numpy arrays read from attributes are not writable

I'm trying to wrap my head around the following issue. Arrays read from Datasets are created with numpy.memmap with "copyonwrite" flag set. So these numpy arrays can be changed in memory but not in the file. This is not true for arrays read from attributes. There we get back a numpy.ndarray with the writable-flag set to FALSE. An MCVE is attached below.

I was already looking where the flag might be set and why, but didn't come far. Any hints much appreciated.

MCVE:

import numpy as np
import h5py
import pyfive

with h5py.File("test.h5", mode="w") as f:
    f.attrs["att"] = np.arange(10)
    f["arr"] = np.arange(10)

with pyfive.File("test.h5") as f:
    # dataset
    arr = f["arr"][:]
    print(type(arr))
    print(arr.flags)
    print(arr.base)
    arr += 1
    print(arr)
    # attribute
    att = f.attrs["att"]
    print(type(att))
    print(att.flags)
    print(att.base)
    att += 1
    print(att)

<class 'numpy.memmap'>
  C_CONTIGUOUS : True
  F_CONTIGUOUS : True
  OWNDATA : False
  WRITEABLE : True
  ALIGNED : True
  WRITEBACKIFCOPY : False
  UPDATEIFCOPY : False

[0 1 2 3 4 5 6 7 8 9]
[ 1  2  3  4  5  6  7  8  9 10]
<class 'numpy.ndarray'>
  C_CONTIGUOUS : True
  F_CONTIGUOUS : True
  OWNDATA : False
  WRITEABLE : False
  ALIGNED : True
  WRITEBACKIFCOPY : False
  UPDATEIFCOPY : False

[0 1 2 3 4 5 6 7 8 9]

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Input In [23], in <cell line: 9>()
     20 print(att.flags)
     21 print(att.base)
---> 22 att += 1
     23 print(att)

ValueError: output array is read-only

Start with the attribute access path reached by f.attrs["att"] and compare it with the dataset read path used by f["arr"][:]. Run the MCVE against the repository to inspect the differing array flags; done means attribute arrays can be modified in memory consistently with dataset arrays and the behavior is covered by a regression test.