tm.makeObjectSeries returns a Series with dtype="datetime64[ns]"

[WillAyd]

>>> import pandas.util.testing as tm
>>> tm.makeObjectSeries()
...
t1Q4jcW3nW   2000-02-09
ElZCr3WTz2   2000-02-10
p43Nqg7u4D   2000-02-11
dtype: datetime64[ns]

@jbrockmendel and I were both surprised that calling makeObjectSeries returns a Series with a datetime dtype. I think it would be more logical for this function to actually return an object dtype Series

[WillAyd]

xref https://github.com/pandas-dev/pandas/pull/27838/files#r322964878

[jmg7173]

Hi, I want to contribute to this issue.

Is it okay just add dtype=object at Series generation for this issue?
Or is there more deep modification about this issue? (I'm not meaning about modification of test code related to makeObjectSeries)

[WillAyd]

Yea I think OK to start with that and see what breaks. Ideally may want to have it contain more than date time objects but not sure what effort that would take; just getting it to object dtype in the first place will be a good start

[jmg7173]

I modified it to object. But performing add at Series, it modifies dates and dtype as datetime64.

How can I forcibly prevent add/sub operator from modifying data and dtype?
Is it natural that automatically changing the dtype when some modification occurs?
I think data modification should not be happend if it modifies dtype as datetime64. (In case of datetime64, it raises exception)

[jbrockmendel]

How can I forcibly prevent add/sub operator from modifying data and dtype?

Can you be more specific about what you mean by "performing add at Series"?

[jmg7173]

I mean, at tests/arithmetic/test_object.py test_object_arr_invalid, it performs add operation.

I simulated it manually, it manipulates Series with dtype=object to dtype=datetime64[ns] and also, it also manipulates data also.

>>> obj_ser = tm.makeObjectSeries()
>>> obj_ser 
YipVeLYwc8   2000-01-03
...
wHtkS7TlO2    2000-02-08 00:00:00
tTYkAjpXUN    2000-02-09 00:00:00
5OGw3PNMfM    2000-02-10 00:00:00
14tcE5tlIl    2000-02-11 00:00:00
dtype: object

>>> operator.add(obj_ser, 1)
YipVeLYwc8   2000-01-04
...
tTYkAjpXUN   2000-02-10
5OGw3PNMfM   2000-02-11
14tcE5tlIl   2000-02-14
dtype: datetime64[ns]

[jbrockmendel]

what happens if you forget the datetimes altogether and fill it with strings?

[jmg7173]

It passes tests.

If I change 1 to "1", it appends "1" to all the data.
And it doesn't convert dtype.

>>> tm.makeObjectSeries() + "1"

sb7X5CxFoh    Zn6Ur8Sd6Z1
...
5ixyCT4dWC    ECQrisl5Ln1
vAvu2yKqsG    HN5LTCpT0V1
Dor4uHXigH    Mx1u5ukW141
dtype: object

[jmg7173]

Is it okay to change filling data as datetime64[ns] to String with dtype: object?

[jbrockmendel]

Is it okay to change filling data as datetime64[ns] to String with dtype: object?

Probably, yes. The one thing I want to double-check is that we aren't losing coverage for cases where we actually do want a datetime64 Series, which I think might be the case in the arithmetic test.

@WillAyd thoughts on how to track this down more generally? grepping shows only 11 usages of makeObjectSeries, but some of those are in fixtures.

[jmg7173]

As I searched, there is no usecase that objectSeries used just for datetime excluding arithmetic test.

I searched it including object_series fixture and instance of TestData(TestData.objSeries).
I hope I found everything :)

[WillAyd]

@WillAyd thoughts on how to track this down more generally? grepping shows only 11 usages of makeObjectSeries, but some of those are in fixtures.

Yea I think @jmg7173 if you can search for the fixture where this is used and check that as well would be helpful.

Mostly could just change and see what breaks though. The naming of this obviously means the intent is to deal with object data, so I don't think should break too much actually making this return an object dtype