SurrealDB open-sources SurrealKV and SurrealMX engines, ships reproducible benchmark harness

SurrealDB has set out a three-part open-source commitment: publish the code, make every benchmark reproducible, and state plainly what was not measured. To back it up, the company has released two new storage engines, SurrealKV and SurrealMX, under the Apache 2.0 licence, together with crud-bench, the harness behind every performance figure it publishes.

The announcement came in a dev.to post by Tobie Morgan Hitchcock. It argues that a database stores your data, determines what persists after a crash and controls who can access what, so any claim about speed, correctness or safety should be checkable rather than taken on trust — especially for infrastructure that autonomous agents depend on while running unattended at scale.

What has been open-sourced

According to the post, SurrealKV 1.0 is a persistent storage engine implemented as a log-structured merge-tree, with snapshot-isolated transactions and optional versioned trees. It uses a concurrent adaptive radix tree as its memtable, plus a commit ring. SurrealDB says it will become the only persistent storage engine in SurrealDB 4.

SurrealMX 1.0 is an in-memory engine that layers multi-version concurrency control over a concurrent adaptive radix tree, offering lock-free reads and optional persistence.

Both engines carry Apache 2.0 licences, as does crud-bench. The company's SDKs, its tools and many of the libraries beneath SurrealDB are also open source under Apache 2.0 or MIT, with the full inventory published on its website. Each engine ships with a whitepaper explaining how its performance is achieved, what was measured and, in line with the stated commitment, what was left out.

Where the core database stands

SurrealDB itself remains source available rather than open source. It ships under the Business Source License 1.1, which the post says costs nothing for internal use and automatically becomes Apache 2.0 after four years. SurrealDB Cloud and Enterprise Edition are commercial products. The post draws a deliberate line between the two labels: open source is reserved for Apache 2.0 or MIT licensing with no restrictions, while source available covers code you can read and run under a licence that restricts certain uses.

Reproducible benchmarks

Every number SurrealDB publishes comes from crud-bench on a single machine: an AMD Ryzen Threadripper 9970X with 32 cores and 64 threads, 128 GB of RAM and a 4 TB NVMe SSD. The same harness and command line drive each comparator, so every comparison pair is measured under identical conditions. The company also discloses whether a given engine ran in memory or on disk.

The post includes the exact commands to rerun the work. SurrealKV is benchmarked with synchronous durability against RocksDB, SlateDB (on disk, not object storage), Fjall, ReDB, LMDB and libmdbx. SurrealMX is measured against LMDB and RocksDB, both running in memory. Changing a single flag in the harness re-runs a comparator, and SurrealDB encourages developers to reproduce results on their own hardware and to open issues or pull requests on crud-bench if they find fairer configurations or engines that should have been included. The reasoning is straightforward: benchmarks that outsiders cannot rerun carry little weight.

Built on published research

The post acknowledges the lineage of both engines. SurrealKV sits in the LSM-tree tradition running from Bigtable through LevelDB to RocksDB, and it is compared against members of that line. SurrealMX draws on the adaptive radix tree and on the multi-version designs behind Hekaton and Silo. The whitepapers cite these influences, and the company notes that several of the engines it benchmarks against are themselves open source.

What comes next

Launches scheduled for 12 to 16 October 2026 will cover the stack from the bottom up: the storage engine, the database layer, and collaborative primitives aimed at agent workloads operating at speed and scale. SurrealDB frames agents as changing what a database has to do — working in volume, sharing state and carrying context between sessions — and says a fuller statement of its infrastructure plans will follow in a separate post.

Why it matters

For developers evaluating databases, this moves the conversation from vendor claims toward evidence anyone can rerun. The licence split — open engines, harness and tooling on Apache 2.0, core database on BSL 1.1 — makes the trust boundary explicit, which matters when assessing long-term lock-in. Publishing the command, hardware and method alongside each number, including whether a comparison was in memory or on disk, sets a standard other vendors can reasonably be held to, since those configuration details are often omitted. And because SurrealKV is slated to become the sole persistent engine in SurrealDB 4, its documented behaviour will largely define the platform's storage story going forward.