Developer Releases KiviDB, an Open-Source Redis-Compatible In-Memory Database and Vector Cache
KiviDB, an AGPLv3 Redis-compatible in-memory database and vector cache, launched with a 21.2M pipelined GET benchmark.
The author said KiviDB began in February 2026 as a way to learn Rust and to build a product without garbage collection pauses and without the memory growth after snapshots and out-of-memory events on running pods that he had faced with Dragonfly. He said he used AI extensively but made the design decisions himself based on tests drawn from his troubleshooting history. KiviDB shards its keyspace across threads and is multi-threaded. It performs best on Linux because it uses io_uring through the tokio_uring crate, with a tokio fallback for other operating systems. The author said he also built a managed cloud service for KiviDB across AWS, GCP and Microsoft Azure, and he chose AGPLv3 rather than a source-available license.
In the benchmark, the author used two c7gn.metal servers in the same placement group in us-east1 and the memtier-benchmark tool, with 10 million keys of 100 bytes each and 32 cores in use. For a pipelined GET workload at a pipeline depth of 16, KiviDB reported 21.2 million operations per second, compared with 9.1 million for Dragonfly 2.0, 2.4 million for Redis 8.10 and 4.7 million for a 32-shard Redis Cluster. When not pipelined, Dragonfly performed slightly better at 3.35 million operations per second versus KiviDB at 3.07 million operations per second.
Memory use per key was 177 bytes for KiviDB, 134 bytes for Dragonfly and 178 bytes for Redis, according to the post. The author said vector ingestion was slower than alternatives, but querying was faster than most alternatives at lower ef and slower than Dragonfly at high ef. He said KiviDB supports about 98-99% of Redis 8.10 commands, with a few issues tracked on GitHub, and invited contributors.
The project website is kividb.io. The author said the quickest way to try KiviDB is to run the Docker command docker run -d -p 6380:6380 kividb/kividb. He said he is looking to build a community around the project and is open to constructive criticism and informed opinion.