False positives from Bloom filters could be decreased exponentially at every GC operation (I'm not sure how quickly, but more or less for free) by using a new k

False positives from Bloom filters could be decreased exponentially at every GC operation (I'm not sure how quickly, but more or less for free) by using a new key for the Bloom filter hashes (e.g. the current timestamp) so that every block that was retained as a false positive last time is aggregated with different pseudorandom sets of other blocks, with a high probability of getting lucky.