issue_comments

It would be wonderful if we could translate this complex xarray issue into a minimally simple zarr issue. Then the zarr devs can decide whether this use case is compatible with the zarr spec or not.

@dmedv and @rabernat - after thinking about this a bit more and reviewing the links in the last post, I'm pretty sure we're bumping into a bug in zarray's directory store pickle support. It would be nice to confirm this with some zarr-only tests but I don't see why the store needs to reference the zgroup files when the object is unpickled.

Also, @dmedv, can you add the output of xr.show_versions() to your original post?

The scenario you are describing--trying to open a file that is not accessible at all from the client--is certainly not something we ever considered when designing this. It is a miracle to me that it does work with netCDF.

True. I think its fair to say that the behavior you are enjoying (accessing data that the client cannot see) is the exception, not the rule. I expect there are many places in our backends that will not support this functionality at present.

The motivation for implementing the parallel feature was simply to shard the fileIO time when opening large collections (>10k) of netcdf files.

Ironically, this dask issue also popped up and has some significant overlap here: https://github.com/dask/dask/issues/5769

In both of these cases, the desire is for the worker to open the file (or zarr dataset), construct the underlying dask arrays, and return the meta object. This requires the object to be fully pickle-able and for any references to be maintained. It is possible, as indicated by your traceback, that the zarr backend is trying to reference the zgroup file and its not there. The logical place to start would be to look into why we can't pickle xarray datasets that come from zarr stores.

Thanks for these detailed reports!

The scenario you are describing--trying to open a file that is not accessible at all from the client--is certainly not something we ever considered when designing this. It is a miracle to me that it does work with netCDF.

I think you are on track with the serialization diagnostics. I believe that @jhamman has the best understanding of this topic. He implemented the parallel mode in open_mfdataset. Perhaps he can give some suggestions.

In the meantime, it seems worth asking the obvious question...how hard would it be to mount the NFS volume on the client? That would avoid having to go down this route.

Hi @dmedv -- thanks a lot for raising this issue here!

One clarification question: is there just a single zarr store you are trying to read? Or are you trying to combine multiple stores, like open_mfdataset does with multiple netcdf files?

Some of the data is only available on the workers, not on the client.

Can you provide more detail about how the zarr data is distributed across the different workers and client.

You can use the pseudocode here: https://xarray.pydata.org/en/stable/io.html#reading-multi-file-datasets and change open_dataset to open_zarr and then things should work (if I understand things correctly)