Efficient conditional selection with masks in very large dataframe

Question

I have a dataframe with some 2 million rows like this:

                    dt   num
0  2019-05-12 10:17:00   135
1  2018-01-16 21:32:00     5
2  2017-11-30 22