Putting Swift's S3 API to Work for AI Workloads on OpenStack Flex
I'll be honest. When someone says "object storage for AI" the picture in most people's heads is a purpose-built S3 appliance with a six-figure invoice attached, or a hyperscaler bucket with an egress meter running. Swift, the OpenStack object store that has been quietly holding petabytes since 2010, rarely makes the list. It's old. It doesn't have a launch video. It speaks S3 through a middleware rather than natively.
Our team has been leaning on that middleware harder every month. Model weights, training shards, and lately the KV cache that vLLM spills out of GPU memory all land in Swift buckets through the S3 API, and at some point "it seems fine" stops being an acceptable answer to "how fast is it." So we did what you're supposed to do. We picked an open source benchmark, built a container around it, ran the object sizes that AI workloads actually produce, and kept the numbers.