Bloom filters

Bloom filters encode membership in bit arrays. A negative result means the item was not inserted into the intact filter; a positive result means possibly present, because false positives are expected. They cannot enumerate members or delete an individual item. Use them to skip unnecessary authoritative lookups, not to enforce uniqueness or access control.

Reserve and query

target/release/ffkv-cli -p 6382 BF.RESERVE docs:bloom 0.01 1000
target/release/ffkv-cli -p 6382 BF.ADD docs:bloom alice
target/release/ffkv-cli -p 6382 BF.MADD docs:bloom bob carol
target/release/ffkv-cli -3 --json -p 6382 BF.EXISTS docs:bloom alice
target/release/ffkv-cli -p 6382 BF.MEXISTS docs:bloom alice dave
target/release/ffkv-cli -3 --json -p 6382 BF.INFO docs:bloom

RESERVE takes error rate then initial capacity. Here 0.01 is a configured false-positive target, not a guarantee for every query/data distribution. ADD reports whether the filter changed, not authoritative proof of a newly unique member: a false positive can make an unseen item look already present.

BF.ADD can create a filter implicitly, but explicit reserve makes sizing visible. BF.INSERT supports batched items and creation controls; NOCREATE is useful when silently creating a missing filter would hide a deployment mistake. Non-scaling and expansion settings determine whether capacity can grow.

Capacity and errors

Scalable filters add subfilters as necessary under bounded expansion. NONSCALING makes exhaustion an error rather than unbounded growth; it conflicts with a nonzero explicit expansion. Error rates must be finite and strictly between zero and one, capacity positive, and backing arrays fit a 64 MiB per-key bound. Hash-count/expansion/work limits and request budgets also apply. Existing incompatible types return WRONGTYPE.

Do not interpret BF.INFO item counts as an exact distinct-member inventory. To expire a time window, expire/delete the entire filter or use rotating keys. If false positives would drop real work, check an exact store before making that decision; a Bloom filter is not an idempotency ledger.

Durability and interchange

With a data directory, resolved filter mutations participate in WAL/checkpoints/backups; without it the filter is volatile. Hybrid mode does not move filter arrays to SSD. SCANDUMP/LOADCHUNK uses ffkv’s versioned checksummed FFKF format with bounded pages, not RedisBloom’s packed-memory dump format. Freeze writes during an export; use ffkv-to-ffkv roundtrips only, not RedisBloom binary migration.

See BF.RESERVE, BF.INSERT, sets, Cuckoo filters, and migration.