Open MarcoGorelli opened 1 month ago
Thanks for the report!
Although this could be easily implemented by summing the timedelta components, I think there was a desire to implement a dedicated libcudf kernel for total_seconds
to avoid the n kernel launches to sum each individual component (IIRC cc @bdice you may have been apart of that passing discussion somewhere)
I think this might be simple, and may not require summing components. I think we can do a conversion/cast to duration_s
and then cast that as a float type to match pandas.
edit: we may need to cast to the smallest duration type (nanos?) and then divide by the appropriate scale factor (1e9), in order to retain subsecond information.
Is your feature request related to a problem? Please describe.
I wish I could use cuDF to do
.dt.total_seconds
on a timedelta columnDescribe the solution you'd like
.dt.total_seconds
Describe alternatives you've considered
Additional context
We're (for now) xfailing tests in Narwhals https://github.com/narwhals-dev/narwhals/pull/951