What is your issue?

Hi folks, I'm not sure what I'm doing wrong here. Context: I have a dataset, adding a coordinate variable with some specified chunking, and then that chunking is reset when writing to_zarr() and opening from disk. Even if I call unify_chunks() before writing or explicitly set the preferred_chunks encoding, the chunks get reset upon write. I have managed to force the chunks by explicitly setting ds.end_date.encoding['chunks']= (2,) before writing (note that before this, no encoding was set, so this isn't the old issue of previous encoding overwriting the dask chunks). It took a while to find this, and I don't understand why the existing behavior disregards the existing chunks on single-dimension variables or coordinates. The issue is that with this default behavior, the dataset written to disk has inconsistent dimensions on read, throws an error, and requires a call to unify_chunks() to be usable.

Reproducer below:

versions: xarray 2023.6.0 dask 2023.6.0 zarr 2.14.2

``` import pandas as pd import numpy as np import xarray as xr from datetime import datetime import dask.array as da

from pandas.tseries.offsets import MonthEnd

create toy dataset

dates = [datetime(2021, 1, 1), datetime(2021, 2, 1), datetime(2021, 3, 1), datetime(2021, 4, 1)] ds = xr.Dataset( data_vars=dict( tas=(["time", "lat", "lon",], np.array([300]) * np.ones((len(dates), 4, 4)), ), ), coords=dict( time=dates, lat=np.array([10, 11, 12, 13]), lon=np.array([20, 21, 22, 23]), ), )

chunk dataset

ds = ds.chunk(chunks={"lat": 2, "lon": 2, 'time': 2})

add end date coordinate

time_df = pd.DataFrame({"time": list(ds['time'].values)}) time_df['end_date'] = time_df['time'] + MonthEnd(0) ds = ds.assign_coords( end_date=("time", da.from_array(time_df['end_date'].values)))

print("---data before unify chunks--- \n", ds.end_date.data) ds = ds.unify_chunks() # surely you'll respect my chunking print("---data before writing to s3--- \n", ds.end_date.data) ds.time.encoding['preferred_chunks']= {'time': 2} # please use this chunking? ds.end_date.encoding['preferred_chunks']= {'time': 2} # pretty please use this chunking?

write data to s3

data_path = "/tmp/mydata.zarr" ds.to_zarr(data_path, mode='w')

read data back from s3

ds_check = xr.open_zarr(data_path)

print("---data after writing to disk and reading back in--- \n", ds_check.end_date.data) ```

output from code above: ---data before unify chunks--- dask.array<array, shape=(4,), dtype=datetime64[ns], chunksize=(4,), chunktype=numpy.ndarray> ---data before writing to s3--- dask.array<rechunk-merge, shape=(4,), dtype=datetime64[ns], chunksize=(2,), chunktype=numpy.ndarray> ---data after writing to disk and reading back in--- dask.array<open_dataset-end_date, shape=(4,), dtype=datetime64[ns], chunksize=(4,), chunktype=numpy.ndarray>