Jepsen, the independent lab that stress-tests databases, published its analysis of ParticleDB and found 96 bugs, from lost committed writes to queries that return the wrong rows. A useful lesson for anyone choosing a database: a "Full ACID" badge on a product page is a promise, not proof.
What Jepsen found in ParticleDB
The Jepsen report was written by Kyle Kingsbury and paid for by Orkes, the company behind the database. ParticleDB is a multi-model database aimed at AI workloads: it combines transactional SQL, a key-value store, vector search, live analytics and user-defined functions called Particles. It speaks the PostgreSQL and Redis protocols and has its own gRPC and HTTP2 APIs. Orkes built it for Conductor, its workflow orchestration platform, which used to keep its data in Redis and Postgres.
Jepsen tested versions 1.1.4 through 1.1.18.11.rc-22 on single nodes (replication was not tested, since the builds provided lacked it). Most surprisingly, the majority of problems showed up on one healthy node, with one client and no injected faults. A few examples from the report:
- the condition
WHERE 4 = 5returned the row whose primary key was 5, - dropped tables could come back in an altered form, with 18.83 turning into 19,
- text values were sometimes returned as integers,
- inserting columns in a different order created rows that ordinary queries could not read,
- up to version 1.1.18.8 there was no way to set a user name for the Postgres protocol, so in practice only disabling authentication worked.
Transactions and durability
The default transaction mode, occ, which Orkes describes as snapshot isolation, lost acknowledged writes during normal operation in the last version tested. The strongest mode, serializable, still allowed anomalies, including lost updates, after a process was killed. Simulated disk faults (reads and writes landing in the wrong place) wiped out all data in one case. The product page promised full ACID and a design formally proven in the TLA+ specification language, but Jepsen never received a TLA+ specification.
The role of AI and the vendor's response
A separate section of the report is about working with language models. According to Jepsen, LLMs let a handful of Orkes engineers build such a large database, but quality paid the price: the AI tools repeatedly claimed a bug was fixed when it was not, and the team usually could not explain what caused a problem. New builds came without changelogs. Orkes later labelled every tested version as an unsupported pre-release build, even though its Docker documentation had described them as ready for production.
The company has released 1.1.19-rc.13 and says it fixes every issue in the report. Jepsen has not tested it. Its advice is clear: stop deploying ParticleDB, move to another database as soon as possible, and check your data, rebuilding it from more reliable sources where you can.
The testing method is worth a look too. A workload called compare generated random table schemas and SQL queries, ran them against both ParticleDB and PostgreSQL, and compared the results. About two thousand lines of these tests were enough to uncover a large share of the bugs.
What this means if you are learning to code
First, make sure you really understand what ACID means and how transaction isolation levels differ. Snapshot isolation and serializable are not decorative words from the docs. They are concrete guarantees: whether two concurrent writes can overwrite each other, or whether a read can see data from a moment ago. You can practise transactions in our SQL course, and read more about the mature database ParticleDB was compared with in our article on PostgreSQL.
Second, when you pick a database for a project, proven tools beat the newest speed claims. Postgres and Redis have years of testing and thousands of deployments behind them. A nice website promising full ACID proves nothing yet.
Third, the report is a warning for anyone who writes code with AI. A model can confidently announce that a bug is fixed. Your job is to check it with a test that failed before. Another good habit is comparing your new code's output with a solution you trust, the way Jepsen compared against Postgres. If you want to work with AI deliberately, try our AI-assisted coding course and the article on vibe coding.
You do not need TLA+ or distributed systems testing yet. SQL, transactions and a testing habit are enough to start.
Also today
- Claude Code 2.1.316. Anthropic's coding agent adds a
CLAUDE_CODE_SUBAGENT_EFFORTvariable that sets the effort level of subagents, and deprecates fast mode for Opus 4.8: from November 30, 2026 such requests will return an error, and Opus 5.5 is the replacement. More about the models in our article on Claude. Source - Prospero, a new harness for coding with AI. Chris Wellons introduced a C++ program with a Dear ImGui interface that lets a model read files and run commands, using, by the author's estimate, two orders of magnitude fewer resources than popular tools. You plug in your own model (llama.cpp, OpenRouter, the Anthropic or OpenAI API), and the Community Edition is free on GitHub. Source
- New OpenAI SDKs for Python and Node.js. openai-python 3.42.0 and openai-node 7.46.0 add settings for agent context compaction, agent session review and agent tools, plus experimental time awareness. openai-node 7.47.0 adds a typed WebSocket warmup for the Responses API. Source
- Electron 45.2.1 and 44.10.1. Release 45.2.1 updates Chromium to 156.0.8078.75 and backports Chromium fixes, while 44.10.1 speeds up
net.fetch()for requests without a body and for repeatedfile:URLs. Source - Firezone introduces sanstun. A Rust library implementing the WireGuard protocol in a sans-IO style, licensed under Apache 2.0, and it replaces the company's fork of Cloudflare's boringtun. The first version, 0.1.0, reached crates.io a day before the announcement. Source
Frequently asked questions
What is Jepsen?
It is Kyle Kingsbury's independent project that tests databases and distributed systems for broken guarantees. Vendors often commission these analyses themselves, as Orkes did here. Jepsen reports are public and very detailed.
Is ParticleDB safe to use?
Jepsen recommends halting deployments and migrating to another database. Orkes says version 1.1.19-rc.13 fixes the problems described, but no one independent has confirmed it, so with important data, wait for further testing.
What does ACID mean in databases?
It stands for four transaction guarantees: atomicity (all or nothing), consistency, isolation (concurrent transactions do not interfere with each other) and durability (committed data does not disappear). In Jepsen's tests ParticleDB mostly broke isolation and durability.
Sources
- Jepsen: ParticleDB 1.1.4, Jepsen, October 31, 2026
- Claude Code v2.1.316, Anthropic (GitHub), October 31, 2026
- Prospero: a new coding harness for a more civilized age, nullprogram.com, October 31, 2026
- openai-python v3.42.0, OpenAI (GitHub), October 31, 2026
- openai-node v7.46.0, OpenAI (GitHub), October 31, 2026
- Electron v45.2.1, Electron (GitHub), October 31, 2026
- Introducing sanstun, Firezone, October 31, 2026
This article was prepared with the help of AI from primary sources and reviewed by the CodeWorlds editorial team.