Benchmarks
KiviDB 1.2.0 against Redis, Redis Cluster, Valkey, Dragonfly, Garnet and kvrocks: throughput, latency, memory, snapshots, vector search, caching and durability, where KiviDB loses, and the report.
KiviDB 1.2.0 against Redis, Redis Cluster, Valkey, Dragonfly, Garnet and kvrocks on the same Azure VMs, in the same hour (Azure VMs and AKS, 1 to 3 November 2026). Every number is in https://github.com/kividbio/kividb/blob/main/benchmarks/1.2.0/REPORT.md; a 15-page PDF report with every comparison, KiviDB Cloud's features and prices can be downloaded from this page.
KiviDB 1.2.0 results
| KiviDB 1.2.0 | Best other | Redis 8.10.3 | |
|---|---|---|---|
| GET 100 B, pipeline 16 | 10.45M ops/s | Garnet 10.69M | 1.41M |
| SET 100 B, pipeline 16 | 7.94M ops/s | Garnet 9.83M | 1.25M |
| MSET of 10 keys | 569K commands/s | Garnet 800K | 144K |
| Lists, hashes, sets, sorted sets, counters, pipeline 16 | 4.2-6.6M ops/s | Redis Cluster x16: level | 3.7-6.0x less |
| Bytes per key, 100 B values, 50M keys | 140.7 | Dragonfly 141.5 | 174.4 |
| Memory added by BGSAVE, 30M keys of 1 KB under writes | +10% | Dragonfly +0.4% | +89% |
| Filtered vector search, 100 clients | 4,478 q/s | Dragonfly 811 | 1,802 |
| Cache hit rate, issue #202 | 60.96% | 60.78% | |
| Acknowledged writes lost to kills, failovers, power loss, forced deletes | 0 | 0 |
Where KiviDB loses
- Garnet Faster on SET at pipeline 16 (9.83M against 7.94M ops/s) and on MSET of 10 keys (800K against 569K). KiviDB's MSET has the lower tail: p99 2.89 against 7.90 ms.
- Redis Cluster With 16 shards, one per vCPU, ahead on HSET and INCR at pipeline 16, where the load generator is the limit.
- Dragonfly Uses less memory per vector and during a snapshot (+0.4% against +10%), and fewer bytes per key for ASCII values (137.1 against 149.2).
- Redis Ingests vectors faster, and serves 3-9% more cache GETs a second in the #202 test.
KiviDB 1.2.0 on bare metal
AWS Graviton3 bare metal: the server on a c7gn.metal confined to 32 cores, the load generator on a second c7gn.metal in the same placement group, one engine at a time. KiviDB 1.2.0, Garnet 2.2.5 (shown only where KiviDB leads it), Dragonfly 2.0.2, Redis 8.10.3. memtier_benchmark, 10 million keys of 100 bytes.
Throughput
| Run | KiviDB | Garnet | Dragonfly | Redis |
|---|---|---|---|---|
| GET, pipeline 16 | 21.24M ops/s | 20.82M ops/s | 9.31M ops/s | 2.36M ops/s |
| SET, pipeline 16 | 17.79M ops/s | - | 8.70M ops/s | 1.94M ops/s |
| GET, no pipelining | 2.99M ops/s | 2.88M ops/s | 3.30M ops/s | 1.47M ops/s |
| SET, no pipelining | 3.22M ops/s | 3.07M ops/s | 3.37M ops/s | 1.26M ops/s |
Tail latency
- GET, pipeline 16, p99 latency: KiviDB 0.36 ms, Dragonfly 0.88 ms, Redis 2.29 ms
- SET, pipeline 16, p99 latency: KiviDB 0.45 ms, Dragonfly 1.02 ms, Redis 2.72 ms
Memory per key
- KiviDB: 143 bytes
- Dragonfly: 134 bytes
- Redis: 178 bytes
What these numbers do not cover
Persistence was off for every engine; one value size (100 bytes) and one machine size (32 cores) are published; Valkey and KeyDB were not measured on bare metal; single nodes without replication or TLS. The scripts are public at https://github.com/kividbio/kivi-benchmark, and an automated run on different hardware is in https://github.com/kividbio/kividb/blob/master/benchmarks/LATEST.md.