Consider renaming DataTree.ds and/or the data argument in DataTree.__init__

[shoyer]

What is your issue?

It is not inherently clear to that DataTree(data=...) and DataTree.ds refer to the same thing, or that it refers specifically to the dataset associated with the node. We also do not use the abbreviation .ds anywhere else in Xarray.

Instead, what about calling this argument node? This makes the data model of DataTree much more transparent, it's just a node and children, with optional name and parent.

Example usage:

tree = DataTree(node=Dataset(...))
tree.node = Dataset(...)

CC @flamingbear @TomNicholas @eni-awowale @owenlittlejohns

[TomNicholas]

It is not inherently clear to that DataTree(data=...) and DataTree.ds refer to the same thing, or that it refers specifically to the dataset associated with the node.

That's a fair criticism.

We also do not use the abbreviation .ds anywhere else in Xarray.

We don't in the API, but we do everywhere in the examples, e.g. ds = open_dataset(...). I agree that the current situation of .ds and .to_dataset() is not ideal though.

This makes the data model of DataTree much more transparent, it's just a node and children, with optional name and parent.

A node and zero or more children, to be really pedantic. As the type of children is not Optional.

Instead, what about calling this argument node?

This is fairly neat idea. The downsides are that "node" doesn't really imply data/dataset-like to me, and that currently we have a lot of explanation (and probably variable names / comments in code too) that uses "node" to mean "DataTree object at a specific location in the full tree", not "dataset data stored at at particular point in the tree".

Essentially if a tree "node" is no longer the thing that stores data, but is the data itself, we need a new word for what used to be called the "node". (In computer science lingo I think we would call each DataTree object a "vertex".)

In computer science terms a DataTree is roughly a "rooted (ordered) labeled tree whose vertices each contain a Dataset, with alignment enforced between Datasets".

[TomNicholas]

I think I'm way too deep in the weeds here, and would appreciate perspective from others in @pydata/xarray on what they think the most intuitive names for these parts of the tree would be. Now, before we release in main, is the only time we can ever change this really.

[shoyer]

Did we ever write the doc on DataTree terminology? That could be a good place to sort this all out.

[TomNicholas]

There is a short, outdated terminology page on the xarray-contrib/datatree docs, but I think that's it. Being able to make separate PRs for these docs additions before releasing I think is a reason to make a dedicated (temporary) branch for the datatree docs (see #9033 (comment)).

[dcherian]

How about simply .dataset? DataTree(dataset=...) seems pretty readable to me.

PS: tree.this struck me as nice but DataTree(this=...) is nonsensical haha.

[TomNicholas]

How about simply .dataset?

That makes the intent quite clear, but

1. It's a lot longer than .ds (or .node), and we expect this method to be called a lot,

2. We also have .to_dataset(), which is different from .ds in that the former returns a mutable copy of the data and the latter returns an immutable view.

tree.this

👎 to this lol

[shoyer]

Thanks @dcherian for the suggestions.

In the DataTree meeting today, we agreed to switch both names to .dataset. We'll leave the .ds property around for now as a (soft deprecated) alias.

[TomNicholas]

Closed by #9476