We seem to get better results when we scale bridge IPs down to multiples of 1 instead of 8 and only mark bridges as blocked if they differ in the 'bad' direction by at least one standard deviation from the mean.
We had been storing all the data as a huge map under the key 'bridges' so that we could iterate over all keys in the map as a way of iterating over all bridges. This caused the database to grow way too big, so we don't do that anymore. Now, we store each bridge's data with the bridge fingerprint as the key, and we store a set of bridge fingerprints under the key 'bridges'.
Vecna