DAOS
What it is
DAOS is a free, open-source, software-defined storage system for artificial intelligence and high-performance computing workloads. Its primary interface, exposed through libdaos, is a scalable key-value-array store. POSIX namespaces, HDF5 datasets, Python, MPI-IO, PyTorch, Spark and Hadoop, and TensorFlow-IO are among the documented data models and integrations. Storage pools can use some or all storage engines, while containers hold user-managed datasets that can be snapshotted. Fine-grained access-control lists apply to pools and containers. Objects may use erasure coding, replication, or no protection, and self-healing rebuilds can start when an engine stops responding. DAOS distributes filesystem data and metadata across storage engines and can use libfabric or UCX for RDMA. It supports S3-compatible access, Kubernetes, hybrid deployment, and on-premises or cloud environments. Version 2.8 targets Linux on x86_64 and lists specific SLES, openSUSE Leap, and EL distributions.
For and against
For
- Free and open source
- Key-value, POSIX, HDF5, MPI-IO and S3 interfaces
- Erasure coding and replication options
- Fine-grained pool and container ACLs
- Self-healing rebuilds for unresponsive engines
Against
- Linux x86_64 focus
- Version 2.8 excludes Ubuntu and Fedora
- Self-hosted deployment requires storage infrastructure
The verdict
DAOS combines a scalable storage interface with several data models, protection choices, and workload integrations. Check the listed Linux distribution limits and infrastructure requirements before adopting it.
Who it suits
DAOS suits organizations running AI or HPC workloads that need scalable storage, snapshots, access controls, and multiple data interfaces. It is intended for Linux-based, self-hosted or hybrid deployments.
On the record
- Access control
- Fine-grained ACLs can be set on storage pools and individual containers.daos.io · read 3 Oct 2026
- Audience
- The site presents DAOS as an open-source storage platform for AI and HPC workloads.daos.io · read 3 Oct 2026
- Data model
- Its core interface is a scalable key-value-array store, with data models including POSIX namespaces and HDF5 datasets built over it.daos.io · read 2 Oct 2026
- Data models
- The overview lists POSIX namespaces, HDF5 datasets, and Python among data models built over libdaos.daos.io · read 3 Oct 2026
- Data protection
- Objects can use erasure coding, replication, or no data protection; DAOS can trigger self-healing rebuilds when an engine becomes unresponsive.daos.io · read 2 Oct 2026
- Datasets and snapshots
- DAOS containers represent datasets, can be snapshotted, and support fine-grained access control.daos.io · read 2 Oct 2026
- Deployments
- DAOS can be deployed on premises or in cloud environments.docs.daos.io · read 3 Oct 2026
- Framework support
- The DAOS v2.8 documentation lists support topics for MPI-IO, HDF5, Python, PyTorch, Spark and Hadoop, and TensorFlow-IO.docs.daos.io · read 3 Oct 2026
- Integrations
- The DAOS client ecosystem includes libdfs for POSIX namespaces, and its user guide documents a PyTorch interface for loading data from DAOS.docs.daos.io · read 2 Oct 2026
- License
- The project repository displays the BSD-2-Clause Plus Patent License.github.com · read 3 Oct 2026
- Network support
- DAOS can use libfabric or UCX and lists Ethernet, RoCE, InfiniBand, Omni-Path, and Slingshot fabrics.daos.io · read 3 Oct 2026
- Notable limits
- DAOS 2.8 does not support Ubuntu, Fedora, CentOS Stream, or RHEL 8 and its clones.docs.daos.io · read 2 Oct 2026
- Performance
- The site describes distributed metadata, ultra-high bandwidth and IOPS, and time-to-first-byte below 20μs when using RDMA where available.daos.io · read 3 Oct 2026
- Protection and recovery
- Objects can use erasure coding or replication, and DAOS can trigger self-healing rebuilds when an engine becomes unresponsive.daos.io · read 3 Oct 2026
- Purpose
- DAOS is an open-source, software-defined storage system for AI and HPC workloads.daos.io · read 3 Oct 2026
- Release
- The DAOS documentation identifies its current documentation set as version 2.8.docs.daos.io · read 3 Oct 2026
- Security
- DAOS supports access control lists on pools and individual containers.daos.io · read 2 Oct 2026
- Snapshots and scale
- DAOS containers are user-managed datasets that can be snapshotted; the site says a POSIX container can hold trillions of files or directories and petabytes of data.daos.io · read 3 Oct 2026
