API: default to microseconds for Timestamp/Timedelta creation from numeric input

[jorisvandenbossche]

After the recent PRs for defaulting to microseconds when possible in date_range/Timestamp/to_datetime etc, there is one main type of input to those functions that still result in nanoseconds, which is numeric input.

For example:

>>> pd.to_datetime([1, 2, 3], unit="s")
DatetimeIndex(['1970-01-01 00:00:01', '1970-01-01 00:00:02',
               '1970-01-01 00:00:03'],
              dtype='datetime64[ns]', freq=None)

(there is the, separate, confusing part that the unit keyword here is define how to interpret the input, and not to specify the unit of the return value)

As a starter, I think there is no reason (apart from that it is currently implemented like that) to return nanoseconds here (unless the input are nanoseconds).
So I would at least default to microseconds when possible.

But then another option would also be to do as minmal conversion as possible, and for example if the input is seconds (or minutes, hours, etc) also return seconds, etc. This gives more variation in the resulting dtype, but causes less conversion.

(I have a slight preference to as much as possible return the same dtype)

[jorisvandenbossche]

cc @jbrockmendel

[jbrockmendel]

I'm thinking something similar, but am still tripped up by floats. We could change the behavior for int-dtype and for object-dtype-of-ints but keep float behavior the same? (xref #53445)

(there is the, separate, confusing part that the unit keyword here is define how to interpret the input, and not to specify the unit of the return value)

xref #62440

But then another option would also be to do as minmal conversion as possible, and for example if the input is seconds (or minutes, hours, etc) also return seconds, etc

I lean toward this option.

[jorisvandenbossche]

Small correction, it seems pd.Timestamp already takes the input unit:

>>> pd.Timestamp(1, unit="s").unit
's'
>>> pd.to_datetime(1, unit="s").unit
'ns'

(and for date_range, if you pass integers, it sees it as nanoseconds, the unit keyword there is for the return value ..)

It is for Timedelta that both are still giving nanoseconds:

>>> pd.Timedelta(1, unit="s").unit
'ns'
>>> pd.to_timedelta(1, unit="s").unit
'ns'

[jorisvandenbossche]

I'm thinking something similar, but am still tripped up by floats. We could change the behavior for int-dtype and for object-dtype-of-ints but keep float behavior the same? (xref #53445)

Yes, indeed we probably have to keep floats default to nanoseconds, or we wouldn't have a consistent dtype, but very value-dependent behaviour .. For a case like this, we would need an "output" unit indication if we would ever want to change the default from nanos to microseconds