encode_cf_datetime() casts dask arrays to NumPy arrays

[andersy005]

Currently, when xarray.coding.times.encode_cf_datetime() is called, it always casts the input to a NumPy array. This is not what I would expect when the input is a dask array. I am wondering if we could make this operation lazy when the input is a dask array?

xarray/xarray/coding/times.py

Lines 352 to 354 in 01462d6

	"""
	dates = np.asarray(dates)

In [46]: import numpy as np                                                                                        

In [47]: import xarray as xr                                                                                       

In [48]: import pandas as pd                                                                                       

In [49]: times = pd.date_range("2000-01-01", "2001-01-01", periods=11)                                             

In [50]: time_bounds = np.vstack((times[:-1], times[1:])).T                                                        

In [51]: arr = xr.DataArray(time_bounds).chunk()                                                                   

In [52]: arr                                                                                                       
Out[52]: 
<xarray.DataArray (dim_0: 10, dim_1: 2)>
dask.array<xarray-<this-array>, shape=(10, 2), dtype=datetime64[ns], chunksize=(10, 2), chunktype=numpy.ndarray>
Dimensions without coordinates: dim_0, dim_1

In [53]: xr.coding.times.encode_cf_datetime(arr)                                                                   
Out[53]: 
(array([[     0,  52704],
        [ 52704, 105408],
        [105408, 158112],
        [158112, 210816],
        [210816, 263520],
        [263520, 316224],
        [316224, 368928],
        [368928, 421632],
        [421632, 474336],
        [474336, 527040]]),
 'minutes since 2000-01-01 00:00:00',
 'proleptic_gregorian')

Cc @jhamman

[jhamman]

Thanks @andersy005 for the clear example. Looking at encode_cf_datetime(), it seems like we didn't intend to make this function dask friendly. Perhaps @spencerkclark or @shoyer have thoughts on the prospects of making this work. My guess is that this will be pretty tricky, mostly because in some cases, we infer the units/dtype/etc on the fly.

[shoyer]

Right, it casts to NumPy because it needs to look at the data to figure out units and calendar.

That said, if you pre-supply units and calendar there's no reason why it couldn't work with dask arrays. That would probably be a worthwhile refactor.

[dcherian]

Closed by #8575