when I originally added first class timedelta support I allowed .astype('m8[unit]') to return a float, except if unit==ns. IOW it was de-facto the same a dividing by Timedelta(1, unit=unit)

In [1]: s = Series(pd.to_timedelta(['1 day', '1 minutes', '1 second'])) In [2]: s Out[2]: 0 1 days 00:00:00 1 0 days 00:01:00 2 0 days 00:00:01 dtype: timedelta64[ns] In [3]: s.astype('timedelta64[s]') Out[3]: 0 86400.0 1 60.0 2 1.0 dtype: float64 In [4]: s / pd.Timedelta('1 s') Out[4]: 0 86400.0 1 60.0 2 1.0 dtype: float64 In [5]: s.astype('timedelta64[ns]') Out[5]: 0 1 days 00:00:00 1 0 days 00:01:00 2 0 days 00:00:01 dtype: timedelta64[ns] 

note that for [5], however we just return a timedelta64[ns] (which happens to be the underlying data representation).

#19224 fixes construction of non-ns (and preserves the astype freq conversions. I think that this is confusing to the user and we should revert this and have a m8[unit] just work the same as it does for M8[unit] namely that you get back a timedelta dtype (and not a float), that is the same

So [3] would be the same as [5]

We do this for datetime types now

In [13]: Series(pd.date_range('20170101', periods=3)).astype('M8[s]') Out[13]: 0 2017-01-01 1 2017-01-02 2 2017-01-03 dtype: datetime64[ns] 

sure this is slightly convenient but a bit confusing.

I would propose that we provide a deprecation warning and change it in the next version.