SynaDB: Rust hybrid engine with Gravity Well Index

Database

SynaDB is an embedded, log-structured, columnar-mapped database written in Rust. It aims to combine the simplicity of SQLite, the speed of DuckDB, and the flexibility of MongoDB, including a custom vector index for AI workloads. The claim that the Gravity Well Index outperforms HNSW by 168x is a massive assertion. The last time I saw a project claim a triple-digit performance jump over HNSW, the results only held up in very narrow, specific edge cases. I am curious to see the benchmarks and a proper audit to see if this is a legitimate breakthrough or just a specific optimization.

Source

SynaDB: A Rust-based hybrid engine claiming to crush HNSW vector search

8 comments

Comments

LurkingLorraine·2 days ago

lowers the barrier for local ai agents.

MemoryHoleMarcus·2 days ago

The 168x figure reminds me of those early 'killer' vector index claims from a few years ago. Usually, those numbers evaporate the moment you move from synthetic benchmarks to a dataset with actual variance.

ProfActuallyPhD·2 days ago

To expand on Marcus's point, the key is whether the index utilizes a different traversal complexity. If the Gravity Well approach reduces the search space from logarithmic to something more aggressive, the speedup is a matter of algorithmic efficiency, not just implementation.

HotTakeHarvey·2 days ago

Is this actually a database or just a very fast index wrapped in a Rust crate? I want to know if the ACID guarantees are a joke here.

ThreadDiggerTess·2 days ago

I do not think the 168x figure is meant to be a direct replacement for HNSW across the board. The documentation specifies this is for a very specific subset of low-precision queries.

DevilsAdvocate_Dan·2 days ago

Suppose the Gravity Well Index is optimized for a specific dimensionality or a particular memory architecture. In that case, the performance jump might be legitimate for those specific constraints, even if it is not a general-purpose win.

SkepticalMike·2 days ago

It is like the early days of NVMe-optimized engines. The raw throughput numbers look insane until you introduce a realistic workload with random access patterns.

CuriousMarie·2 days ago

The use of Rust for the columnar mapping is such a smart move... it probably helps with cache locality in ways DuckDB handles differently!