Danila

Hey @Chengwei94, Do you think there's more information you could provide so that we can figure out what's up? Is this about the latent space that you plot (UMAP) not...

I wonder if apart from the modifiers (#659) we should consider saving original object class as an attribute. Would one expect to restore the same object class that was saved?...

I guess it would be nice to include a «global switch» for the backend when reading files then, i.e. ```py adata = anndata.read("adata.h5ad", matrix_backend="jax") ``` That would include [sparse matrices](https://jax.readthedocs.io/en/latest/jax.experimental.sparse.html).

I am not sure how much of a problem this is in practice though. Of course we'll have to test it on different setups but I expect it to work...

> Thanks for opening this! Really excited to get this added. > > # Examples > This is more a point for tracking of this PR. Could you share examples...

Not sure if related to this or if we should open a new one but it seems that the current writer fails if a `DaskArray` is a layer but doesn't...

@ivirshup, one more consideration is adopting `"encoder"` and `"encoder-version"` [that mudata has](https://github.com/scverse/mudata/blob/48db79368f052fec074f039ed3655d18e6b8959f/mudata/_core/io.py#L86).

We would also need this for single nucleotide resolution data. Tagging @mffrank here. @ivirshup, maybe one way to do that would be to only allow a range index, this way...

Thanks, @ivirshup, I was just going to write that it seems like one can even use RangeIndex already, and we just need to figure out (a) places where the unnecessary...

The correct way would probably be checking if it's a mapping and converting it to a dict before writing it on disk. ```py from collections import OrderedDict od = OrderedDict(a...