Jepsen Analysis: ParticleDB 1.1.4 Safety and Correctness

Executive Summary

ParticleDB 1.1.4, a multi-model database marketed for AI workloads, failed to meet its claims of "Full ACID" compliance during Jepsen's testing. The analysis uncovered 96 distinct issues, ranging from basic SQL syntax failures to severe data corruption, lost writes, and transactional anomalies that occurred even in single-node, fault-free environments. While subsequent release candidates addressed many of these bugs, the findings suggest fundamental deficiencies in the database's development and quality assurance processes.

The State of ParticleDB

ParticleDB was designed by Orkes to support their Conductor orchestration platform, aiming to unify SQL, key-value, and vector search workloads. Despite marketing claims of being "formally proven" and "orders of magnitude faster than Postgres," the version tested (1.1.4 through 1.1.18.11.rc-22) exhibited profound instability.

Key findings include:

• Data Corruption: The database frequently lost committed writes, returned incorrect results for simple queries, and allowed "zombie" rows that could not be deleted or updated.
• Transactional Failures: Even under strong isolation settings, the system exhibited violations of Snapshot Isolation and Serializability, including Lost Updates and Stale Reads.
• SQL Incompatibilities: Basic operations like WHERE clauses, JOINs, and aggregate functions frequently returned incorrect data or threw internal errors.
• Systemic Instability: The database suffered from spontaneous deadlocks and crashes, even without fault injection.

Insights into Quality Assurance

Jepsen’s report highlights a concerning gap between the database's marketing and its actual implementation. The "formally proven" claims were unsupported by any available TLA+ specifications, and the documentation frequently contained fabrications, such as incorrect default usernames and non-existent SDK dependencies.

As one commenter on Hacker News noted, the sheer volume of "low-hanging fruit" bugs—such as WHERE 4 = 5 returning rows—suggests a lack of basic automated testing. Commenters speculated that the project may have been "vibecoded" or heavily reliant on LLMs without sufficient human oversight or verification. As noted by one user:

"The sheer hubris of their claims combined with the specific reality of their broken garbage software... is literally breathtaking."

Recommendations

Jepsen recommends that users halt the deployment of ParticleDB and migrate to more reliable alternatives. For developers, the report serves as a stark reminder of the risks associated with rapid, LLM-assisted development cycles that bypass rigorous quality assurance and formal verification.

Orkes has since updated their documentation to label the modes tested as "experimental" and claims that version 1.1.19-rc.13 resolves the reported issues. However, Jepsen maintains that the burden of proof lies with the vendor to demonstrate correctness through transparent, reproducible testing and accurate documentation.