Found a graph database that uses S3 instead of Raft/etcd for coordination
Posted by HighwayMedium39@reddit | Python | View on Reddit | 0 comments
Just found this on GitHub and thought it was a pretty interesting approach to distributed databases.
NamiDB is a graph database written in Rust where the source of truth is object storage (S3 / R2 / GCS / Azure). Instead of using Raft, etcd, or ZooKeeper, it apparently relies on S3 conditional writes for coordination.
You can run it embedded like DuckDB, as a single binary server, or through their hosted cloud.
The benchmarks they shared claim warm reads were between 1.3x and 3.9x faster than Kùzu on some validated LDBC IC queries. They also openly mention Neo4j on local NVMe is still faster on raw latency, but their focus seems to be object-storage-native deployments: scale-to-zero, multi-tenant setups, cheap backups, etc.
The architecture itself is what caught my attention more than the benchmarks. Haven’t seen many databases treating object storage as the primary persistence layer instead of an archive layer.
Curious what people here think about this design tradeoff.