Writing to zarr fails with message "specified zarr chunks would overlap multiple dask chunks"

[TonioF]

For some data it is necessary to rechunk it in order to deal with it efficiently. This rechunking process might result in that the final chunk of a dimension has a smaller size than the previous ones. E.g., if we have a dimension of 500, originally split into ten chunks of size 50, we can rechunk it to sizes of 200, 200, and 100. This is perfectly valid and supported by xarray.
However, when the affected dimension/variable is latitude and the latitude is ascending, the current implementation of the normalization in xcube will revert the order of chunks and cause that the smaller chunk will be at the start - which is not supported by xarray and might result in an error as the one in the description, for example when writing to zarr.

For the time being, I will work around this issue by ensuring that all chunks are of the same size after rechunking, which will make this issue a bit harder to reproduce. I open this issue to document that the current solution of dealing with ascending latitudes is not optimal. The best solution will probably be to get rid of the part where latitudes are reverted during the normalization and support ascending latitudes in the data.

See also #251 and #327.

[AliceBalfanz]

Hi Tonio, I am facing a similar issue with the same error message.

When executing the example jupyter NB examples/store/open_data_directory.ipynb
line
store.write_data(new_dataset, 'cube-1-250-250-subset.zarr', writer_id='dataset:zarr:posix')
results in

NotImplementedError: Specified zarr chunks encoding['chunks']=(1, 250, 250) for variable named 'quality_flags' would overlap multiple dask chunks ((1, 1, 1, 1, 1), (90, 90, 90, 90), (90, 90, 90, 90, 90, 90, 90, 90, 60)). This is not implemented in xarray yet. Consider either rechunking using `chunk()` or instead deleting or modifying `encoding['chunks']`.

xarray does not adjust the encoding of varialbes when rechunking a dataset.

new_dataset.c2rcc_flags.encoding

{'chunks': (1, 250, 250),
 'compressor': Blosc(cname='lz4', clevel=5, shuffle=SHUFFLE, blocksize=0),
 'filters': None,
 '_FillValue': nan,
 'dtype': dtype('float64')}

This problem is already known:

pydata/xarray#2300

Maybe we could implement in xcube, that if a dataset is rechunked, then the encoding is adjusted as well?

[TonioF]

Maybe we could implement in xcube, that if a dataset is rechunked, then the encoding is adjusted as well?

We already did: https://github.com/dcs4cop/xcube/blob/bc4cdb4aa5e88557d920d71b7dff4100015c3512/xcube/core/chunk.py#L10 (Make sure you set the format when you use this, otherwise the encodings won't be updated).

However, I still couldn't run your notebook cell successfully, but that seems to be due to another problem.

[AliceBalfanz]

Thanks for your suggestion, I changed the notebook to use xcube chunk_dataset instead of xarrays way to chunk a dataset. It now works like a dream :)

[rabernat]

This can and should be fixed in xarray. But it's also very easy to just delete the encoding.

del new_dataset.c2rcc_flags.encoding['chunks']

as a simple workaround.

[forman]

@rabernat thanks! The xcube function chunk_dataset() is basically xarray.chunk() plus correction of "chunks" encoding in each variable.

@AliceBalfanz

Thanks for your suggestion, I changed the notebook to use xcube chunk_dataset instead of xarrays way to chunk a dataset.

We should make sure that any xr.Dataset returned from higher-level xcube functions should always have a chunking compatible with Zarr, as this is our standard I/O format. Users should not be forced to rechunk just for the purpose of writing to Zarr. This is counter-intuitive. I suggest we provide a utiliy function that ensures "valid" chunking (including possible deletion of the "chunks" encoding property, as suggested by @rabernat), ensure_valid_chunks(). Then we'll apply that to all datasets before we return them from xcube functions.

I guess there are good reasons why encoding is not adjusted in xarray.chunk().

[rabernat]

The xcube function chunk_dataset() is basically xarray.chunk() plus correction of "chunks" encoding in each variable.

🙌 Could I convince you to submit this as a PR to xarray itself? 😁

[forman]

I'm convinced. Hope to find a little time next week. Maybe there is a related issue already?

[rabernat]

pydata/xarray#2300 is the main one.