Show HN: KiviDB, a faster Redis-compatible memory database and vector cache

KiviDB is a high-performance, multi-threaded Redis-compatible key-value store written in Rust that exposes RESP2/RESP3 so existing Redis clients work unchanged. It shards work across cores with an in-memory design, supports native snapshots and an optional append-only file, and includes built-in full-text (BM25) and vector search (FLAT and HNSW) with Redis Stack-style FT.* commands. Vectors are stored as raw FLOAT32 little-endian bytes in hash fields. Builds are configurable via Cargo features to include Lua scripting and TLS; the server listens on port 6380 by default, uses io_uring on Linux when available, and is released under AGPL‑3.0.

Benchmarks on AWS Graviton3 (server confined to 32 cores) show pipelined GET throughput of 21.2M ops/s (pipeline 16), versus 9.1M for Dragonfly and 2.4M for Redis 8, with p99 latency 0.36 ms for GET. Vector-search tests on glove-25-angular (1.18M vectors) report 59.3k q/s at ef=64 with recall@10 0.943, outperforming Dragonfly and Redis in many ef settings; ingest took 31s and memory after ingest was 1.66 GB (higher than competitors). Snapshots persist keys, index schemas and HNSW graphs so search quality is preserved across restarts. The published figures come from a single controlled run; users should benchmark on their own hardware and workloads.