@kmuehlbauer - genius! Yes. That pull request should fix this issue exactly! And it explains why I see this issue and you don't - with undefined behavior anything can happen. Since we are on different OSes, our systems behave differently.

I just double checked with pandas and this fix will do the right thing: python import pandas as pd print(pd.to_timedelta([np.nan, 0],"ns") + np.datetime64('1970-01-01')) DatetimeIndex(['NaT', '1970-01-01'], dtype='datetime64[ns]', freq=None) I see that the pull request with the fix has been sitting since December of last year. Is there some way to somehow get someone to look at that pull request who can merge it?

Ah hah! Well, I don't know why this is working for you @kmuehlbauer, but I can see why it is not working for me. I've been debugging through the code and it looks like the problem is the _decode_datetime_with_pandas function. For me, it's converting a float NaN into an integer, which results in a zero value.

It all starts in the open_zarr function, which by default sets the use_cftime parameter to None by default: https://github.com/pydata/xarray/blob/25d9a28e12141b9b5e4a79454eb76ddd2ee2bc4d/xarray/backends/zarr.py#L701-L817

There's a bunch of stuff that gets called, but eventually we get to the function decode_cf_datetime, which ironically (given the name) also takes this use_cftime parameter, which is still None. Because use_cftime is None, the function calls _decode_datetime_with_pandas:

https://github.com/pydata/xarray/blob/25d9a28e12141b9b5e4a79454eb76ddd2ee2bc4d/xarray/coding/times.py#L265-L289

and then, in _decode_datetime_with_pandas, the code casts a float NaN value to zero:

https://github.com/pydata/xarray/blob/979b99831f5d34d33120312a15dad3e6a0830f32/xarray/coding/times.py#L216-L262

In line 254, flat_num_dates is array([ nan, 1.6726176e+18]). After line 254, flat_nuM-dates_ns_int is array([ 0, 1672617600000000000]).

@kmuehlbauer - I ran https://github.com/pydata/xarray/issues/7790#issuecomment-1529894939 and I get an incorrect fill value:

```

Created with fill value 1900-01-01 <xarray.DataArray 'time' (time: 2)> array([ 'NaT', '2023-01-02T00:00:00.000000000'], dtype='datetime64[ns]') Coordinates: * time (time) datetime64[ns] NaT 2023-01-02

Read back out of the zarr store with xarray <xarray.DataArray 'time' (time: 2)> array(['1970-01-01T00:00:00.000000000', '2023-01-02T00:00:00.000000000'], dtype='datetime64[ns]') Coordinates: * time (time) datetime64[ns] 1970-01-01 2023-01-02 {} {'chunks': (2,), 'preferred_chunks': {'time': 2}, 'compressor': Blosc(cname='lz4', clevel=5, shuffle=SHUFFLE, blocksize=0), 'filters': None, '_FillValue': -2208988800000000000, 'units': 'nanoseconds since 1970-01-01', 'calendar': 'proleptic_gregorian', 'dtype': dtype('int64')}

Read back out of the zarr store with zarr <zarr.core.Array '/time' (2,) int64 read-only> <zarr.attrs.Attributes object at 0x132802a50> [-2208988800000000000 1672617600000000000]

and here is my show_versions, since it may have changed because I've added some new libraries. It looks like my ipython version is slightly different, but I can't see how that would affect things. INSTALLED VERSIONS

commit: None python: 3.11.3 | packaged by conda-forge | (main, Apr 6 2023, 08:58:31) [Clang 14.0.6 ] python-bits: 64 OS: Darwin OS-release: 22.4.0 machine: arm64 processor: arm byteorder: little LC_ALL: None LANG: en_US.UTF-8 LOCALE: ('en_US', 'UTF-8') libhdf5: None libnetcdf: None

xarray: 2023.4.2 pandas: 2.0.1 numpy: 1.24.3 scipy: None netCDF4: None pydap: None h5netcdf: None h5py: None Nio: None zarr: 2.14.2 cftime: None nc_time_axis: None PseudoNetCDF: None iris: None bottleneck: None dask: None distributed: None matplotlib: None cartopy: None seaborn: None numbagg: None fsspec: None cupy: None pint: None sparse: None flox: None numpy_groupies: None setuptools: 67.7.2 pip: 23.1.2 conda: None pytest: None mypy: None IPython: 8.13.1 sphinx: None ```

Oops! Yes. You are right. I had some cross-wording on the variable names. So I started a new notebook. Unfortunately, I think you may have also gotten some wires crossed? You set the time fill value to 1900-01-01, but then use NaT in the actual array?

Here is a fresh notebook with a stand-alone cell with everything that I think you were doing, but I'm not 100%. The fill value is still wrong when it gets read out, but it is at least different? The fill value is now set to the units for some reason. This seems like progress?

