Measured 27 Aug 2026 · one node · one tenant

Isolation is not expensive.

These numbers exist to show that per-tenant isolation does not cost an order of magnitude — not to rank DBX against a tuned cluster of something else. Measured on one node, single tenant. Measure on your own hardware before you rely on any of this.

Lab log

Field Value
Date 27 Aug 2026
OS Windows 10 build 26200
CPU 18 logical processors
RAM 15.6 GiB
Topology Single-node, local. One tenant.
Purpose Windows-only certification. This page does not claim Linux numbers.

SET

186,147 ops/s

RESP, 64 connections, pipeline 64, WAL everysec

GET

284,785 ops/s

Same run, 128k operations

Vector ingest

7,233 vec/s

100k × 128-dim, batches of 1,000, 8-way HNSW

ANN search

2.304 ms p50

p95 3.132 ms · p99 3.730 ms · 100k, 128-dim, efSearch=80, 50 queries

Recall@10

0.920 mean

p05 0.800 · SQ8 HNSW vs float32 brute force

Same numbers, as a table
Operation Number Measured with
SET 186,147 ops/s RESP, 64 conns, pipeline 64, WAL everysec
GET 284,785 ops/s Same run, 128k ops
Vector ingest 7,233 vec/s 100k × 128-dim, batches 1000, 8-way HNSW
ANN search p50 2.304 / p95 3.132 / p99 3.730 ms 100k, 128-dim, efSearch=80, 50 queries
Recall@10 mean 0.920 / p05 0.800 SQ8 HNSW vs float32 brute force

How to read these numbers

  1. 01

    Hardware is attached

    Date, OS build, logical processors, RAM, client shape, connection count, and whether the run was pipelined. A number without those is marketing.

  2. 02

    Single tenant, said out loud

    This run is one node, one tenant, local. Density — how many active tenants fit on one node, and what an idle tenant costs — is the interesting product metric. Do not derive a cluster ranking from these rows.

  3. 03

    Recall travels with latency

    SQ8 is a quality trade. Publishing ANN milliseconds without Recall@10 would make the page untrustworthy. Mean 0.920, p05 0.800 against float32 brute force, on this corpus.

  4. 04

    Windows, not Linux

    Certification was Windows 10 build 26200 as documented. This page does not claim Linux numbers.

We do not quote a competitor’s number we did not measure ourselves under the same conditions. We do not derive throughput from a latency percentile. If a claim requires naming a competitor to make sense, it is a weak claim.

Methodology notes in the repo →