The current way is sensitive to floating point precision (#1823, #4465). We could make a smarter check: compare dataset values vs expected linear interpolated with a relative tolerance, this should be more resilient. related to #4466, #1828
