# openlake-project/openlake

OpenLake is a high performance storage engine for efficient LLM inference and GPU Training

Repository: https://github.com/openlake-project/openlake
Canonical: https://ross.abutalabs.com/products/openlake
Homepage: https://theopenlake.com
Language: Rust
License: Apache-2.0
License Family: permissive
Topics: gpu, high-performance, rdma, rust, storage, throughput, blackwell, gpt, llm, llm-training, model-serving
Last push: 2026-08-25T08:12:50+00:00

## Health v2 (maintenance only)
Score: 81/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 99, release rhythm 98, longevity 9
- inputs: {"age_days": 128, "days_push": 8, "days_rel": 19, "gap_med": 9.0, "n_releases_24m": 9}
- flags: young
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2330, forks 413 (observed 2026-08-28T04:06:37.642923+00:00)

## What it is
OpenLake is a high-performance distributed storage engine written in Rust (built on io_uring, RDMA, and GPUDirect) designed to feed GPUs during LLM inference and training. It provides petabyte-scale KV cache offload, fast checkpointing, vector serving, and S3-compatible object storage with sub-millisecond, million-IOPS performance.

## Use cases
- offload LLM KV cache from GPU memory to persistent storage
- speed up model checkpointing for RL and ML training runs
- reduce GPU idle time during training with fast small-file I/O
- store and reuse KV cache across vLLM or SGLang inference requests
- build a fast vector database for retrieval workloads
- store massive conversation context and memory for AI agents
- replace S3 bottlenecks in GPU data pipelines

## When to choose
- you run LLM inference at scale and KV cache recompute is hurting throughput
- training jobs are bottlenecked on checkpoint or small-file storage I/O
- you need GPU-direct data paths (RDMA/GPUDirect) to saturate accelerators
- you want S3-compatible APIs with far higher throughput on GPU hosts

## When to avoid
- you need a general-purpose cloud object store with multi-region replication
- your workloads are CPU-bound with no GPU acceleration
- you need a turnkey managed service rather than self-hosted infrastructure
- you run primarily on Windows or macOS, since it targets Linux with io_uring

## Facets
- artifact type: library
- maturity: active
- function: caching, object-storage, serialization, gpu-computing, llm-inference, llm-training, rag
- domain: large-language-models, databases, gpu-computing, infrastructure-as-code, microservices, machine-learning, artificial-intelligence
- platform: rust, self-hosted
- tags: storage-engine, kv-cache-offload, io-uring, rdma, gpudirect, checkpointing, vector-database, model-serving, high-performance-storage, linux, gpu, docker

## Member repositories
- openlake-project/openlake (main) score 81

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:06:37.642923+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T02:38:06.648799+00:00, confidence not recorded.
  - readme: https://github.com/openlake-project/openlake (fetched 2026-08-28T04:06:37.642923+00:00, sha b6af5745b00a)
  - homepage: https://theopenlake.com (fetched 2026-08-29T10:18:27.015510+00:00, sha 65e978cf72fc)
- Data as of 2026-08-30T08:39:29.467469+00:00.