The method to_netcdf does not preserve chunks

[yt87]

What happened?

Methods to_zarr and to_netcdf behave inconsistently for chunked dataset. The latter does not preserve existing chunk information, the chunks must be specified within the encoding dictionary.

What did you expect to happen?

I expected the behaviour to be consistent for for all to_XXX() methods.

Minimal Complete Verifiable Example

import xarray as xr
import dask.array as da

rng = da.random.RandomState()
shape = (20, 20)
chunks = [10, 10]
dims = ["x", "y"]
z = rng.standard_normal(shape, chunks=chunks)
ds = xr.DataArray(z, dims=dims, name="z").to_dataset()
ds.chunks
# This one is rechunked
ds.to_netcdf("/tmp/test1.nc", encoding={"z": {"chunksizes": (5, 5)}})
# This one is not rechunked, also original chunks are lost
ds.chunk({"x": 5, "y": 5}).to_netcdf("/tmp/test2.nc")
# This one is rechunked
ds.chunk({"x": 5, "y": 5}).to_zarr("/tmp/test2", mode="w")

Frozen({'x': (10, 10), 'y': (10, 10)})
<xarray.backends.zarr.ZarrStore at 0x7f3669f1af80>

xr.open_mfdataset("/tmp/test1.nc").chunks
xr.open_mfdataset("/tmp/test2.nc").chunks
xr.open_mfdataset("/tmp/test2", engine="zarr").chunks

Frozen({'x': (5, 5, 5, 5), 'y': (5, 5, 5, 5)})
Frozen({'x': (20,), 'y': (20,)})
Frozen({'x': (5, 5, 5, 5), 'y': (5, 5, 5, 5)})

MVCE confirmation

• Minimal example — the example is as focused as reasonably possible to demonstrate the underlying issue in xarray.
• Complete example — the example is self-contained, including all data and the text of any traceback.
• Verifiable example — the example copy & pastes into an IPython prompt or Binder notebook, returning the result.
• New issue — a search of GitHub Issues suggests this is not a duplicate.
• Recent environment — the issue occurs with the latest version of xarray and its dependencies.

Relevant log output

No response

Anything else we need to know?

I did get the same results for h5netcdf and scipy backends, so I am not sure whether this is a bug or not.
The above code is a modified version of #2198.
A suggestion: the documentation provides only examples of encoding styles. It would be helpful to provide links to a full specification.

Environment

Details

INSTALLED VERSIONS ------------------ commit: None python: 3.11.6 | packaged by conda-forge | (main, Oct 3 2023, 10:40:35) [GCC 12.3.0] python-bits: 64 OS: Linux OS-release: 6.5.5-1-MANJARO machine: x86_64 processor: byteorder: little LC_ALL: None LANG: en_US.UTF-8 LOCALE: ('en_US', 'UTF-8') libhdf5: 1.14.2 libnetcdf: 4.9.2

xarray: 2023.10.1
pandas: 2.1.1
numpy: 1.24.4
scipy: 1.11.3
netCDF4: 1.6.4
pydap: None
h5netcdf: 1.2.0
h5py: 3.10.0
Nio: None
zarr: 2.16.1
cftime: 1.6.2
nc_time_axis: None
PseudoNetCDF: None
iris: None
bottleneck: 1.3.7
dask: 2023.10.0
distributed: 2023.10.0
matplotlib: 3.8.0
cartopy: 0.22.0
seaborn: None
numbagg: 0.5.1
fsspec: 2023.10.0
cupy: None
pint: None
sparse: 0.14.0
flox: 0.8.1
numpy_groupies: 0.10.2
setuptools: 68.2.2
pip: 23.3.1
conda: None
pytest: None
mypy: None
IPython: 8.16.1
sphinx: None

[NikosAlexandris]

I am experimenting the last hours with this too. Here a confirmation :

In [43]: xr.open_dataset('SDUds202001010000004231000101MA.nc')['SDU'].encoding
Out[43]:
{'zlib': True,
 'szip': False,
 ..
 'contiguous': False,
 'chunksizes': (1, 1, 2600),
 ..
 'original_shape': (1, 2600, 2600),
 ..}

Rechunking and writing out

sdu20200101_rechunked = xr.open_dataset(
    'SDUds202001010000004231000101MA.nc',
    decode_cf = True,  # useful at this step ?
    chunks = {'time': 1, 'lat': 10, 'lon': 10}
)
sdu20200101_rechunked.to_netcdf('sdu20200101_rechunked.nc')
sdu20200101_rechunked.to_zarr('sdu20200101_rechunked.zarr')

Diagnosing

In [48]: xr.open_dataset('sdu20200101_rechunked.nc').chunksizes
Out[48]: Frozen({})

In [49]: xr.open_zarr('sdu20200101_rechunked.zarr')['SDU'].chunksizes
Out[49]: Frozen({'time': (1,), 'lat': (10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10), 'lon': (10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10)})

[yt87]

Maybe this should be rephrased as a usage question: why netcdf backends do not pull the chunk information from the variable? At least for my example I can make it work by adding

if 'chunksizes' not in encoding:
    encoding['chunksizes'] = tuple(max(c) for c in variable.chunks)

after else: statement on line 509 in https://github.com/pydata/xarray/blob/main/xarray/backends/netCDF4_.py#L509
I suspect there must be a reason why it can't be done.

[NikosAlexandris]

By the way, I still cannot find a practical way using Python to rechunk data in a NetCDF file. I can do so using the standard nccopy utility.

[doblerone]

By the way, I still cannot find a practical way using Python to rechunk data in a NetCDF file. I can do so using the standard nccopy utility.

Hi. I did recently rechunk a bunch of 25GB NetCDF files from chunks (time,lat,lon): 1,500,1000 to
chunks (time,lat,lon): 8760,1,1
using

import xarray as xr

ds = xr.open_dataset('field_access_opt.nc') # chunks (time,lat,lon): 1,500,1000

#re-chunk variable 'GHI'
ds.to_netcdf("point_access_opt.nc",
encoding={'lat': {'zlib': False, '_FillValue': None},
'lon': {'zlib': False, '_FillValue': None},
'time': {'zlib': False, '_FillValue': None, 'dtype': 'double'},
'GHI': {'chunksizes': [len(ds['GHI'].time),1,1], 'zlib': True,
'complevel': 1}})

Original files:
https://thredds.met.no/thredds/catalog/sunpoint/ML-Optimized-Maps/hourly/field_access/catalog.html

Rechunked files:
https://thredds.met.no/thredds/catalog/sunpoint/ML-Optimized-Maps/hourly/point_access/catalog.html

(MEM usage was approx. 2x file size)

Feel free to remove this comment if it is not helpful.