Changelog¶

Changes to the protocol, newest first. Engine versions are not listed here; they are on every result row.

2026-11 (in development)¶

The benchmark was renamed from DBBench to MMDBBench on 2026-11-07 and refocused on multi-model databases. The protocol version is unchanged.

The protocol version, in development until its first release. The method is described on the Methodology page.

Method changes made while it is in development, before any result is published with them:

Served ArcadeDB leaves the initial heap (-Xms) unset, like embedded ArcadeDB. Neo4j and Elasticsearch keep a pinned initial heap, as their vendors recommend.
Neo4j's heap and page cache come from its own neo4j-admin server memory-recommendation for each memory cap. Elasticsearch's heap is the smaller of the tier heap, 50% of the cap and 26 GiB.
Container io figures count a stacked block device once. Earlier rows summed every device line of the cgroup io.stat, which counted the benchmark machine's NVMe disk and the device-mapper device on top of it twice. The harness now stamps io_stat_dedup: 1 into each row, and the frozen table halves the io figures of rows without the stamp (see Result rows).
Milvus runs as 3.0.3 on the vendor's Docker Compose standalone (three containers sharing the cell's memory cap), ingests by bulk import, waits for every segment to be flushed and the index to be finished before timed searches, and records the storage mode, the output fields and the filter. Storage V3 is a sensitivity setting.

A development release, superseded by 2026-11 before any comparison was published from it.