save "encoding" when using open_mfdataset

[sbiner]

I like the automatic decoding of the time variable when reading netcdf files but I often need to keep the calendar attribute of the time variable.

Could it be possible to keep those attributes in the DataSet/DataArray return by open_dataset?

[rabernat]

Do you know you can access them in the .encoding namespace? Is that not sufficient for your needs?

[sbiner]

It would be ok but it is (or looks) empty when I use open_dataset()

[spencerkclark]

@sbiner are you looking at the encoding attribute of the full Dataset or the time variable? The time variable should retain the calendar encoding (the Dataset will not). E.g.:

In [1]: import cftime

In [2]: import numpy as np

In [3]: import xarray as xr

In [4]: units = 'days since 2000-02-25'

In [5]: times = cftime.num2date(np.arange(7), units=units, calendar='365_day')

In [6]: da = xr.DataArray(np.arange(7), coords=[times], dims=['time'], name='a')

In [7]: da.to_netcdf('data-noleap.nc')

In [8]: ds = xr.open_dataset('data-noleap.nc')

In [9]: ds.encoding['calendar']
---------------------------------------------------------------------------
KeyError                                  Traceback (most recent call last)
<ipython-input-38-677c245c7bb8> in <module>()
----> 1 default.encoding['calendar']

KeyError: 'calendar'

In [10]: ds.time.encoding['calendar']
Out[10]: u'noleap'

[sbiner]

@spencerkclark Yes I was looking at time.encoding. Following you example I did some tests and the problem is related to the fact that I am opening multiple netCDF files with open_mfdataset. Doing so time.encoding is empty while it is as expected when opening any of the files with open_dataset instead.

[TomNicholas]

If open_mfdataset() is actually dropping the encoding, then this is an issue related to #1614. That's because in open_mfdataset() while the attrs are explicitly set to those of the first supplied dataset, I don't see any similar explicit treatment of the encoding. I think that means the encoding is being set by what happens inside the core of auto_combine(), and is presumably being lost upon some of the concat or merge operations which happen inside auto_combine().

So I think to fix this then either open_mfdataset() should contain explicit treatment of the encoding, or the rules for propagating the encoding through the auto_combine() should be solidified.

[TomNicholas]

@sbiner I know it's been a while, but I expect that #3498 and #3877 probably resolve your issue?

[sbiner]

@TomNicholas I forgot about this sorry. I just made a quick check with the latest xarray master and I still have the problem ... see code.

Related question but maybe out of line, is there any way to know that the snw.time type is cftime.DatetimeNoLeap (as it is visible in the overview of snw.time)?

snw = xr.open_mfdataset(l_f, combine='nested', concat_dim='time')['snw']
ipdb> xr.__version__                                                                                                              
'0.15.2.dev29+g6048356'
ipdb> snw.time                                                                                                                    
<xarray.DataArray 'time' (time: 277393)>
array([cftime.DatetimeNoLeap(2006-01-01 00:00:00),
       cftime.DatetimeNoLeap(2006-01-01 03:00:00),
       cftime.DatetimeNoLeap(2006-01-01 06:00:00), ...,
       cftime.DatetimeNoLeap(2100-12-30 18:00:00),
       cftime.DatetimeNoLeap(2100-12-30 21:00:00),
       cftime.DatetimeNoLeap(2100-12-31 00:00:00)], dtype=object)
Coordinates:
  * time     (time) object 2006-01-01 00:00:00 ... 2100-12-31 00:00:00
Attributes:
    long_name:           time
    standard_name:       time
    axis:                T
    coordinate_defines:  point
ipdb> snw.time.encoding                                                                                                           
{}

[TomNicholas]

@TomNicholas I forgot about this sorry.

No worries!

I just made a quick check with the latest xarray master and I still have the problem ... see code.

#3498 added a new keyword argument to open_mfdataset, to choose which file to load to attributes from, can you try using that?

time.encoding is empty while it is as expected when opening any of the files with open_dataset instead

If this is the case, then to solve your original problem, you could also try using the preprocess argument to open_mfdataset to store the encoding somewhere where it won't be lost? i.e.

def store_encoding(ds):
    encoding = ds['time'].encoding
    ds.time.attrs['calendar_encoding'] = encoding
    return ds

snw = xr.open_mfdataset(l_f, combine='nested', concat_dim='time', 
                        master_file=lf[0], preprocess=store_encoding)['snw']

Related question but maybe out of line, is there any way to know that the snw.time type is cftime.DatetimeNoLeap (as it is visible in the overview of snw.time)?

