Single node Valkey in mostly memory safe Rust. Verified against Valkey's test suite, uses memory safe TLS with rustls.

Live benchmarks & coverage → valdr.dev

Valdr is a port of Valkey aiming to be fully compatible with Redis/Valkey clients in memory safe Rust. The motivation for this project is to attempt to build mostly memory safe alternatives to core web infrastructure and to explore architecture choices that will enable faster performance in the long run.

This repo heavily leveraged coding agents in the process. This was largely inspired by the changing landscape of memory safety attacks as agentic cyber capabilities increase.

Performance

Server config: no .conf file — both servers are launched from explicit flags by the bench harness, persistence off, otherwise stock defaults.

Per-command (default valkey-benchmark suite)

• Wins (ratio ≥ 1.2×): ping_inline, ping_mbulk, set, get, lrange_100, lrange_300, lrange_500, lrange_600, mset, mget.
• Parity (0.95×–1.2×): incr, lpush, lpop, sadd, hset, spop, zadd, zpopmin.
• Behind (< 0.95×): rpush, rpop, xadd, fcall

Pipeline-depth curve (GET/SET/PING/INCR at p=1/16/100)

How the numbers are produced

• Warmup: 1,000 PING_MBULK requests (1 client, pipeline 1) before every measured row.
• Refresh: make bench-release (fresh local artifacts) then make site-data, which regenerates docs/perf-data.json, docs/perf.svg, docs/perf.png, and rewrites the table above from it. The valdr.dev landing page fetches the same JSON, so the site and this README can never disagree — they share one source of truth.
• Switch the Valkey adversary: git -C reference/valkey checkout <tag> && make -j MALLOC=jemalloc BUILD_TLS=no, then re-run the refresh.

Coverage

Valkey's own test suite full breakdown by category → valdr.dev/coverage.html.

Counted assertions (54-file run)  [████████████████████] 100.0%   3,035 / 3,035
Single-node core blocks           [████████████████████]  99.96%  2,525 / 2,526
Full upstream suite               [███████████░░░░░░░░░]   59%     2,534 / 4,281

The only unproven single-node core source block is unit/replybufsize.tcl, which the current tag policy filters to a 0/0 run. The full-suite figure is lower because ~41% of upstream tests cover surfaces valdr does not claim yet: cluster, modules, Sentinel, HA replication, and platform-specific integration.

Status

Compatibility

Features

Run

cargo build --release
./target/release/redis-server --port 6379 --bind 127.0.0.1

Docker

docker pull ghcr.io/ianm199/valdr:alpha
docker run --rm -p 6379:6379 ghcr.io/ianm199/valdr:alpha

Benchmark Commands

A rough, runnable head-to-head against the official valkey/valkey:8-alpine image (both sides ship jemalloc). Benchmarking inside a container is noisy — especially Docker Desktop on macOS, where per-run throughput can swing ~2×. The Performance chart above is the measurement of record: native binaries, warmed, jemalloc on both sides, vs Valkey 9.1.0. For a stable head-to-head, run this on a Linux host.

docker network create valdr-bench

docker run -d --rm \
  --name valkey-ref \
  --network valdr-bench \
  valkey/valkey:8-alpine

docker run -d --rm \
  --name valdr \
  --network valdr-bench \
  ghcr.io/ianm199/valdr:alpha

sleep 1

docker run --rm \
  --network valdr-bench \
  valkey/valkey:8-alpine \
  valkey-benchmark \
    -h valkey-ref \
    -p 6379 \
    -n 100000 \
    -c 50 \
    -P 100 \
    -d 64 \
    -t ping_inline,ping_mbulk,set,get,incr,lpush,rpush,lpop,rpop,sadd,hset,spop,zadd,zpopmin,lrange_100,lrange_300,lrange_500,lrange_600,mset,mget,xadd,function_load,fcall \
    --csv \
    --precision 3

docker run --rm \
  --network valdr-bench \
  valkey/valkey:8-alpine \
  valkey-benchmark \
    -h valdr \
    -p 6379 \
    -n 100000 \
    -c 50 \
    -P 100 \
    -d 64 \
    -t ping_inline,ping_mbulk,set,get,incr,lpush,rpush,lpop,rpop,sadd,hset,spop,zadd,zpopmin,lrange_100,lrange_300,lrange_500,lrange_600,mset,mget,xadd,function_load,fcall \
    --csv \
    --precision 3

docker rm -f valkey-ref valdr
docker network rm valdr-bench

Test Commands

bash scripts/setup-reference.sh
cargo build -p redis-server
cargo test --workspace
bash harness/oracle/smoke.sh --skip-build
python3 harness/oracle/rdb-diff --direction=all
bash harness/oracle/run-single-node-tcl-suite.sh --skip-build
bash harness/bench/official-warm-run.sh

The single source of truth for what these prove — counted TCL passes, single-node source-block coverage, and how the full 4,281-block upstream denominator is bucketed — is valdr.dev/coverage.html.

Roadmap

Full roadmap → valdr.dev/roadmap.html