Jepsen Crash-Tested ParticleDB, but Only on One Machine

Jepsen has published an analysis of ParticleDB version 1.1.4, a database built to store data for AI applications. The report, dated 31 October 2026, covers only how the database behaves on a single computer. The builds that Orkes provided did not include replication, the feature that copies data across several machines.
Orkes hired Jepsen to test refactored versions of ParticleDB. Version 1.1.4 is the first release in that effort, and this report is the first result of that work. Its scope is set by the version that was tested, not by the full product.
The test setup could simulate crashes by shutting down the ParticleDB process without warning. It also used a tool called LazyFS to simulate the loss of writes that had not yet been saved to the disk.
Together, these tests check durability. That means whether data the database told a client was saved still exists after a sudden stop.
A database can confirm a write before that write is safely on the disk. If a crash hits at that moment, the data can disappear. This gap is where a single-node database's promises either hold up or fail.
Replication raises harder questions. These include how writes travel between computers, what a user sees when the network splits, and how a change of leader affects data that was already confirmed. None of these fall within this report, because the builds did not include replication.
Readers should treat the findings as a statement about one ParticleDB process on one disk. Any setup that relies on data copied across machines will need separate evidence.
Orkes publishes official ParticleDB content on orkes.io.
The broader context here is that developers increasingly use databases to hold the memory of AI agents, the information they look up, and records of their workflows. If a write is lost, an agent can change its behavior without anyone noticing. Tests that target crashes and unsaved writes catch a kind of failure that ordinary performance benchmarks often miss.
In my view, the most useful part of this work is the test setup itself. A tool that kills the process and throws away unsaved data can be reused on other systems. The open question for anyone evaluating ParticleDB is how the same discipline will be applied once replicated builds are available for testing.


