The false positive rate can be made arbitrarily small. I have an implementation with a fp rate of 1:1.000.000 with a filter size of just 1.8 GB (https://github.com/adewes/have-i-been-bloomed). The cost of lowering the rate is logarithmic so you could go to much lower values for little cost (1:1.000.000.000 would be 2.8 GB) so false positives are not really a problem in practice. If you really want you could still perform an exact check against a DB if the filter returns true to rule out false positives with certainty, though at one false positive for one billion requests this might be exaggerated.