BUG: conversion precision on Timedeltas

[jreback]

I think this is rounding prematurely, though not entirely sure. Could also just be loss of precision when converting from a float -> int64.

In [12]: pd.to_datetime(1434743731.8770001,unit='s')
Out[12]: Timestamp('2015-06-19 19:55:31.877000')

In [13]: pd.to_datetime(1434743731.8770001,unit='s').value
Out[13]: 1434743731877000000

In [14]: (pd.Timestamp('2015-06-19 19:55:15.877000')+pd.Timedelta('1us')).value
Out[14]: 1434743715877001000

[jorisvandenbossche]

Copying from #19680

It happens in this function:

In [126]: pd._libs.tslibs.timedeltas.cast_from_unit(1.5e-6, 's')
Out[126]: 2000

because we round at 6 decimals:

pandas/pandas/_libs/tslibs/timedeltas.pyx

Lines 210 to 212 in 405ed25

	elif unit == 's':
	m = 1000000000L
	p = 6

pandas/pandas/_libs/tslibs/timedeltas.pyx

Line 234 in 405ed25

frac = round(frac, p)

Not fully sure why we choose 6 decimals to be allowed, and not 9 in case of 's' (and 6 in case of 'ms', 3 for 'us', ..)

[Sup3rGeo]

Isn't this the same as #12690?

[jorisvandenbossche]

Indeed, will close the other in favor of this one. Thanks for noticing!