Mehran Toosi · Principal Software Engineer · Copenhagen

2025 – now · Rust · open source · 197 stars · about 7,600 crates.io downloads

FeOxDB

problem

Caches, sessions, queues and other derived state have to be fast and should survive a restart, but rarely need an fsync on every write. Embedded stores usually either block on fsync or keep everything in memory.

solution

A lock-free in-memory index with write-behind persistence through io_uring, a durability window that is bounded and documented, and an explicit flush for writes that must be synchronous. feox-server adds the Redis protocol on top.

results

GET in about 200 ns and INSERT under 1 µs on an M3 Max. feox-server runs about twice Redis’s throughput on mixed workloads. It is also the storage engine under Orvena. feoxdb has about 200 stars on GitHub and feox-server about 85.

Architecture

Reads are answered from an SCC hash map and never touch the disk; a crossbeam skip list serves ordered scans. Writes update the index and land in per-shard buffers, and background workers batch them to the device through io_uring.

reads in blue, writes in amber · latencies from the README’s M3 Max benchmarks · flush cadence slowed down

durability

A durability window you can calculate

Buffers flush every 100 ms or when a shard holds 16 MB. Shards drain in parallel, so the worst case is 100 ms plus the time to write 16 MB at the device’s 4 KB random-write speed, with at most 16 MB per shard at risk. A caller that needs fsync semantics calls flush and waits.

four shards on two example devices · formula and examples from the README

recovery

Crash-safe updates

An update persists the new record before it frees the old record’s sectors, so a crash at any point leaves either the old value or the new one. Records and metadata are checksummed, allocation is journaled, and on open the store rebuilds its index and free-space map by scanning the device.

one update to key K, killed at three different moments, then reopened · sector layout simplified

benchmark

feox-server against Redis

feox-server runs one thread per core on top of FeOxDB. Same machine, redis-benchmark with 1M random keys, 50 clients and pipeline depth 64, plus mixed workloads. Pub/Sub doesn’t touch storage and is slightly slower than Redis.

feox-serverRedis

from the feox-server README benchmark tables