Description
What is your issue?
In #8733 (comment) and #9286 (comment) it's become fairly clear that the ChunkManager abstraction isn't quite right - it's "too greedy" as @dcherian said. #9286 will fix this by removing .chunks and .rechunk from the ChunkManager's responsibilities, but the result will be a "ChunkManager" that doesn't explicitly handle chunks!
I think a better name to describe the new interface is a "ComputeManager", as it still handles the creation of lazily-computed parallel arrays, distribution of computation over parts of those arrays, and triggering the materialization of the arrays.
JAX is also an interesting potential use case because there you don't have chunks, but you do still have to manage dividing computation up over multiple devices. See #9286 (comment)
Renaming ChunkManagerEntrypoint to ComputeManagerEntrypoint will be a breaking change but:
a) this is a very advanced feature,
b) the docs for it have a fat "experimental" warning on them,
c) I'm only aware of 2 libraries using this outside of xarray itself: cubed (tagging @tomwhite), and @hmaarrfk's chunked data structure. The dask ChunkManager ships with xarray, so there is no breaking change there. (Users may have to pip install again to re-register entrypoints if upgrading a development version of xarray inside existing environments though.)
I'm separating this out from #9286 because that PR shouldn't be a breaking change, and the follow-up that closes this issue will be the minimal possible breaking change (i.e. just renaming ChunkManagerEntrypoint -> ComputeManagerEntrypoint).