Jepsen Finds 96 Bugs in ParticleDB 1.1.4
Original: Jepsen: ParticleDB 1.1.4
Why This Matters
ACID claims are foundational for AI agent workloads; silent data loss is a dealbreaker.
Kyle Kingsbury's Jepsen analysis of ParticleDB 1.1.4–1.1.18.11.rc-22 found 96 issues, including data loss, deadlocks, and lost committed writes, undermining the database's 'Full ACID' and 'formally proven' claims.
Jepsen tested ParticleDB, an AI-workload-focused multi-model database built by Orkes in Rust, across versions 1.1.4 through 1.1.18.11.rc-22. The report, funded by Orkes itself, catalogues 96 distinct issues. Many are mundane—crashes, deadlocks, data loss—but several are genuinely strange: rows that were partially invisible yet couldn't be deleted, dropped tables re-emerging in distorted forms, strings silently converting to integers, and a WHERE 4 = 5 anomaly. Critically, most bugs required only a single server and a single client with no injected faults.
By version 1.1.18.11.rc-22, Orkes had fixed 86 of the 96 issues. The remaining ten are serious: ParticleDB still loses committed writes during normal operation under the default OCC transaction mode, and allows internal anomalies and Lost Update even under the serializable isolation setting—its strongest option. The database's homepage had advertised a TLA+ formal proof of correctness, but Jepsen reports no specification was ever made available during the engagement. Orkes has since released 1.1.19-rc.13 and claims it resolves all remaining issues in the report.