[QUESTION] Sharing a Python cudf, or better, dask-cudf?

Hi @trxcllnt & team!

Curious if any pointers on this. Ultimately, we're trying to setup a basic ~LRU over Python-managed dask-cudf objects (raw ddf or published) and have node do reads+writes with it. That's probably too big a leap for now, but I'm wondering if we can do something in the near term by scoping down to just CUDA IPC + cudf, or something similar?

Ultimate ideal case: multi-gpu dask

Ideally for the read direction, for example:

Dask-based Python publish:

ddf = dask.dataframe....
client.persist(ddf)
client.publish_dataset(xyz=ddf)

Dask-based JS load:

var ddf = nodeDaskCudf.getDataset('xyz'); // does not exist
var gdf = ddf.compute(); // does not exist
var arr = ddf.toArrow();

Near-term workaround via cudf + CUDA IPC

So I'm wondering: If we scoped down to python cudf <> node cudf, would this become doable, say via CUDA IPC?

Ex: Read direction

Python cudf pointer publish:

ddf = dask.dataframe....
gdf = client.compute(ddf)
handle : str = gdf.get_ipc_handle() # RMM / UCX basically do this already afaict
request.post(f'/nodejs/?handle={handle})

JS cudf readback:

function cb(handle) {
   var gdf = cudf.fromIPCHandle(handle); // ???
   var arr = gdf.toArrow();
}

This is a design question rather than a scoped implementation task; no files or tests are named. Start by checking the RMM/UCX CUDA IPC paths and the Python cudf-to-Node.js boundary described here, then define a minimal cudf IPC read path and its validation before tackling dask-cudf or published datasets.