Jepsen tests ParticleDB 1.1.4, a new database unifying SQL, key-value, and vector search

ParticleDB Overview

ParticleDB is a new database engine developed by Orkes, the creators of the workflow orchestration platform Conductor. It is designed to unify diverse data workloads, including transactional SQL, key-value storage, vector search, and live analytics, within a single system. The database also supports user-defined functions called Particles and offers integration with existing systems via Postgres, Redis, gRPC, and HTTP2 protocols. ParticleDB is written in Rust and claims performance advantages over Postgres.

Development and Release

Orkes initially built ParticleDB to improve performance for their Conductor platform, which previously relied on Redis and Postgres. The 1.0.0 version had a limited feature set tailored for Conductor's queries and was deployed in some Conductor clusters. A major refactoring followed, adding a custom wire protocol, more datatypes, and additional OLAP queries. This refactored version, starting with 1.1.4, was soft-launched on particledb.ai, leading to Jepsen's engagement for testing.

Jepsen's Initial Assessment

Jepsen's report focuses on ParticleDB 1.1.4 and covers only single-node behavior, as the provided builds did not support the enterprise cloud-only replication feature. ParticleDB's homepage advertised a "formally proven" status with a TLA+ proof, but no TLA+ specification was available during the collaboration, leaving the scope and applicability of any such proof unclear. The database offers three Write-Ahead Log (WAL) sync modes: fsync, which calls fsync on each commit, and groupsync, which batches multiple transactions.