TL;DR
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Jepsen tested ParticleDB versions 1.1.4 through 1.1.18.11.rc-22 and reported 96 issues, including data loss and transaction anomalies. Most reported problems appeared resolved by the last version tested, but Jepsen still found committed-write loss in the default transaction mode and anomalies under serializable mode. Orkes says ParticleDB 1.1.19-rc.13 fixes all remaining issues discussed in the report; Jepsen’s supplied findings do not include results validating that release.
Jepsen reported 96 issues in ParticleDB releases it tested, including lost committed writes and transaction anomalies in version 1.1.18.11.rc-22, putting the database’s advertised ACID guarantees under scrutiny. Orkes, ParticleDB’s maker, says a later release, 1.1.19-rc.13, resolves all remaining issues discussed in the report, but the findings provided do not show whether Jepsen has independently tested that build.
The Jepsen analysis covers ParticleDB versions 1.1.4 through 1.1.18.11.rc-22. It lists failures including crashes, deadlocks and data loss, as well as less conventional behavior: rows that became partly invisible and could not be deleted, dropped tables returning in altered forms, confused columns and strings changing into integers. Jepsen said most issues required only one server and one client, with no injected faults.
By version 1.1.18.11.rc-22, Jepsen said all but ten issues appeared resolved. But it continued to observe committed writes being lost during normal operation with the default occ transaction mode. With the strongest setting, described in the report as serializable, the database still showed internal anomalies and a Lost Update. Those findings concern the single-node behavior Jepsen could test; the builds supplied by Orkes did not support the product’s replicated-cluster feature.
Jepsen said its research was funded by Orkes, Inc. and conducted under the organization’s ethics policy. Orkes’ statement that version 1.1.19-rc.13 resolves the remaining reported problems is a company claim; the report material does not document independent test results for that release.
Transaction Safety Under Test
The findings matter because transaction correctness is central to a database’s role as an authoritative store. Losing a write after a client has committed it can leave applications operating on missing or inconsistent state. Anomalies under the strongest advertised transaction setting may also affect applications that rely on serializable execution to prevent conflicting updates.
ParticleDB is positioned as a multi-model database for AI workloads, combining transactional SQL with key-value storage, vector search, analytics and user-defined functions. A single system may reduce the work involved in coordinating separate databases and services, but only if its guarantees match the needs of the applications using it. Jepsen’s results do not establish that every deployment or workload will encounter these failures; they do show why users should examine the tested version, transaction mode and durability configuration rather than relying on a broad “Full ACID” description.
The distinction between an announced fix and a verified fix also matters to teams considering deployment. Orkes says the later release addresses the outstanding issues, but readers evaluating the repair need details about the affected cases and independent validation, particularly for committed-write durability and serializable behavior.
Popular right now
Dyson Car+Boat Cordless Handheld Vacuum
De'Longhi 15-Bar Pump Espresso Machine with Milk Frother & Temperature Control
Breville Bambino Espresso Machine, Brushed Stainless Steel (BES450BSS)
As an affiliate, we earn on qualifying purchases.
A Database With Multiple Modes
ParticleDB was created by Orkes, initially to support its workflow orchestration platform, Conductor. The company developed it as an alternative to coordinating data across systems such as Redis and Postgres. The product supports SQL, key-value access, vector search, live analytics and connections through interfaces including Postgres and Redis protocols, according to the report.
Jepsen began testing during a major refactoring ahead of a broader release, starting with version 1.1.4. The report says transaction isolation depended on a startup setting, --txn-mode, rather than being controlled through standard SQL isolation-level statements. The documented modes included occ, the default; zero-copy-occ; fast; and table-2pl. Jepsen said the documentation did not clearly specify the guarantees offered by each mode, complicating efforts to determine what application developers could safely expect.
The report also describes conflicting documentation about write-ahead-log durability. ParticleDB offered sync, groupsync and nosync options. Jepsen noted that some documentation described groupsync as providing the same durability as sync, while another description said sync offered stronger durability. This discrepancy is separate from the reported transaction anomalies and leaves users needing precise, version-specific guidance on what a successful commit guarantees.
Although ParticleDB also offered a Raft-replicated cluster in enterprise cloud deployments, the builds supplied for Jepsen’s work did not support replication. The analysis therefore addresses single-node behavior and cannot establish how the replicated configuration performs.
““Full ACID” claims were “somewhat premature.””
— Jepsen report
The New Release Still Needs Verification
The report does not provide Jepsen test results for version 1.1.19-rc.13, so it remains unclear from the supplied material whether the release fixes the committed-write loss and serializable-mode anomalies under the same test conditions. Orkes’ assertion is not a substitute for published validation of those cases.
The findings also leave open how the product behaves in replicated deployments, because the tested builds did not support replication. The report’s listed WAL documentation conflicts further leave the practical durability guarantees of different sync settings unclear. The supplied findings do not say whether Orkes has clarified those descriptions or whether every issue applies to every version, workload and configuration.
Track Fixes and Replication Testing
The next point for users and evaluators is whether Jepsen or another independent tester publishes results for 1.1.19-rc.13, especially for the failures involving committed writes and serializable transactions. Until then, readers should distinguish Orkes’ stated fixes from independently reported verification and check release-specific guidance before relying on the database for durable transactional workloads.
Further published testing would also be needed to assess replicated-cluster behavior, which was outside the scope of the supplied Jepsen results. Clearer documentation of transaction-mode guarantees and WAL sync semantics would help operators understand which protections each configuration provides. No date for additional testing or a follow-up report is given in the material.
Key Questions
What did Jepsen find in ParticleDB?
Jepsen reported 96 issues across versions 1.1.4 through 1.1.18.11.rc-22, including crashes, deadlocks, data loss and unusual table or data behavior. In the latest tested version, it still found committed-write loss in default occ mode and anomalies under serializable mode.
Did Jepsen test ParticleDB 1.1.19-rc.13?
The supplied report material does not include test results for 1.1.19-rc.13. Orkes says that release resolves all remaining issues discussed in the analysis, but independent confirmation is not provided here.
Was ParticleDB’s replicated cluster tested?
No. Jepsen said the builds supplied by Orkes did not support replication, so its findings cover single-node behavior, not the Raft-based replicated cluster offered in enterprise cloud deployments.
What is the concern with the default transaction mode?
Jepsen reported that ParticleDB continued to lose committed writes during normal operation with its default occ mode in version 1.1.18.11.rc-22. The report also found internal anomalies and Lost Update under the strongest, serializable setting.
Source: jepsen.io
