What the NVMe benchmark tells us
Read the measured results, their conditions, and the questions they leave open.
The September 6 report compares a historical Lavik build with tuned Redis and Valkey using a working set that fits in memory. Its Lavik commit is 29dc8e6b87c40196dc397759690252944f1196f0. The report is preserved with the 0.1.0 documentation as historical evidence.
In the published September 6 experiment, Lavik reached 828,502 GET requests/s, compared with 964,267 for tuned in-memory Redis: 85.9% of Redis throughput.
In that experiment, Lavik reached 984,452 SET requests/s, compared with 874,879 for tuned in-memory Redis, a 12.5% higher measured peak.
The comparison used an AMD EPYC 9V74 server, 125 GiB RAM, six raw NVMe drives, 10 million 1 KiB values, and 30-second measurement windows. Redis persistence was disabled and Lavik defragmentation was paused. These measurements do not establish equivalent crash durability, production cost, or an application SLA.
Look beyond one number
At their selected peaks, reported GET p99 latency was 4.543 ms for Lavik and 4.671 ms for Redis. Lavik’s GET p99.99 was higher. Each point was measured once; the report recommends repeated, interleaved trials. Similar p99 in this experiment does not establish an SLA for another workload.
The same report also contains a separate one-billion-key storage-tier experiment. It uses a different dataset and measurement window. Keep those results separate from the in-memory comparison.