A1

AIStore 1.4 Scalable Object Storage for AI and ML Datasets on Ubuntu 24.04 LTS

Azure Storage

AIStore is a lightweight object storage system built for AI and machine learning datasets, serving them over its own API and an S3 compatible API from storage you control.

Base
Hardened build
minimal ports, security patches applied at build time
Access
Unique credentials
generated on first boot, readable only by root
Verified
Boots working
services pass a health gate before release
Support
24/7, 365 days
by email and live chat, 24 hour response SLA

Overview

AIStore (AIS) is an open source object storage system from NVIDIA, designed for the read heavy, high throughput access patterns of AI and machine learning training. It stores objects in buckets on local disks and presents them two ways at once: through its own cluster aware API and command line client, and through an S3 compatible REST API that unmodified S3 tools and SDKs such as the AWS command line client and boto3 can use directly. That makes it a drop in home for training datasets, model checkpoints, shards and other large artifacts, without rewriting the code that reads them.

Because it speaks S3, existing data pipelines can point at it unchanged; because it is a real storage system rather than a gateway, the data lives on volumes you own and can be listed, versioned, checksummed and served at speed to compute nodes on the same network.

Why the cloudimg image

cloudimg ships AIStore as a complete single node cluster, a gateway and a storage target already wired together and managed by the system service manager, so buckets can be created and objects read and written the moment first boot finishes, with nothing to compile and no container platform required. Object data lands on a dedicated volume rather than the operating system disk, so storage and the operating system are kept apart from the start. Most importantly the appliance ships closed: AIStore is unauthenticated in its default upstream form, so cloudimg turns its own token authentication on in the shipped configuration and generates the signing key and the administrator password uniquely on each instance at first boot into a root only file. Nothing binds to the network at all until those secrets exist, so there is no window in which an open object store is reachable, and an unauthenticated request is refused on both the native and the S3 compatible endpoint. The base is fully patched with unattended security upgrades enabled, and every deployment is paired with a step by step deploy guide, backed by 24/7 cloudimg support.

Common uses

  • Serving training datasets and shards to machine learning compute nodes
  • An S3 compatible bucket store for pipelines you keep on your own network
  • Holding model checkpoints, artifacts and other large objects