Data types

ffkv’s current taxonomy contains 19 families, counting bitmaps and bitfields separately even though both operate on string bytes. Native FQ.* queues are an ffkv extension beyond that catalog. Search is a separate indexing capability, not an extra family.

Choose a model

Family	Use it for	Result model
Strings	Bytes, counters, expiring values	Exact bytes/integer arithmetic
Lists	Ordered payloads and remove-on-pop queues	Exact order
Sets	Membership and algebra	Exact unique members
Hashes	Field/value records and field TTL	Exact fields
Sorted sets	Leaderboards, ordered schedules	Exact member/score index
Bitmaps	Dense boolean flags	Exact addressed bits
Bitfields	Packed fixed-width integers	Exact within chosen overflow policy
Geospatial	Point proximity and regions	Encoded coordinates and computed distance
HyperLogLog	Unique cardinality with bounded space	Estimate, no member enumeration
Streams	Append history and consumer groups	Ordered entries/pending ownership
JSON	Path-addressable document trees	Exact document operations
Arrays	Sparse unsigned-index byte strings	Exact slots, holes, and cursor
Time series	Timestamped samples and aggregation	Samples with floating-point aggregates
Vector sets	Embedding similarity	Approximate HNSW search; explicit exhaustive mode
Bloom filters	Compact membership prefilters	Possible false positives
Cuckoo filters	Fingerprint membership/counting/deletion	Probabilistic membership/counts
Count-min sketch	Frequency estimates	Collision-biased estimates
Top-K	Heavy hitters	Approximate ranked candidates
t-digest	Quantiles/CDF over measurements	Approximate distribution

Prefer an exact set/hash/counter when an incorrect answer changes authorization, money, or irreversible work. Probabilistic structures are useful for ranking, telemetry, and prefiltering with an authoritative fallback; they are not interchangeable with exact collections.

Shared semantics

Examples use target/release/ffkv-cli -p 6382 against the quickstart instance. Use a fresh dataset; commands mutate namespaced example keys. A command on an incompatible existing type generally returns WRONGTYPE; this is different from a missing key. Some families share physical representations: bitmaps/bitfields/HyperLogLog use strings, and geo uses sorted sets. Avoid unrelated commands that corrupt their encoded meaning.

All nineteen families participate in key TTLs, WATCH, ACLs, WAL recovery, checkpoints, and backups in both storage profiles. No data directory means no restart durability. Hybrid offloads only large list and native-job bodies: other families stay in RAM. Request/reply limits apply independently of family limits. See persistence, limits, and compatibility.

A SCAN-family cursor is opaque and not a snapshot; COUNT is a hint. ARSCAN is a different index/value range API. Prefer bounded reads to whole-collection commands for large data.

For workflows rather than individual types, see list queues, native leased jobs, transactions, and search indexes.