We persist the IDs to disk with RocksDB itself when that happens, periodically pruning them away when the messages are complete. The Bloom filter is mostly an optimization - even though it does the job most of the time. You're right - we intentionally omitted further discussion of that piece.