```python import numpy as np import xarray as xr import zarr

Create a time array with one fill value, NaT

time = np.array([np.datetime64("NaT", "ns"), '2023-01-02 00:00:00.00000000'], dtype='M8[ns]')

Create xarray with this fill value

xr_time_array = xr.DataArray(data=time,dims=['time'],name='time') xr_ds = xr.Dataset(dict(time=xr_time_array)) print("****") print("xarray created with NaT fill value") print("----------------------") print(xr_ds["time"])

Save as zarr

location_with_units = "xarray_and_units.zarr" encoding = { "time":{"_FillValue":np.datetime64("NaT","ns"),"dtype":np.int64,"units":"nanoseconds since 1970-01-01"} } xr_ds.to_zarr(location_with_units,mode="w",encoding=encoding)

Read it back out again

xr_read = xr.open_zarr(location_with_units) print("****") print("xarray created read with NaT fill value") print("----------------------") print(xr_read["time"]) print(xr_read["time"].attrs) print(xr_read["time"].encoding)

```

```

xarray created with NaT fill value

<xarray.DataArray 'time' (time: 2)> array([ 'NaT', '2023-01-02T00:00:00.000000000'], dtype='datetime64[ns]') Coordinates: * time (time) datetime64[ns] NaT 2023-01-02

xarray created read with NaT fill value

<xarray.DataArray 'time' (time: 2)> array(['1970-01-01T00:00:00.000000000', '2023-01-02T00:00:00.000000000'], dtype='datetime64[ns]') Coordinates: * time (time) datetime64[ns] 1970-01-01 2023-01-02 {} {'chunks': (2,), 'preferred_chunks': {'time': 2}, 'compressor': Blosc(cname='lz4', clevel=5, shuffle=SHUFFLE, blocksize=0), 'filters': None, '_FillValue': -9223372036854775808, 'units': 'nanoseconds since 1970-01-01', 'calendar': 'proleptic_gregorian', 'dtype': dtype('int64')}

```

Where in the code is the time array being decoded? That seems to be where a lot of the issue is?

The zarr store does indeed use an integer in this case according to the .zmetadata file: { "metadata": { ".zattrs": {}, ".zgroup": { "zarr_format": 2 }, "time/.zarray": { "chunks": [ 2 ], "compressor": { "blocksize": 0, "clevel": 5, "cname": "lz4", "id": "blosc", "shuffle": 1 }, "dtype": "<i8", "fill_value": -25567, "filters": null, "order": "C", "shape": [ 2 ], "zarr_format": 2 }, "time/.zattrs": { "_ARRAY_DIMENSIONS": [ "time" ], "calendar": "proleptic_gregorian", "units": "days since 1900-01-01 00:00:00" } }, "zarr_consolidated_format": 1 } Once again the values in the zarr store are correct given the units, but xarray misreads the fill value for some reason: python z_new_fill = zarr.open('from_xarray_new_fill.zarr','r') z_new_fill["time"][:] array([ 0, 44926])

@kmuehlbauer - I think I'm not understanding what you are suggesting because the zarr store is still not being read correctly when I switch the fill value to a different date:

