There is no set rule, but a 50/50 split between both variants is a good default.
If you're testing a non-standard variant with a sizeable difference from your standard, you can start with less traffic on the non-standard variant. Though you should keep in mind that it will take longer to converge if that's the case.