---
type: "post"
id: "jepsen-analysis-of-particledb-1-1-4-warns-4c017d"
slug: "jepsen-analysis-of-particledb-1-1-4-warns-4c017d"
title: "Jepsen analysis of ParticleDB 1.1.4 warns LLMs erode software trust"
description: "Jepsen’s analysis of ParticleDB 1.1.4 through 1.1.18.11.rc-22 reports 96 issues, including lost committed writes under the default transaction mode and anomalies under a mode labeled serializable."
retrieval_nugget: "Jepsen’s analysis of ParticleDB 1.1.4 through 1.1.18.11.rc-22 reports 96 issues, including lost committed writes under the default transaction mode and anomalies under a mode labeled serializable. The report matters to AI builders because ParticleDB is positioned for AI workloads, and because Jepsen argues that LLM-assisted development can make traditional software-trust signals less reliable."
published_at: "2026-11-05"
updated_at: "2026-11-05"
record_date: "2026-11-05"
date_kind: "published_at"
topics: ["agent-security"]
entities: ["Jepsen","ParticleDB","Orkes","Kyle Kingsbury","Conductor"]
related_patterns: ["agent-security-moves-to-runtime-boundaries"]
novelty_label: "new_fact"
what_changed: "The Jepsen report adds concrete reliability findings: 96 issues across ParticleDB versions, including committed-write loss and serializable anomalies; the prior focused on Kingsbury’s warning that LLMs erode software trust."
prior_refs: []
source_url: "https://jepsen.io/analyses/particledb-1.1.4"
source_urls: ["https://jepsen.io/analyses/particledb-1.1.4"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:jepsen-analysis-of-particledb-1-1-4-warns-4c017d"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/jepsen-analysis-of-particledb-1-1-4-warns-4c017d/","markdown":"https://newruntime.com/posts/jepsen-analysis-of-particledb-1-1-4-warns-4c017d.md","json":"https://newruntime.com/posts/jepsen-analysis-of-particledb-1-1-4-warns-4c017d.json"}
markdown_url: "https://newruntime.com/posts/jepsen-analysis-of-particledb-1-1-4-warns-4c017d.md"
---

# Jepsen analysis of ParticleDB 1.1.4 warns LLMs erode software trust

## Retrieval answer

Jepsen’s analysis of ParticleDB 1.1.4 through 1.1.18.11.rc-22 reports 96 issues, including lost committed writes under the default transaction mode and anomalies under a mode labeled serializable. The report matters to AI builders because ParticleDB is positioned for AI workloads, and because Jepsen argues that LLM-assisted development can make traditional software-trust signals less reliable.

## Summary
Jepsen’s analysis of ParticleDB 1.1.4 through 1.1.18.11.rc-22 reports 96 issues, including lost committed writes under the default transaction mode and anomalies under a mode labeled serializable. The report matters to AI builders because ParticleDB is positioned for AI workloads, and because Jepsen argues that LLM-assisted development can make traditional software-trust signals less reliable.

## What happened
Kyle Kingsbury published the Jepsen analysis on October 31, 2026. ParticleDB is Orkes’ multi-model database, designed to consolidate transactional SQL, key-value storage, vector search, live analytics and user-defined functions called Particles. It was built to support Orkes’ Conductor workflow orchestration platform. Jepsen tested versions 1.1.4 through 1.1.18.11.rc-22 after Orkes engaged the firm ahead of a broader release.

The report catalogues 96 issues, ranging from crashes, deadlocks and data loss to partially invisible rows that could not be deleted, dropped tables reappearing in distorted forms, confused columns and strings becoming integers. Most issues required only one server and one client, without injected faults. By version 1.1.18.11.rc-22, Jepsen said all but ten of the identified issues appeared resolved. However, it continued to observe loss of committed writes during normal operation with the default `occ` transaction mode. It also found internal anomalies and Lost Update behavior with `serializable`, described as the strongest setting.

ParticleDB offered several transaction modes, and the report documents confusing or inconsistent descriptions of their guarantees. Its marketing claimed “Full ACID” and “159K TPS,” while nearby text referred to snapshot isolation, which is not full isolation. Jepsen said SQL isolation-level statements were ignored or rejected in some cases and silently ignored in others; the effective behavior depended on a startup `--txn-mode` setting. Documentation and operator guidance also disagreed about transaction and write-ahead-log sync behavior. The database offered `sync`, `groupsync` and `nosync` modes for WAL syncing.

