How does xarray write a dask array to zarr using the distributed scheduler?

[ivirshup]

Hey,

This is much more of an implementation question than anything else.

Over on anndata we ran into the issue of serializing locks when trying to write a dask array to a hdf5 or zarr store:

• Dask Distributed Write Fix For Zarr scverse/anndata#1079

We came up with a solution based on what xarray does for HDF5, though in our case this only works for zarr I believe since we don't have xarrays CachingFileManager (#4242). However, xarray is not currently using this approach for zarr as far as I can tell.

Question: Could someone explain how xarray writes a dask array to a zarr store when using the distributed schedueler? I'm not seeing an obvious use of a lock or synchronizer in the zarr backend code.

[ivirshup]

Looking into it a bit, I believe xarray ends up passing dask.store(..., lock=None) for the zarr backend. Still working on figuring out where the chunks are aligned to make it safe to write without a lock though.

[TomNicholas]

The writing should all go through .store yes. For the other questions @jhamman might know.