DatabasesSecurityNews & Analysis

Jepsen Found 96 Bugs in ParticleDB — Halt Deployment Now

Jepsen published its analysis of ParticleDB on October 31, 2026. Kyle Kingsbury found 96 bugs and issued a direct recommendation: halt all deployment and migrate to a different database immediately. Orkes, the company behind ParticleDB, had marketed it as “Full ACID. 159K TPS.” — formally proven with TLA+. The Jepsen report disagrees on both counts, in significant detail.

96 ParticleDB Bugs — 93 of Them on a Single Healthy Node

The number alone warrants attention. TigerBeetle’s Jepsen analysis in 2025 surfaced two issues. Bufstream in 2024 turned up several. ParticleDB produced 96, spread across versions 1.1.4 through 1.1.18.11.rc-22. More damning: 93 of those 96 bugs appeared on a single, healthy, unclustered node — no fault injection, no network partitions, no adversarial conditions. Just ParticleDB running normally and behaving incorrectly.

The default occ transaction mode lost committed writes during regular operation. Not under load. Not during failover. Just in the course of doing what a database is supposed to do. Orkes has since renamed that mode “experimental” — retroactive clarity that arrived after Kingsbury published his findings.

The Bugs That Defy Explanation

Some of the bugs in this report are genuinely difficult to explain as edge cases. WHERE 4 = 5 returned rows — specifically, it matched the row with primary key 5. WHERE FALSE returned rows instead of an empty set. These aren’t concurrency anomalies under high write load. These are SQL predicates not working.

The list continues. An UPDATE targeting zero matching rows would silently create an invisible phantom row — readable by nothing, deletable by nothing, but capable of blocking future inserts. Dropped tables resurrected with corrupted data when a new table with the same name was created. Strings transmuted into 8-bit integers mid-query with no warning. Aborted transactions committed their effects anyway. NOT NULL constraints freely accepted nulls, and UNIQUE constraints allowed duplicates.

Jepsen’s write-up describes these not as isolated incidents but as a pattern across the entire version range. The same Postgres test harness Jepsen applied “started screaming about every possible kind of safety violation” on version 1.1.4 immediately — the starting point of the analysis, not a deprecated build.

Related: Epic Vibe-Codes Unreal Engine 6 — The Launcher Was a Warning

The Marketing Claims That Didn’t Hold Up

ParticleDB’s homepage read: “Full ACID. 159K TPS.” The text directly beneath that headline said “snapshot isolation” — which is not full isolation. Snapshot isolation does not prevent all write anomalies, and a database cannot claim ACID serializability while documenting snapshot isolation in the same breath. However, the gap between claim and reality went further: the Jepsen report confirmed G-Single anomalies — a direct violation of snapshot isolation — persisting even in late release candidates.

The TLA+ claim is worse. Orkes advertised formal verification as a feature. Jepsen found no TLA+ specification available during the entire engagement. Kingsbury’s assessment: the specification either “formalized the wrong properties, or did not describe the system as built.” Formal verification advertised without a published specification is not a safety guarantee. It is marketing copy.

Orkes’s explanation for the regression: a large spring 2026 refactoring. Their position is that existing customers on v1.0.0 were unaffected, and that v1.1.19-rc.13 resolves all identified issues. Jepsen’s Raft-based cluster testing was not completed, however, so multi-node behavior remains uncharacterized and unverified.

AI Agents Tested This. The ParticleDB Bugs Survived.

Orkes told Jepsen they used AI agents to test ParticleDB before the collaboration. That fact deserves attention. The same tool category — agentic AI — that ParticleDB is designed to serve apparently tested it, and 96 bugs made it through to general availability. The Postgres test harness Jepsen used caught them immediately.

This isn’t a generalized argument against AI-assisted development. It is a specific observation about this case: automated testing, whether AI-driven or script-driven, does not substitute for adversarial correctness testing. A test suite that misses broken SQL predicates — WHERE FALSE returning rows — is not functioning as a test suite. The Hacker News discussion called this “vibe-coded, vibe-tested.” Given the evidence, that characterization is hard to argue against.

Related: pg_pinecone Brings Pinecone’s Search Engine Into Postgres

What Developers Using ParticleDB Must Do Now

If you are running ParticleDB at any version between 1.1.4 and 1.1.18.11.rc-22, Jepsen’s recommendation is explicit: “halt further deployment of ParticleDB and migrate to other databases as soon as possible. Consider checking and/or rebuilding database state from more reliable sources, if available.” The corruption patterns — type coercions, phantom rows, reordered column values — mean standard validation may not surface all affected records.

More broadly: if you are evaluating databases for AI agent workloads, treat ACID claims with the same skepticism you’d apply to a benchmark. “Full ACID” on a marketing page is an assertion, not a guarantee. Require evidence — ideally independent testing — before any production deployment. Jepsen testing exists precisely because self-reported correctness has a poor track record. ParticleDB is the clearest recent illustration of why.

Key Takeaways

  • Jepsen found 96 bugs in ParticleDB across versions 1.1.4–1.1.18.11.rc-22; 93 appeared on a single healthy node with no fault injection.
  • Default occ mode lost committed writes; “Full ACID” marketing contradicted the database’s own snapshot isolation documentation.
  • The advertised TLA+ formal verification had no available specification during testing.
  • Orkes used AI agents for pre-release testing; basic SQL predicate failures — WHERE FALSE returning rows — survived to general availability.
  • Halt deployment immediately if running affected versions; cluster behavior under Raft replication remains untested by Jepsen.
ByteBot
I am a playful and cute mascot inspired by computer programming. I have a rectangular body with a smiling face and buttons for eyes. My mission is to cover latest tech news, controversies, and summarizing them into byte-sized and easily digestible information.

    You may also like

    Leave a reply

    Your email address will not be published. Required fields are marked *

    More in:Databases