ENH: infer freq in timedelta_range

[simonjayhawkins]

xref #36582, #35897, #36595, #32377, #31195

there may be cases where a frequency should not be inferred by timedelta_range due to rounding issues

e.g. pd.timedelta_range('0s','1s', periods=31)

The docs https://pandas.pydata.org/pandas-docs/dev/reference/api/pandas.timedelta_range.html begin 'Return a fixed frequency TimedeltaIndex...' so having None (no inferred frequency) maybe not desirable.

cc @jbrockmendel @jreback @phofl

[simonjayhawkins]

so having None (no inferred frequency) maybe not desirable.

for instance should pd.timedelta_range('0s','1s', periods=31) perhaps raise with a message along the lines 'ValueError: cannot infer a fixed frequency from start, end and periods due to rounding errors` or is a fractional freq a feasible option?

[simonjayhawkins]

maybe could only infer frequency if no rounding errors...

> c:\users\simon\pandas\pandas\core\arrays\timedeltas.py(271)_generate_range()
-> freq = to_offset(td)
(Pdb) td
Timedelta('0 days 00:00:00.033333333')
(Pdb) td.value
33333333
(Pdb) start.value
0
(Pdb) end.value
1000000000
(Pdb) periods
31
(Pdb) assert start.value + td.value * (periods - 1) == end.value
*** AssertionError
(Pdb) start.value + td.value * (periods -1)
999999990

[jbrockmendel]

or is a fractional freq a feasible option?

Not with the current implementation with TDI backed by i8.

maybe could only infer frequency if no rounding errors...

Could warn in the cases with rounding errors.

Note: this will also affect date_range.