独立测试机构 Jepsen 对 Orkes 公司开发的 AI 原生数据库 ParticleDB（版本 1.1.4）进行了深度安全性与一致性测试 1。报告显示，该数据库存在数百个关键 Bug，其中 Jepsen 明确发现 96 个问题，且绝大多数发生在健康的单节点无故障实例中 1。核心缺陷包括 WHERE 子句被忽略或错误匹配、NOT NULL 及 UNIQUE 约束失效、事务死锁以及无法删除或更新的“僵尸行” 1。尽管厂商宣称其具备 ACID 特性且经过形式化验证，但实际行为不符合预期，常规操作即可导致数据静默损坏或丢失 1。
针对上述问题，Jepsen 强烈建议用户停止部署 ParticleDB，进行数据恢复并迁移至其他数据库 1。报告指出，LLM 辅助开发可能导致代码库充满正确性错误，且开发者往往难以解释 Bug 的根本原因 1。此外，虽然文档声称有 TLA+ 形式化证明，但实际未提供规范；厂商宣称支持全 ACID 和快照隔离，但实际表现亦未达到标准 1。在测试后，厂商 Orkes 更新了文档，将相关事务模式标记为“实验性”，并声称后续版本已修复大部分问题 1。
Independent testing firm Jepsen has released a report detailing severe consistency and safety failures in ParticleDB version 1.1.4, an AI-native database developed by Orkes 1. The audit identified 96 distinct issues, the majority of which occurred within healthy single-node instances without external faults or network partitions 1. These findings contradict the vendor's claims that the software supports full ACID compliance, snapshot isolation, and formal verification, as the actual behavior failed to meet these documented standards 1.
The core defects discovered include significant logical errors such as WHERE clauses being ignored or matching incorrectly, alongside ineffective enforcement of NOT NULL and UNIQUE constraints 1. Jepsen also reported transaction deadlocks and the emergence of "zombie rows," records that cannot be deleted or updated through standard operations 1. Despite documentation asserting the existence of TLA+ formal proofs, no corresponding specifications were provided to support these claims 1. The researchers noted that LLM-assisted development appeared to have introduced numerous correctness errors into the codebase, with developers often unable to explain the root causes of the bugs 1.
In response to the test results, Orkes updated its documentation to label certain transaction patterns as "experimental" and claimed that subsequent versions had resolved most of the identified problems 1. However, due to the risk of silent data corruption or loss during routine operations, Jepsen strongly advised users to cease deploying ParticleDB immediately 1. The firm recommended that existing users prioritize data recovery and migrate to alternative database systems 1.
评论
还没有评论，欢迎留下第一条。