The report covers single-node behavior only. ParticleDB also offered replicated clusters using Raft, but replication was available only in enterprise cloud deployments, and the builds Orkes supplied to Jepsen did not support it. Orkes subsequently released ParticleDB 1.1.19-rc.13 and told Jepsen it resolves all remaining issues discussed in the report. Jepsen’s research was funded by Orkes and conducted under its ethics policy.

Jepsen also uses the engagement to make a broader argument about LLM-assisted development. Its concern is not that a model caused these defects. Rather, it argues that generated code, documentation and surface polish can make conventional signals of a mature project easier to counterfeit. That makes independently testable behavior, clear specifications and reproducible evidence more important when evaluating software that presents broad guarantees.

## What the archive adds
This is a new fact on a broader software-trust thread, not a repeat of earlier coverage. [Cohere’s sandbox report on Qwen3.8-Flash-Next](https://newruntime.com/posts/cohere-s-sandbox-caught-qwen3-8-flash-next-e7776b/) documented model behavior under a security prompt; this Jepsen analysis instead tests a database’s transactional guarantees and raises a distinct concern about LLM-assisted software development.

## Trend signal
The report supports [agent security moves to runtime boundaries](https://newruntime.com/patterns/agent-security-moves-to-runtime-boundaries/) only indirectly: the core finding is database correctness, not an agent runtime boundary. It reinforces the practical need to evaluate systems by observed behavior rather than trust claims or polished surfaces, but it does not establish the pattern’s thesis directly.

## Why now
Orkes engaged Jepsen while preparing ParticleDB for a broader release, after a major refactoring added a custom wire protocol, more data types and additional OLAP queries. That timing makes the test a release-readiness examination, rather than a retrospective on a mature, static system.

## Why it matters
Teams building AI applications often want one store to serve several roles: transactional database, vector index, analytics engine and agent-state backend. Consolidation can simplify architecture, but it also concentrates risk. If committed state can disappear or isolation claims do not match behavior, failures can propagate across an application’s memory, workflow and retrieval paths.

This is especially consequential where agent actions write state that later steps treat as authoritative: task records, approvals, inventory changes, user preferences or the evidence behind a retrieved answer. A system can produce plausible outputs while its durable state is inconsistent. The relevant question is therefore not whether a database has vector search or accepts familiar protocols, but which transaction and durability guarantees apply to the exact workload, version and configuration being deployed.

The report also cautions against treating documentation polish, feature breadth or formal-sounding guarantees as substitutes for independent validation. Jepsen says ParticleDB’s home page claimed a TLA+ proof, but no specification was made available during the collaboration, leaving unclear what properties were proved or whether they described the shipped system. For builders, the concrete response is to test the exact version, transaction mode and durability configuration they intend to run, and to ask vendors to substantiate guarantees with reproducible artifacts.

## Limits and uncertainty
This is Jepsen’s account of specific versions and tests, not proof that every ParticleDB deployment has the same failures. The supplied builds did not support replication, so the report does not establish the behavior of ParticleDB’s Raft clusters or enterprise cloud deployments. Orkes claims version 1.1.19-rc.13 resolves the remaining reported issues, but the supplied material does not include an independent Jepsen retest of that release. Nor does the report establish that LLMs caused the database defects; its broader argument is about how LLMs may make trust signals easier to counterfeit.

## Skill to build now
Create a small acceptance suite for any datastore that will hold agent or workflow state. Include concurrent updates, a committed-write durability check, recovery after restart, and assertions for the isolation property the application expects. Run it against the exact release and startup transaction mode planned for production. Keep the results with deployment records, so an upgrade or configuration change is evaluated as a behavioral change rather than accepted from release notes alone.

## Reader takeaway
ParticleDB’s wide feature set and “Full ACID” positioning did not settle whether its transaction modes preserved the guarantees a team might rely on. Jepsen found concrete failures, while Orkes says a later release fixes them. Treat the report as a reason to validate the exact release and workload you plan to use—and as a reminder that credible claims need inspectable evidence.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Agent Security](https://newruntime.com/topics/agent-security.md): Explore the agent-security topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)