// the find
daos-stack/daos
DAOS Storage Stack (client libraries, storage engine, control plane)
DAOS is Intel's distributed object storage stack built for HPC and exascale clusters, designed around NVMe and persistent memory rather than spinning disk or generic cloud storage. It's for people running supercomputer-scale clusters who need a storage layer under MPI-IO, HDF5, Spark, or a POSIX namespace via dfuse — not for anyone building a general-purpose app backend.
The dependency footprint is honest about complexity instead of hiding it: PMDK for persistent memory, SPDK for userspace NVMe, mercury/argobots for RPC and async execution, with its own patches tracked explicitly in deps/patches rather than silently vendored. CI is unusually thorough for a C codebase this size — clang-format enforcement, Coverity static analysis, Trivy scanning, bandit on the Python tooling, and a dedicated functional/unit/storage test split in ci/, which matters a lot for a project touching raw block devices. It backs a genuinely wide set of application interfaces (MPI-IO, HDF5, TensorFlow-IO, native Python dict bindings, Spark/Hadoop connector) rather than inventing a bespoke API everyone has to adopt. Now governed as a Linux Foundation project, which is a real signal against single-vendor abandonment risk for something this infrastructure-critical.
988 stars after years of Intel backing and supercomputer deployments tells you the audience is tiny — this is not something you evaluate for a side project or even most company infra. The README itself is almost content-free; it exists to route you to docs.daos.io and a JIRA tracker, so you can't assess the project from the repo alone. Getting a working build requires real hardware or careful emulation of SCM/NVMe (there's a qemu-vms.md in QSG for exactly this reason), plus RPM packaging and cluster provisioning scripts — there's no docker-compose-up path to try it in five minutes. The control plane, storage engine, and client libraries are all in one monorepo with SConstruct as the build system, which means onboarding a new contributor means learning SCons idioms most developers haven't touched since 2010, on top of the domain-specific storage internals.