Keith Kraus
Keith Kraus
I'm happy to join this team as well (not sure if I'm steering or not currently...)
> Interchange between dask and ray could most likely be addressed by requiring ip-addresses and disallowing dask's more flexible identifiers. However, interchange with MPI is more difficult since MPI requires...
> Yes, one goal is definitely to minimize such overlap as much as possible. `__partitioned__` tries to only define the "address" or "location" in which the local object is accessible....
> Yes, that's why the spec defines that for MPI(-likes) the location must be the rank. That's an inherent characteristic of MPI and that's why that information must be in...
> Agree. The generalized concept could be the process id. I had been thinking about replacing `` in `'location' ` with a tuple of `(, )`. An additional, optional elements...
> As for providing a process ID, why is that not a granular enough. Can you elaborate? It requires domain knowledge about the producing distributed framework from the consumer. There's...
> * It feels a little unfortunate to go "halfway to Arrow" with adding the string `offsets` buffer, requiring serialization always for variable-size binary data. I haven't thought through what...
What should the behavior be if the library has its data on say a GPU instead of CPU? Throw? Copy to CPU? If we support this CPU only protocol why...
I am still a strong -1 on this proposal. The array API didn't implement `__array__` or `__array_interface__` despite being more widely adopted and used than the equivalent Arrow protocols being...