Before we begin… yes, we know.
DuckDB is OLAP. KeraDB is OLTP.
One is built to slice through massive analytical datasets like a lightsaber.
The other is optimized for fast, tiny transactional operations.

This benchmark is not a fair fight — and we knew that.
But developers kept asking questions, and curiosity got the best of us.
So we ran it anyway using Rust’s criterion.rs benchmarking suite.

And in the spirit of DuckDB:
Quack quack. 🦆
Let's see what happened.

We're excited to share that KeraDB is now available on PyPI, making it easy to use KeraDB directly in Python applications.

You can now install KeraDB with a single command:

pip install keradb

This release makes KeraDB accessible to Python developers building local-first apps, agents, data tools, and embedded systems that need fast document storage and vector search without running an external database.

This post shares our KeraDB vs RocksDB benchmark results, and let’s just say it upfront:
RocksDB absolutely outperformed us in most categories.

But that’s okay.

Because RocksDB has:

• 10+ years of optimization,
• dozens of Facebook engineers behind it,
• and billions of production workloads tuning it daily.

KeraDB has:

• one maintainer with coffee,
• dreams,
• and version v0.1.0.

So yes — RocksDB is faster. But KeraDB is just getting started, and this benchmark gives us a map for where to improve next.
(Also: we did win one category. Let us enjoy that.)

This post summarizes the results of our v0.1.0 benchmark comparison between KeraDB, an embedded NoSQL + vector search database, and SQLite, the widely used embedded relational engine.
The goal is to highlight the strengths of each system and provide transparency on where KeraDB excels today—and where we plan to optimize next.

We're thrilled to announce the official launch of KeraDB — a lightweight, embedded database with MongoDB-compatible APIs and built-in vector search capabilities.