I'm not familiar with these classes, but presumably you mean more than just checking with isinstance()? e.g.

from cftime import DatetimeNoLeap
print(isinstance(snw.time.values, cftime.DatetimeNoLeap))

[sbiner]

#3498 added a new keyword argument to open_mfdataset, to choose which file to load to attributes from, can you try using that?

#3498 says something about a master_file keyword but xr.open_mfdataset does not accept it and I do not see anything else similar in the documentation except attrs_file but it is the first file by default and it did not return the calendar even when I specified attrs_file=l_f[0].

If this is the case, then to solve your original problem, you could also try using the preprocess argument to open_mfdataset to store the encoding somewhere where it won't be lost? i.e.

def store_encoding(ds):
    encoding = ds['time'].encoding
    ds.time.attrs['calendar_encoding'] = encoding
    return ds

snw = xr.open_mfdataset(l_f, combine='nested', concat_dim='time', 
                        master_file=lf[0], preprocess=store_encoding)['snw']

I tried and it did not work ...

ipdb> ds = xr.open_mfdataset(l_f, combine='nested', concat_dim='time', preprocess=store_encoding)                                 
ipdb> ds.time                                                                                                                     
<xarray.DataArray 'time' (time: 2920)>
array([cftime.DatetimeNoLeap(2006-01-01 00:00:00),
       cftime.DatetimeNoLeap(2006-01-01 03:00:00),
       cftime.DatetimeNoLeap(2006-01-01 06:00:00), ...,
       cftime.DatetimeNoLeap(2006-12-31 15:00:00),
       cftime.DatetimeNoLeap(2006-12-31 18:00:00),
       cftime.DatetimeNoLeap(2006-12-31 21:00:00)], dtype=object)
Coordinates:
  * time     (time) object 2006-01-01 00:00:00 ... 2006-12-31 21:00:00
Attributes:
    long_name:           time
    standard_name:       time
    axis:                T
    coordinate_defines:  point
ipdb> ds.time.attrs                                                                                                               
{'long_name': 'time', 'standard_name': 'time', 'axis': 'T', 'coordinate_defines': 'point'}

Related question but maybe out of line, is there any way to know that the snw.time type is cftime.DatetimeNoLeap (as it is visible in the overview of snw.time)?

I'm not familiar with these classes, but presumably you mean more than just checking with isinstance()? e.g.

Yes, I was more thinking of something like type(ds.time) which would return cftime.DatetimeNoLeap

from cftime import DatetimeNoLeap
print(isinstance(snw.time.values, cftime.DatetimeNoLeap))

[keewis]

unfortunately, numpy does not allow us to put cftime object into dtypes (yet!), so ds.time.values is a numpy.ndarray with dtype object, containing cftime objects. To make the code work, use it with ds.time.values[0]. Of course, that won't help if the array contains objects of more than one type.

>>> import cftime
>>> isinstance(ds.time.values[0], cftime.DatetimeNoLeap)
True
>>> type(ds.time.values[0])
<class 'cftime._cftime.DatetimeNoLeap'>

In #3498, the original proposal was to name the new kwarg master_file, but later it was renamed to attrs_file. If l_f is a list of file paths, you used it correctly.

Before trying to help with debugging your issue: could you post the output of xr.show_versions()? That would help narrowing down on whether it's a dependency issue or a bug in xarray.

Also, could you try to demonstrate your issue using a synthetic example? I've been trying to reproduce it with:

In [14]: units = 'days since 2000-02-25' 
    ...: times = cftime.num2date(np.arange(7), units=units, calendar='365_day') 
    ...: for x in range(5): 
    ...:     ds = xr.DataArray( 
    ...:         np.arange(x, 7 + x).reshape(7, 1), 
    ...:         coords={"time": times, "x": [x]}, 
    ...:         dims=['time', "x"], 
    ...:         name='a', 
    ...:     ).to_dataset() 
    ...:     ds.to_netcdf(f'data-noleap{x}.nc') 
    ...: paths = sorted(glob.glob("data-noleap*.nc")) 
    ...: with xr.open_mfdataset(paths, combine="by_coords") as ds: 
    ...:     print(ds.time.encoding) 
    ...:
{'zlib': False, 'shuffle': False, 'complevel': 0, 'fletcher32': False, 'contiguous': True, 'chunksizes': None, 'source': '.../data-noleap0.nc', 'original_shape': (7,), 'dtype': dtype('int64'), 'units': 'days since 2000-02-25 00:00:00.000000', 'calendar': 'noleap'}

[dcherian]

This example works without attrs_file specified.

[keewis]

I removed it since it doesn't change anything.