```python

Create a numpy array of type np.datetime64 with one fill value and one date

time_fill_value = np.datetime64("1900-01-01") time = np.array([time_fill_value,'2023-01-02'],dtype='M8[ns]')

Create a dataset with this one array

xr_time_array = xr.DataArray(data=time,dims=['time'],name='time') xr_ds = xr.Dataset(dict(time=xr_time_array))

print("******") print("Created with fill value 1900-01-01") print(xr_ds["time"])

Save the dataset to zarr

location_new_fill = "from_xarray_new_fill.zarr" encoding = { "time":{"_FillValue":time_fill_value,"dtype":np.int64} } xr_ds.to_zarr(location_new_fill,encoding=encoding)

xr_read = xr.open_zarr(location) print("******") print("Read back out of the zarr store with xarray") print(xr_read["time"]) print(xr_read["time"].encoding)

Created with fill value 1900-01-01 <xarray.DataArray 'time' (time: 2)> array(['1900-01-01T00:00:00.000000000', '2023-01-02T00:00:00.000000000'], dtype='datetime64[ns]') Coordinates: * time (time) datetime64[ns] 1900-01-01 2023-01-02

<xarray.DataArray 'time' (time: 2)> array(['2023-01-02T00:00:00.000000000', '2023-01-02T00:00:00.000000000'], dtype='datetime64[ns]') Coordinates: * time (time) datetime64[ns] 2023-01-02 2023-01-02 {'chunks': (2,), 'preferred_chunks': {'time': 2}, 'compressor': Blosc(cname='lz4', clevel=5, shuffle=SHUFFLE, blocksize=0), 'filters': None, '_FillValue': -9.223372036854776e+18, 'units': 'days since 2023-01-02 00:00:00', 'calendar': 'proleptic_gregorian', 'dtype': dtype('float64')} ```

Interestingly, xarray is also perfectly happy to read a numpy.datetime64 array out of a zarr store as long as the xarray metadata is present. xarray even helpfully creates an '_FillValue" attribute for the array so there is no confusion:

```

Create a zarr store directly with numpy.datetime64 type

location_zarr_direct = "from_zarr.zarr" root = zarr.open(location_zarr_direct,mode='w') z_time_array = root.create_dataset( "time",data=time,shape=time.shape,chunks=time.shape,dtype=time.dtype, fill_value=time_fill_value )

Add xarray metadata

z_time_array.attrs["_ARRAY_DIMENSIONS"] = ["time"] zarr.convenience.consolidate_metadata(location_zarr_direct)

Use xarray to read this data out

xr_read_from_zarr = xr.open_zarr(location_zarr_direct) print(xr_read_from_zarr["time"]) <xarray.DataArray 'time' (time: 2)> array([ 'NaT', '2023-01-02T00:00:00.000000000'], dtype='datetime64[ns]') Coordinates: * time (time) datetime64[ns] NaT 2023-01-02 Attributes: _FillValue: NaT ```

So I am extremely confused as to why xarray encodes time arrays so strangely when it creates the zarr store itself! (Hence https://github.com/pydata/xarray/discussions/7776)

Ah! Okay. I did not know about the .encoding option, which does indeed have the fill value. Thank you.

Interestingly, -9.223372036854776e+18 is just the float equivalent of numpy.datetime64('NaT'):

python float(np.datetime64('NaT').view('i8')) -9.223372036854776e+18

And I know this isn't an issue with zarr and NaT because I can create the zarr store directly with the zarr library and it's perfectly happy: ```python

Create a zarr store directly with numpy.datetime64 type

location_zarr_direct = "from_zarr.zarr" root = zarr.open(location_zarr_direct,mode='w') z_time_array = root.create_dataset( "time",data=time,shape=time.shape,chunks=time.shape,dtype=time.dtype, fill_value=time_fill_value ) zarr.convenience.consolidate_metadata(location_zarr_direct)

Read it back out again

read_zarr = zarr.open(location_zarr_direct,mode='r') print(read_zarr["time"][:]) [ 'NaT' '2023-01-02T00:00:00.000000000'] ```