freaking fast KV: a Redis-compatible datastore in Rust, built for queues and background jobs.
Handbook ·
Quickstart ·
Try it in your browser ·
Command reference ·
Benchmarks
ffkv speaks RESP2/RESP3, so existing Redis clients, redis-cli and Sidekiq work
unchanged. It adds the parts a background-job backend usually bolts on: durable
logs, a RAM/SSD hybrid mode, native leased jobs and a management console.
- Redis 8.10 command surface. All 486 catalog entries and 19 datatype families, from strings, lists and streams to JSON, vector sets, time series and probabilistic sketches.
- Fast on multi-core. Faster than Dragonfly in all 27 matched benchmark cases. See Performance.
- Durable when you want it. A checksummed WAL with
strict(fsync per write) oreverysec, plus checkpoints, online backups and offline restore. - Built for jobs. Blocking list queues, native leased jobs with acknowledgements, and Lua, transactions and Pub/Sub. The pinned Sidekiq integration is tested.
- Operable. TLS, ACLs, Prometheus metrics, an authenticated React console, native replication, cluster mode and an MCP sidecar for AI tools.
Status: alpha. ffkv is extensively tested (Rust tests plus a BDD suite that covers every catalog entry), but it is not production-certified or fully Redis-binary-compatible. Read compatibility before relying on it.
No install needed to look around: the browser playground runs ffkv's real command engine in your browser via WebAssembly (in-memory only).
docker run -d --name ffkv -p 6379:6379 -v ffkv-data:/var/lib/ffkv ghcr.io/akhil-gautam/ffkv
docker exec ffkv ffkv-cli PING # PONG
docker exec ffkv ffkv-cli SET greeting 'hello ffkv' # OKThe image runs on amd64 and arm64, uses durable strict mode, and generates TLS
and an admin password on first start, stored in the volume. To connect from your machine:
docker cp ffkv:/var/lib/ffkv/docker/ca.crt ./ffkv-ca.crt
REDISCLI_AUTH="$(docker exec ffkv cat /var/lib/ffkv/docker/admin-password)" \
redis-cli --tls --cacert ./ffkv-ca.crt --user admin PINGSettings such as your own password, extra certificate names and Compose with the web console are covered in the Docker guide.
You need Rust 1.99 or newer and a C toolchain. No Redis installation is required.
cargo build --release --bins
target/release/ffkv --bind 127.0.0.1:6380 # memory-only; data is gone on exitIn another terminal:
target/release/ffkv-cli PING # PONG
target/release/ffkv-cli SET greeting 'hello ffkv' # OK
target/release/ffkv-cli INCR visits # 1
target/release/ffkv-cli LPUSH jobs '{"id":1}' # 1
target/release/ffkv-cli BRPOP jobs 0 # jobs, {"id":1}ffkv-cli defaults to 127.0.0.1:6380. Any Redis client works too, for example
redis-cli -p 6380.
Keep your data. Add a data directory and choose a durability policy:
target/release/ffkv --bind 127.0.0.1:6380 --data-dir ./ffkv-data --durability strict| Goal | Flags | Guide |
|---|---|---|
| Fsync before every reply | --data-dir DIR --durability strict |
Persistence |
| Fsync about once a second | --data-dir DIR --durability everysec |
Persistence |
| Spill large list items and job bodies to SSD | --storage hybrid --data-dir DIR |
Hybrid storage |
| TLS and named users | --tls-cert, --tls-key, --acl-file |
Security |
| Web console and management API | --management-bind, --management-token-file |
Console |
| Docker Compose / Kubernetes | deploy/ |
Docker, Kubernetes |
Every flag is listed in the configuration reference.
