ParticleDB’s marketing page touts “Full ACID. 159K TPS.” The Jepsen report disputes the claim.
PARTICLEDB'S FULL ACID CLAIMS PROVEN PREMATURE BY JEPSEN TESTS
*Jepsen’s analysis of ParticleDB 1.1.4‑1.1.18 reveals dozens of stability and consistency failures. Orkes’ promised fix arrives in 1.1.19‑rc.13, but the damage to credibility is clear.*
ParticleDB entered the AI‑focused database market with bold promises: “Full ACID” performance at 159,000 transactions per second. Built in Rust by Orkes to power its Conductor workflow platform, the engine claimed formal verification and multi‑model capabilities. Jepsen’s October 31, 2026 analysis shattered that narrative. Testing versions from 1.1.4 through 1.1.18.11.rc-22, the team uncovered 96 defects, many of which appeared without any fault injection. The findings expose a gap between marketing hype and operational reality, especially for enterprises eyeing ParticleDB as a single‑node or Raft‑replicated solution.
What did Jepsen find in ParticleDB’s early releases?
Jepsen examined ParticleDB versions 1.1.4 through 1.1.18.11.rc-22. The team logged 96 distinct issues, ranging from crashes and deadlocks to bizarre data anomalies such as partially‑invisible rows, WHERE 4 = 5 queries, re‑emerging dropped tables, column confusion, and type‑conversion errors. Most bugs manifested on a single server with a single client and without injected faults. By version 1.1.18.11.rc-22, only ten of those problems remained unresolved. The report also noted that even after most bugs were fixed, the database continued to lose committed writes under the default OCC transaction mode and exhibited Lost Update anomalies under the strongest serializable setting.
Does ParticleDB actually deliver Full ACID guarantees?
ParticleDB’s homepage boasted “Full ACID. 159K TPS.” The same page qualified the claim with “snapshot isolation,” a level that falls short of true isolation. Jepsen observed that the engine ignored or rejected standard SET TRANSACTION ISOLATION LEVEL statements, sometimes forcing READ COMMITTED, sometimes silently discarding higher‑level requests. Isolation was instead dictated by the –txn-mode flag, which offered four opaque modes—fast, occ, zero‑copy‑occ, and table‑2pl—without clear guarantees. The documentation warned that the flag “changes correctness and throughput characteristics more than SQL text does,” yet failed to define the exact consistency model. Consequently, the advertised Full ACID label proved premature.
How does ParticleDB’s durability model compare to traditional databases?
ParticleDB provides three WAL sync tactics: fsync (sync on each commit), groupsync (batched fsync), and nosync (no fsync). The product’s own help text claimed groupsync “COMMIT waits for flush,” matching the durability of sync, while the transaction engine documentation described sync as offering “stronger durability” than groupsync. This contradictory messaging mirrors the broader inconsistency in the product’s safety claims. Traditional databases such as Postgres expose clear durability options, whereas ParticleDB’s mixed messages leave operators guessing about data loss risk, especially given the observed loss of committed writes in normal operation.
What is the path forward for Orkes and ParticleDB?
Orkes responded to the Jepsen findings by releasing version 1.1.19‑rc.13, asserting that it resolves all remaining issues highlighted in the report. The research was funded by Orkes, Inc., and conducted under the Jepsen ethics policy. While the new release may close the known bugs, the earlier failures raise questions about the robustness of ParticleDB’s testing regime and the reliability of its marketing claims. Future adopters will need to verify the fixes independently, especially before deploying the Raft‑based replicated cluster, which was not covered in the single‑node tests.
Frequently asked questions
How many issues did Jepsen find in ParticleDB versions 1.1.4 to 1.1.18?
Jepsen logged 96 distinct issues across those versions, ranging from crashes to data anomalies.
Did ParticleDB lose committed writes during testing?
Yes, the database continued to lose committed writes under the default OCC transaction mode even after most bugs were fixed.
What version did Orkes release to address the reported problems?
Orkes released version 1.1.19‑rc.13, claiming it resolves all remaining issues discussed in the Jepsen report.
Orkes’ rapid release of 1.1.19‑rc.13 shows a willingness to patch, but the record of lost writes, invisible rows, and contradictory durability documentation erodes trust. Stakeholders must demand transparent specifications and independent validation before relying on ParticleDB for mission‑critical AI workloads. The Jepsen report serves as a cautionary benchmark for any database that trades bold claims for unproven engineering.
Sources: https://jepsen.io/analyses/particledb-1.1.4