Open
Description
Is your feature request related to a problem? Please describe.
NVTX allows attaching arbitrary data to NVTX ranges. We should take advantage of this in libcudf. We could send many types of data, and some choices could risk leaking sensitive data. One idea is we could start by sending the row counts for any input tables or columns - this would be useful for understanding data partitioning and other details about libcudf calls.
Describe alternatives you've considered
All ideas for how to use extended payloads in libcudf are welcome!