Sam Levang

I don't have an MCVE but can say I've also run into this before. The hack to run `isdir` twice does work, so I've used that since I first saw...

I have a use case with a tree where all the nodes share the initial i/o ops and many intermediate compute tasks, and noticed how slow this is to map...

This would be a huge improvement, would love to see this get implemented. This is the only impediment to reducing my `dvc.yaml` lines by about 5x and making everything much...

Yeah I think that's a reasonable distinction. Totally agree `mode="a"` is very useful, my point is just that we try to push a lot of different (and potentially surprising) functionality...

Your mention of merge gets me thinking in terms of the standard xarray ops every knows well. `mode="a"` basically covers what `xr.merge` does, and between `compat`, `join`, and `combine_attrs` you...

Looks like things are better but still very slow. The example in the OP now takes just over a minute on latest versions writing to GCS. `DataTree.to_zarr` hasn't changed so...

Sorry @dcherian I haven't had a chance to look at this in detail yet, but overall it seems really nice. One random thought on the API: is there a way...

> Alternatively perhaps https://github.com/pydata/xarray/pull/8877 is too liberal and we should only drop indexes that are in region. Perhaps. Curious to hear other perspectives on this, and whether there are applications...

Or what about an opt-in kwarg, `clean_for_region` (someone please come up with a better name if we go this route)? If provided, we assume the user knows what they want...

I agree that #8460 should generally be the recommended way once that's merged. I think I do have some patterns where that doesn't totally apply though and I'm still going...