Tom Nicholas

Zarr v3 currently does not, but will do after the extension dtypes are available. Which is why we in virtualizarr have been eagerly awaiting adding these data types back in...

That's a very good point @keewis - I forgot about the `rename_paths` method. Yeah having `virtualizarr` instead of `virtualize` would be a minor improvement, as the verb `virtualize` is misleading...

We could also just use `.vz` as a shorthand for `virtualizarr`, to correspond with a convention of importing top-level functions via `import virtualizarr as vz`.

So I think maybe we can virtualize Parquet. Crucially (from the [parquet docs](https://parquet.apache.org/docs/concepts/)): > - Column chunk: A chunk of the data for a particular column. They live in a...

ah. good point. damn.

> That functionality wasn't in obstore when I started ManifestStore IIRC but now that it's released we should definitely use it. Ohh cool. > There are likely a large number...

Omg are you making fsspec but with an actual specification!?! (well a protocol) (See https://github.com/fsspec/filesystem_spec/issues/1446)

Thinking more about how we might prototype this idea in Cubed today... In https://github.com/cubed-dev/cubed/issues/187 @tomwhite added support for TensorStore as an alternative on-disk Zarr store for writing intermediate results to....

> Cubed was designed from the beginning as a fully distributed, horizontally scalable, serverless execution framework. Sure you could do better by designing something for the vertical case for the...

The other reason to do this is that then it's obvious where to parameterize these tests with hypothesis strategies - i.e. for all netCDF files created by a strategy, the...