[BUG] DatetimeArray is not writable in cudf.pandas

[galipremsagar]

Describe the bug
When a DatetimeArray is created in cudf.pandas, it appears to be that the internal ndarray is not writable.

Steps/Code to reproduce bug

import pandas as pd
i = pd.date_range("2016-01-01 01:01:00", periods=5)
print(i)
print(i._data)
print(i._data._ndarray.flags)

output:

(cudfdev) pgali@dgx19:/datasets/pgali/cudf$ python y1.py 
DatetimeIndex(['2016-01-01 01:01:00', '2016-01-02 01:01:00',
               '2016-01-03 01:01:00', '2016-01-04 01:01:00',
               '2016-01-05 01:01:00'],
              dtype='datetime64[ns]', freq='D')
<DatetimeArray>
['2016-01-01 01:01:00', '2016-01-02 01:01:00', '2016-01-03 01:01:00',
 '2016-01-04 01:01:00', '2016-01-05 01:01:00']
Length: 5, dtype: datetime64[ns]
  C_CONTIGUOUS : True
  F_CONTIGUOUS : True
  OWNDATA : False
  WRITEABLE : True
  ALIGNED : True
  WRITEBACKIFCOPY : False

(cudfdev) pgali@dgx19:/datasets/pgali/cudf$ python -m cudf.pandas y1.py 
DatetimeIndex(['2016-01-01 01:01:00', '2016-01-02 01:01:00',
               '2016-01-03 01:01:00', '2016-01-04 01:01:00',
               '2016-01-05 01:01:00'],
              dtype='datetime64[ns]', freq='D')
<DatetimeArray>
['2016-01-01 01:01:00', '2016-01-02 01:01:00', '2016-01-03 01:01:00',
 '2016-01-04 01:01:00', '2016-01-05 01:01:00']
Length: 5, dtype: datetime64[ns]
  C_CONTIGUOUS : True
  F_CONTIGUOUS : True
  OWNDATA : False
  WRITEABLE : False # Expected True
  ALIGNED : True
  WRITEBACKIFCOPY : False

[mroeschke]

It appears pyarrow is making this not writeable e.g.

In [4]: import pyarrow as pa

In [5]: import datetime

In [7]: pa.array([datetime.datetime(2020, 1, 1)]).to_pandas().array._ndarray.flags
Out[7]: 
  C_CONTIGUOUS : True
  F_CONTIGUOUS : True
  OWNDATA : False
  WRITEABLE : False
  ALIGNED : True
  WRITEBACKIFCOPY : False

[mroeschke]

Opened apache/arrow#45341 to track this

[galipremsagar]

Though the problem is in pyarrow, the issue only shows up with usage of cudf.pandas, can we not avert it?

[mroeschke]

to_pandas is being called during i._data which triggers the behavior shown in #17743 (comment)

But I think we can call pd.arrays.ArrowExtensionArray(pa_array).to_numpy(copy=) in to_pandas to sidestep arrow always avoiding a copy and making the result writable

[mroeschke]

OK I realized we can use pyarrow.Array.to_numpy(zero_copy_only=False, writable=True) in mode.pandas_compatible to make the numpy array writable, but then we'll always be making a CPU copy of the data during to_pandas. Is that acceptable?

[vyasr]

So our options are

1. No copy but the result is not writeable
2. With a copy and the result is writeable

If we take a step back for a second and forget about cudf.pandas, how do we expect this to behave if we did this explicitly in cudf? If we created a cudf.DatetimeIndex and called to_pandas on it, should the resulting object have a writeable _data array? Should that behavior perhaps be dependent on mode.pandas_compatible? It seems like this behavior is something we'd want to be decided in the to_pandas implementation regardless of whether we're using cudf.pandas or not.