openlake-project/openlake — reverse-engineered prompt

Build me a Rust based object storage service for AI workloads that’s focused on getting data from NVMe to GPU memory as fast as possible. I want it to feel like a real storage engine for training and inference clusters, with very high throughput, low latency, and an S3 compatible endpoint so normal S3 tools can talk to it.

Please include a server I can run on multiple nodes with a simple per node TOML config, plus a CLI that can do quick local benchmarks against a filesystem backend so I can confirm the build works and measure throughput fast. It should run well on Linux, and be able to build on macOS for development. If the platform supports it, use modern async I O and zero copy style data movement, and keep the hot path efficient.

I also want a practical quickstart, example configs, and a simple way to start a node and then create a bucket, upload a file, and list it with an S3 client. Look up current docs online if you need to.