OpenLake
OpenLake projectOpenLake is a distributed storage engine for LLM inference and GPU training, built in Rust on io_uring and aimed at small-I/O throughput and sub-millisecond reads. Two surfaces sit on the same engine: an offload pool that spills KV cache off GPU nodes, and a petabyte-scale S3-compatible object store. A vLLM connector ships in the repository.
Openness
5 high confidence- license
- Apache-2.0(OSI)
- source
- public
- core features withheld
- no
Apache-2.0 license body confirmed. The repository is public and unarchived and builds the whole product, and the README describes no paid tier, enterprise edition or license-gated build beside it, so source is public and the core ungated.
- https://raw.githubusercontent.com/openlake-project/openlake/main/LICENSE recorded 2026-09-30
LICENSE file is the Apache-2.0 text: "Apache License" and "Version 2.0, January 2004" and "TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION".
- https://raw.githubusercontent.com/openlake-project/openlake/main/README.md recorded 2026-09-30
README carries the Apache-2.0 badge and ends with "[Apache 2.0](LICENSE)." and describes building the whole store from source with "cargo build --release --locked -p openlake_server --bin openlaked"; searched the body for enterprise, paid, license key, commercial and proprietary: none.
Adoption
2 low confidenceOpenLake is built from source or run from container images and Helm charts, with no package-registry channel of its own, so GitHub stars are the only adoption signal available. A star reflects interest rather than confirmed use, so it cannot show how many deployments are actually running OpenLake.
- https://api.github.com/repos/openlake-project/openlake recorded 2026-08-18
Repo metadata - stargazers_count = 2,330 - for openlake-project/openlake.
Capability
4 medium confidenceOpenLake is a durable, cluster-capable store: GPU nodes contribute to one cluster, and data and parity are erasure-coded across drives, the same kind of scale-out store as AIStore and JuiceFS. It stores and returns objects and cached KV blocks rather than retrieving across modes, and hosts no model in its serving path, so it sits one step below Vespa.
- https://raw.githubusercontent.com/openlake-project/openlake/main/README.md recorded 2026-10-08
README: "Distributed storage for GPU workloads"; "fully persistent and durable"; "GPU nodes contribute to create an OpenLake cluster"; "SIMD Reed Solomon erasure coding distributes data and parity across drives"; a Helm guide places "one OpenLake instance on each selected node"; sections for KV offload and a PB-scale object store, with a vLLM connector.
Verified 2026-09-27