KiviDB Claims 21.2 Million GETs a Second, but Its Own Table Has Dragonfly Ahead Without Pipelining

KiviDB, a Rust in-memory database that speaks the Redis protocol, publishes a benchmark in which it serves 21.2 million pipelined GETs a second against 2.4 million for Redis 8.10 and 9.1 million for Dragonfly 2.0.2. The same README shows Dragonfly ahead of KiviDB on unpipelined throughput, at higher vector-search accuracy settings, and on memory per key.

The repository, which reached Hacker News as a Show HN post, has 208 stars, three forks and an empty contributors list. Its copyright belongs to KIVIDB PRIVATE LIMITED, and the page names no individual author. Every figure below is vendor-supplied: the KiviDB README links the company's own kivi-benchmark repository for method, and names no independent party who ran it.

What the benchmark ran on

The server was an AWS c7gn.metal Graviton3 machine with 64 vCPUs, confined to 32 cores, with a second c7gn.metal in the same placement group generating load. Each engine ran alone. The key-value test used memtier_benchmark with 10 million keys, 100-byte values, 30 seconds per scenario and the mean of two rounds, at 300 connections when pipelined and 1,200 when not.

Versions were KiviDB 1.1.0, Dragonfly 2.0.2 and Redis 8.10, with Redis Cluster at 32 shards as a fourth column. Valkey, the other Redis-compatible server people compare against, was not benchmarked.

KiviDB's --threads default is cores minus two. The README gives no thread settings for Redis or Dragonfly, so a reader cannot tell whether the comparison gave each engine the same CPU.

1. KiviDB 1.1.021.2 million ops/s
2. Dragonfly 2.0.29.1 million ops/s
3. Redis Cluster, 32 shards4.7 million ops/s
4. Redis 8.102.4 million ops/s

Source: KiviDB README benchmark section, accessed 2026-10-19

The headline ratio depends on which Redis is the baseline. Against the single Redis 8.10 process, 21.2 million is 8.8 times higher. Against the 32-shard Redis Cluster in the same table, which reaches 4.7 million, it is 4.5 times. Against Dragonfly, a multithreaded design with its own shared-nothing architecture, it is 2.3 times.

Pipelining sends 16 commands per round trip, so this scenario favours engines that can keep many cores busy with batched work. It is a legitimate workload, but it is not the one an application making single requests sees.

Where KiviDB's own table has it behind

Without pipelining, KiviDB served 3.07 million GETs a second and Dragonfly 3.35 million. On SETs the figures were 3.12 million and 3.38 million. Redis 8.10 managed 1.32 million GETs and 1.54 million SETs.

Memory per key was 177 bytes for KiviDB, 134 for Dragonfly, 178 for Redis 8.10 and 183 for Redis Cluster. The README marks its comparison table "As of October 2026" and asks readers to report errors.

Vector search, glove-25-angular	KiviDB 1.1.0	Dragonfly 2.0.2	Redis 8.10
ef 64, queries/s (recall@10)	59.3k (0.943)	41.3k (0.912)	31.8k (0.912)
ef 512, queries/s (recall@10)	18.7k (0.998)	29.8k (0.995)	13.7k (0.995)
Ingest and index build	31 s	13 s	38 s
Memory after ingest	1.66 GB	1.13 GB	1.07 GB

At ef 64, KiviDB's recall of 0.943 is also higher than the 0.912 the other two reach, so its 59.3k queries a second is not a like-for-like comparison with their 41.3k and 31.8k. At ef 256, KiviDB scores 31.5k against Dragonfly's 36.2k.

The README's own caution reads: "Treat published figures as one controlled capture, not a guarantee, and measure on your own hardware and workload."

Licence: AGPL, a commercial option and a CLA

KiviDB is AGPL-3.0-only. The README says "Using KiviDB does not put your application under the AGPL," because network clients are separate programs, but anyone who modifies it and offers it as a service must offer that source to the service's users. KIVIDB PRIVATE LIMITED also sells a commercial licence for organisations that cannot accept AGPL, and contributors sign a CLA, a contributor licence agreement.

That is the same shape Redis chose. Rowan Trollope wrote on the Redis blog on May 1, 2025 that Redis added AGPL as an option starting with Redis 8, after moving to the SSPL in March 2024. A CLA plus a commercial licence also appears in the VoiceStudio repository, where an issue asks contributors to sign over commercial relicensing rights.

What is missing before anyone runs it in production

The README makes no production-readiness claim. Lua scripting works only in builds with the lua feature, and TLS needs the tls feature: a build without it refuses to start if given --tls-port. While search indexes rebuild at startup, most commands return LOADING, and only commands such as PING and INFO work.

Persistence is a dump.kdb snapshot plus an optional append-only file. With --aof, a crash loses at most about one second of writes. The default port is 6380, not Redis's 6379, and the supported command list lives in a COMMANDS.md file rather than the README.

Dragonfly, by comparison, has 31.8k stars and, per its repository, implements about 185 Redis commands, roughly the Redis 5.0 API. KiviDB's 208 stars measure attention, not adoption, as would any count; the site's report on CLI-Anything's 54,400 stars covers a repository at the other end of that scale.

KiviDB's next test is outside its own repository: no one other than the company has been named as running its numbers, and the README invites readers to measure on their own hardware.