# OpenLake

A storage engine, written in Rust on io_uring, for AI infrastructure that keeps GPUs fed by offloading and serving KV cache, vector indexes and model checkpoints with high-throughput, low-latency I/O.

OpenLake is a distributed storage engine for AI infrastructure, built in Rust on io_uring. It targets the I/O patterns of GPU workloads, small, random reads and writes at high throughput and low latency, while remaining fully persistent and durable.

It offloads and serves LLM inference KV cache at large scale, reducing time to first token for long or repeated prompts by keeping a distributed cache across GPU hosts. It also provides a vector store for fast index building and serving, fast checkpoint storage and retrieval for reinforcement-learning and machine-learning training, small-file I/O for model training, and storage for large conversation and context histories used in agentic retrieval.

A connector integrates it with vLLM, and it installs as a Python package alongside a background daemon process.

Website: https://github.com/openlake-project/openlake