Is the .attrs slot a stable feature?

[LucaMarconato]

Background (you can jump to "Question" for a "TLDR;")

Hi! In a library I co-develop (https://github.com/scverse/spatialdata/) we use a series of classes that have a .attrs attribute, which we use to store lightweight metadata: xarray.DataArray, xarray.DataTree, geopandas.GeoDataFrame, dask.dataframe.DataFrame and anndata.AnnData (which has a slot called .uns, but it's essentially working like .attrs).

Unfortunately, the support of .attrs in pandas, geopandas and dask is spotty, for example:

• the .attrs semantic in pandas is not clear, and the attribute almost got deprecated DEPR: attrs pandas-dev/pandas#52166
• the serialization in .attrs in geopandas has bugs BUG: to_parquet does not seem to preserve attrs geopandas/geopandas#3320
• dask removed .attrs altogether: Dask 2024.5.1 removed .attrs dask/dask#11146

Also, we saw that polars doesn't plan to implement a .attrs equivalent: pola-rs/polars#14815.

All the above is creating problems to us and to our users.

Since we already use xarray data structures in our projects, and since xarray.Dataset is general enough to model also some of the data that we model with the other classes (or conversions are possible, for instance here is (xarray.Dataset -> geopandas.GeoDataFrame) a solution from our end could be to increasingly rely on xarray classes. This would not just be motivated by having .attrs, but also by the fact we could have multiscale support via xarray.DataTree for the objects that currently we don't represent as multiscale.

Question

So my question, how stable is .attrs in xarray? Do you see any risk of it being removed like it happened to dask? Or is there any problem in specifying its semantic, like it happens with pandas or polars?

Thank you!

[jhamman]

.attrs is quite core to Xarray's data model and, as far as I now, we have had no serious discussions of removing it.

The one issue to keep in mind is how Xarray propagates .attrs as objects are used for computation. There's a good issue describing the challenges here: #1614

[LucaMarconato]

Thank you for your answer and for linking to the discussion on the semantic, sounds great to me!