Should xpystac attempt to stack?

[jsignell]

The more I have been thinking about this library the more I am wondering if it is maybe too many layers of indirection.

I do think there is value in making it easier to open assets.

Compare:

without xpystac

import xarray as xr

xr.open_zarr(
    asset.href,
    **asset.extra_fields["xarray:open_kwargs"],
    storage_options=asset.extra_fields["xarray:storage_options"]
)

with xpystac

import xarray as xr

xr.open_dataset(asset)

But I am less confident about the stacking. There is just so much going on in odc-stac and stackstac and all that needs to be configured carefully. I'm wondering if it might make more sense to return an iterable of datasets or a xr.DataTree if the input is a collection or an item collection and leave it to the user to figure out how they want to stack things....

For an item containing assets where each asset is a single-band COG I think open_mfdataset makes a lot of sense, but then you'd have to use that internally since the input wouldn't strictly be a list. So at that point you are kind of just writing another version of a stacking library...

[keewis]

as mentioned in #45 I do believe that the fewer things the backend for open_dataset does the better (not the least because that gets rid of the weird behavior reported in #45), but that doesn't have to keep you from adding helpers that users would be able to call explicitly if they'd want to stack things.

For example:

# just open the data
xr.open_dataset(asset, engine="stac", ...)  # as usual
xr.open_dataset(item, engine="stac", ...)  # if there's only a single data asset
xr.open_datatree(item, engine="stac", ...)
xr.open_datatree(item_collection, engine="stac", ...)

# open + stack
xr.open_dataset(item, engine="stac", ...).pipe(xpystac.stack, ...)  # (or .stac.stack or whatever)
xr.open_datatree(item_collection, engine="stac", ...).pipe(xpystac.stack, ...)

Regarding using open_mfdataset: this is basically a bunch of open_dataset calls (optionally parallelized using dask) followed by a call to combine_*, so it may be possible to just add this as another helper function?

Alternatively (and maybe preferably), stackstac / odc-stac could be taught to work with the objects returned by xpystac.

[jsignell]

Yeah I like these ideas a lot. This ticket came directly out of #45.