Original ticket http://projects.scipy.org/numpy/ticket/1912 on 2011-07-18 by @chrisjordansquire, assigned to unknown.

I'm using the 2.0.0 development version of numpy on ububtu 10.10 .

When string arguments for the dtype are given for a record array, and one of the individual dtypes is datetime64[D], the day part is lost and the datetime become generic. This makes it impossible to assign anything to it.

For example, look at
np.zeros(4, dtype = ','.join(['datetime64[D]', 'i4']))

This gives an array with dtype dtype=[('f0', <M8), ('f1', '<i4')], where the <M8 should be <M8[D].

A different bug happens when putting datetimes into masked arrays. For example,
np.ma.zeros(4,dtype= 'datetime64[D],i4')
gives a ValueError: mismatch in size of old and new data-descriptor with traceback

/home/intern/my26/lib/python2.6/site-packages/numpy/ma/core.pyc in call(self, a, _args, *_params)
7142 _extras[p] = params.pop(p)
7143 # Get the result

-> 7144 result = self._func.call(a, _args, *_params).view(MaskedArray)
7145 if "fill_value" in common_params:
7146 result.fill_value = _extras.get("fill_value", None)

/home/intern/my26/lib/python2.6/site-packages/numpy/ma/core.pyc in array_finalize(self, obj)
2794 if self.dtype.names:
2795 if self._fill_value is None:
-> 2796 self._fill_value = _check_fill_value(None, self.dtype)
2797 return
2798

/home/intern/my26/lib/python2.6/site-packages/numpy/ma/core.pyc in _check_fill_value(fill_value, ndtype)
392 if fields:
393 descr = ndtype.descr
--> 394 fill_value = np.array(_recursive_set_default_fill_value(descr),
395 dtype=ndtype,)
396 else:

/home/intern/my26/lib/python2.6/site-packages/numpy/ma/core.pyc in _recursive_set_default_fill_value(dtypedescr)
360 deflist.append(tuple(_recursive_set_default_fill_value(currenttype)))
361 else:
--> 362 deflist.append(default_fill_value(np.dtype(currenttype)))
363 return tuple(deflist)
364

ValueError: mismatch in size of old and new data-descriptor