we should raise an exception instead (but not sure if we need a deprecation cycle for this). The message itself looks good to me, although I would mention Dataset(ds) explicitly.

[kkollsga]

I've opened #11139 to address this. It raises TypeError when a Dataset is passed as data_vars, with a message pointing to ds.copy():

>>> xr.Dataset(ds)
TypeError: Passing a Dataset as `data_vars` to the Dataset constructor is
not supported. Use `ds.copy()` to create a copy of a Dataset.

Went with a direct error, but happy to switch to a deprecation warning if that's preferred.

[headtr1ck]

I feel like we could just implement the copy constructor instead of raising an error?

[kkollsga]

Thats a good point, I think you're right — making it work correctly is better than blocking it. Since dict(d) and pd.DataFrame(df) both produce copies, Dataset(ds) should too.

I can rework #11139 to do this instead. Something like:

if isinstance(data_vars, Dataset):
    if coords is None:
        coords = data_vars.coords
    if attrs is None:
        attrs = data_vars.attrs
    data_vars = data_vars.data_vars

This way partial overrides like Dataset(ds, attrs={"new": True}) also work naturally.