issue_comments

IIUC the issue Ryan & Tom are talking about is tied to reading from files.

For example, we read from a zarr store using zarr, then wrap that zarr.Array (or h5Py Dataset) with a large number of ExplicitlyIndexed Classes that enable more complicated indexing, lazy decoding etc.

IIUC #4628 is about concatenating such arrays i.e. neither zarr.Array nor ExplicitlyIndexed support concatenation, so we end up calling np.array and forcing a disk read.

With dask or cubed we would have dask(ExplicitlyIndexed(zarr)) or cubed(ExplicitlyIndexed(zarr)) so as long as dask and cubed define concat and we dispatch to them, everything is 👍🏾

PS: This is what I was attempting to explain (not very clearly) in the distributed arrays meeting. We don't ever use dask.array.from_zarr (for e.g.). We use zarr to read, then wrap in ExplicitlyIndexed and then pass to dask.array.from_array.

Cubed implements concat, but perhaps xarray needs richer concat functionality than that?

Cubed should define a concatenate function, so that should be OK

@rabernat just pointed out to me that in order for this to work well we might also need lazy concatenation of arrays.

Xarray currently has it's own internal wrappers that allow lazy indexing, but they don't yet allow lazy concatenation. Instead dask is what does lazy concatenation under the hood right now.

This is a problem - it means that concatenating two cubed-backed DataArrays will trigger loading both into memory, whereas concatenating two dask-backed DataArrays will not. If #4628 was implemented then xarray would never load the underlying array into memory regardless of the backend.

Not really a generic and parallel execution back-end, but Open-EO looks like an interesting use case too (it is a framework for managing remote execution of processing tasks on multiple big Earth observation cloud back-ends via a common API). I've suggested the idea of reusing the Xarray API here: https://github.com/Open-EO/openeo-python-client/issues/334.

Might I propose Arkouda?

https://github.com/Bears-R-Us/arkouda https://chapel-lang.org/presentations/Arkouda_SIAM_PP-22.pdf

Have they improved recently to support more than 1D arrays?

Might I propose Arkouda?

https://github.com/Bears-R-Us/arkouda https://chapel-lang.org/presentations/Arkouda_SIAM_PP-22.pdf

Thanks for opening this @TomNicholas

The challenge will be defining a parallel computing API that works across all these projects, with their slightly different models.

Agreed. I feel like there's already an implicit set of "chunked array" methods that xarray expects from Dask that could be formalised a bit and exposed as an integration point.

at SciPy i learned of fugue which tries to provide a unified API for distributed DataFrames on top of Spark and Dask. it could be a great source of inspiration.

Sounds good to me. The challenge will be defining a parallel computing API that works across all these projects, with their slightly different models.

Another parallel framework would be Ramba

cc @DrTodd13

This sounds great! We should finish up https://github.com/pydata/xarray/pull/4972 to make it easier to test.