← Back to issue10 / 10 · Week of Nov 2, 2026

Jepsen finds 96 issues in ParticleDB, a database built with LLM help, and recommends halting its deployment

Jepsen's Oct 31 report tested ParticleDB 1.1.4 through 1.1.18.11.rc-22, a Rust database from Orkes whose home page advertised "Full ACID". It lists 96 issues, and most needed only one server, one client and no injected faults: lost committed writes, rows returned by `WHERE 4 = 5`, strings turning into integers. Roughly two thousand lines of generative tests, which ran random SQL against ParticleDB and Postgres and compared the answers, found many of them. By rc-22 all but ten appeared resolved, but the default `occ` mode still lost committed writes and the strongest `serializable` setting still allowed Lost Update. Orkes says 1.1.19-rc.13 fixes the rest; Jepsen has not tested that version. Orkes funded the work. Why it matters: Jepsen says a handful of Orkes engineers built ParticleDB with LLM help, and that the LLM tooling repeatedly claimed bugs were fixed when they were not. The website looked like a mature project, yet the docs seemed to invent class names, gave wrong default usernames and even made up an email address for the report's author. Jepsen's point is that the signals people used to read as signs of a mature system are now easier to counterfeit, so the question becomes how maintainers can show their software works.

Try this: Pick one feature your team or a coding agent shipped this month and find the test that failed before it existed and passes now. If there is none, write it before the next release, as Jepsen's section 4.3 recommends.

Source
Jepsen: ParticleDB 1.1.4
View source →

Get the field brief every week.

Important AI developments, useful explanations, and practical resources in one weekly read. Context to understand what matters, with links to the original sources and deeper reading.

Subscribe free →
Free weekly·No spam·Unsubscribe anytime