These are commands per second for 32 clients at pipeline depth 16. Each figure is the median of two runs, memory-only. Each engine was a fresh server on four pinned CPUs.
| Workload | ffkv | Dragonfly 2.0.2 | Redis 8.2.1 | vs Dragonfly | vs Redis |
|---|---|---|---|---|---|
| GET 256 B | 2,941,042 | 2,603,375 | 2,082,643 | 1.13× | 1.41× |
| SET 256 B | 2,634,198 | 2,427,828 | 1,736,991 | 1.09× | 1.52× |
| GET 4 KiB | 1,840,583 | 1,155,995 | 795,787 | 1.59× | 2.31× |
| SET 4 KiB | 1,641,558 | 1,097,966 | 494,416 | 1.50× | 3.32× |
| INCR | 3,076,196 | 2,971,252 | 2,079,936 | 1.04× | 1.48× |
| HGET | 2,826,854 | 2,599,985 | 1,689,266 | 1.09× | 1.67× |
| HSET | 2,628,055 | 2,447,099 | 1,514,368 | 1.07× | 1.74× |
| ZADD | 2,398,487 | 1,835,926 | 548,041 | 1.31× | 4.38× |
| ZRANGE (10 members) | 1,786,067 | 1,485,170 | 1,094,335 | 1.20× | 1.63× |
Across all 27 cases (nine workloads × three client profiles), ffkv beat Dragonfly in 27/27 and Redis in 18/27. With 32 clients and no pipelining it was 1.35–1.81× Dragonfly. With a single client it was 1.61–1.67× Dragonfly. In that single-client case, single-threaded Redis stays 0.5–4% ahead.
All engines used one ARM64 Linux VM on an Apple M4, with redis-benchmark 8.2.1
on separate CPUs. The results are memory-only with two runs per case, so they
show no statistical significance. The
benchmark guide has every
table, the cross-check with other load generators, the method and the
reproduction commands. Raw data is in
benchmarks/results/.
flowchart LR
clients["Redis clients<br/>ffkv-cli · Sidekiq"]
mcp["ffkv-mcp<br/>(MCP sidecar)"]
console["Management console<br/>REST + live events"]
subgraph server["ffkv server"]
direction LR
listener["Listener<br/>TCP / TLS"]
subgraph frontends["Frontend runtimes (one per CPU)"]
direction TB
parse["RESP parser<br/>+ ACL / pause checks"]
route{"Route by key"}
inline["Inline execution<br/>under shard lease"]
reply["Reply encoder<br/>+ output ledger"]
parse --> route
route -- "single shard,<br/>uncontended" --> inline --> reply
end
subgraph core["Main runtime"]
direction TB
coord["Coordinator<br/>cross-shard atomicity"]
workers["Shard workers<br/>batch contended keys"]
shards[("Shards 1..N<br/>keyspace")]
gate["Ordered commit gate"]
coord --> shards
workers --> shards
end
persist[("WAL · checkpoints<br/>hybrid segments")]
repl["Replication ·<br/>cluster bus"]
end
clients --> listener -- "round-robin accept" --> parse
mcp --> listener
console --> server
route -- "contended shard" --> workers
route -- "multi-key" --> coord
inline --> shards
shards --> gate --> persist
gate --> repl
- Connections: each connection is pinned to one frontend runtime. It parses and runs most commands inline under an exclusive shard lease, with no extra hop.
- Contended keys: when several frontends compete for one shard, its worker runs the queued commands in batches. A hot key therefore doesn't bounce between CPUs.
- Cross-shard commands: commands such as
MSET,RENAMEand set stores go through the coordinator, which keeps them atomic. - Durability: a write is committed through one ordered gate to the WAL before any reader sees it. Replication streams the same resolved effects.
More detail: architecture.
| I want to… | Read |
|---|---|
| Install and try it | Browser playground, Installation, Quickstart, CLI |
| Use a datatype | Datatypes, Command reference |
| Run jobs | List queues, Native jobs, Sidekiq |
| Operate it | Backup & restore, Replication, Clustering, Observability |
| Connect AI tools | MCP sidecar |
| Know the limits | Compatibility, Limits |
cargo test --locked --all-targets
cargo clippy --locked --all-targets -- -D warnings
npm --prefix apps/docs ci && npm --prefix apps/docs run dev # handbook at localhost:4321 (needs: rustup target add wasm32-wasip1)See development,
testing (including the BDD
suite) and benchmarks. The
handbook source is in apps/docs.
ffkv uses Semantic Versioning. Changes are listed in CHANGELOG.md, and the release steps are in the development guide.