Atlas
One API for the storage you have. A filesystem for the storage you need.
Atlas is two things in one Apache-2.0 platform. A control plane that maps volume, snapshot and bucket intent to Ceph, NFS, ZFS, Longhorn and a Mac's own APFS disks through pluggable drivers. And Atlas Native, a distributed filesystem with its own data path: striped and erasure-coded extents, sharded Raft metadata, NFS, SMB and S3 gateways, replication to a second site, and a single-node edge profile.
4 backends + Native · 150+ REST endpoints · REST · gRPC · SSE · Two-way DR, clean failback · 6 DB engines in DataBridge
Quickstart · Docs · Architecture · API · DR · Status · Comparison
Why teams pick Atlas
| When this happens… | Atlas gives you… |
|---|
| Every product talks to Ceph pools, NFS exports and ZFS datasets its own way | One intent API (volumes, snapshots, clones, CephFS RWX, S3 buckets) over REST and gRPC |
| You need a scale-out filesystem but not another vendor's appliance | Atlas Native: your own nodes, Raft metadata, erasure coding, NFS/SMB/S3/CSI, under Apache-2.0 |
| A second site exists, but nobody trusts the failback | Two-way RBD mirroring and native replication, with a promote guard and a status endpoint that says when it is safe |
| Edge sites have one small box and no storage team | The Native edge profile: one pod, small caches, replication or snapshot export back to the core |
| Tenants share storage with no guardrails | Per-tenant quotas, rate limiting, OIDC/SSO, Vault-backed secrets and token revocation |
| Databases must move from the cloud to the edge | DataBridge: discovery, full load, CDC and cutover onto edge databases on Ceph |
| You want to try it without a storage cluster | make run starts the gateway against a fake driver so you can click through the whole console |
One API, every backend
Products call stable Atlas APIs, and the StorageDriver trait maps the intent to a backend. Drivers cover Ceph (RBD, CephFS, RGW), NFS, ZFS, read-only Longhorn, macOS disks (APFS containers and volumes, external drives, a fleet of Macs through a read-only agent, guarded erase/add/delete/eject on external disks; docs) and Atlas Native, each with fake and real modes; real drivers never fabricate data. Around it: DataBridge cloud-to-edge database migration, Day 2 operations (alerts, maintenance, upgrade preflight, quotas, audit export), and a read-only Ops Advisor that scores risk and plans capacity, also exposed as MCP tools. Architecture →
Agent-ready operations
Atlas speaks the Model Context Protocol so an AI agent can look at your storage the way an operator would, with the same roles and tenant limits as the REST API. The gateway's MCP endpoint (/api/atlas/v1/mcp) exposes 27 read-only tools: inventory, the Ops Advisor (risk, incidents, anomalies, what-if), eBPF-backed I/O risk, Ceph status, DR status and preflight, DataBridge, the audit trail, jobs and snapshots. The atlas-io node agent serves its own authenticated /mcp with six I/O tools. The only writes are opt-in (ATLAS_MCP_WRITE_ENABLE=1): create_snapshot and ack_alert (record that an operator saw an alert), each two-phase with a single-use, five-minute confirmation token bound to the caller and arguments, and audited. Failover, delete, restore, disk wipe and token minting are never exposed. MCP can never trigger a call to an external LLM. MCP guide → · Ops Advisor → · I/O risk → · Agent pack →
Run the filesystem yourself
Atlas Native is a distributed filesystem Atlas runs itself: striped, checksummed, copy-on-write extents with direct client reads and io_uring on raw devices; the namespace sharded across Raft groups; 3 replicas or Reed-Solomon k+m with a rebuild controller; cold extents tiered to S3; POSIX over FUSE, NFS, SMB, S3 and a Kubernetes CSI driver; quotas, POSIX ACLs and a single-node edge profile. Native storage →
Fail over, fail back cleanly
GET /dr/mirrors/{id}/status shows each site's live rbd mirror image status and whether a promote would be clean. Atlas refuses a non-forced promote until the local rbd-mirror has replayed the peer's demotion, and ?force=1 stays available for real disasters. Native clusters get asynchronous replication with read-only replicas, promote and demote. DR guide → · Native replication →
Proof, not promises
Everything below ran on real infrastructure in the Atlas lab. Lab verification is not production support (STATUS.md).
- Two-way RBD mirroring between two Rook Ceph clusters, clean failback, unsafe promote refused, checksums matched at every step (DR.md).
- Edge profile: catalog memory flat at about 26 MB from 200k to 400k files, where defaults reach about 196 MB (NATIVE_EDGE.md).
- Erasure coding: about 1.5 GiB/s encode and 2 GiB/s two-shard rebuild per core on 4 MiB extents (NATIVE_STORAGE.md).
- NFS, SMB, S3 and CSI on Native verified with real clients and a single-node k3s (NATIVE_NFS.md).
- DataBridge through CDC and cutover for six source engines (DATABRIDGE.md).
Not yet: NVMe and RDMA benchmark numbers, an IO500 submission, GPUDirect Storage, and cross-site mirroring of RBD consistency groups. See COMPARISON.md.
Atlas vs Rook alone, and vs WEKA
| Atlas | Rook alone |
|---|
| What it is | Storage control plane with drivers per backend, plus its own filesystem | Kubernetes operator that runs Ceph |
| Interface for products | Intent-based REST, gRPC and SSE APIs | Kubernetes custom resources and CSI PVCs |
| Backends | Ceph, NFS, ZFS, Longhorn, macOS disks, Atlas Native, generic S3 | Ceph |
| Cross-site DR | Two-way RBD mirroring with a guarded failback; native replication | RBD mirroring CRDs |
| Database migration | DataBridge (discovery, full load, CDC, cutover) | Not in scope |
| Choose Rook alone when | | You only run Ceph inside Kubernetes and its own dashboard and CRDs are enough |
Compared with a parallel filesystem such as WEKA, including where WEKA is the better choice today: docs/COMPARISON.md.
Leaving VMware
Atlas is the storage layer, not a hypervisor, so it sits next to VMware's vSAN rather than replacing the whole hyperconverged stack. VMs move with the rest of the Zyvor suite: Scout scores readiness, h2kvm repairs the guest offline and converts it to qcow2, and the VM lands on Kairon, Machina or KubeVirt with its disks on Atlas-managed Ceph, NFS, ZFS or Native storage. For raw images, an image.import job streams the disk into an atlas-native volume with resume and a read-back checksum (IMAGE_IMPORT.md); it does not read qcow2 or write to Ceph RBD. See vSAN and migration in COMPARISON.md.
Application recovery drills
Capture coordinated recovery points, restore into new drill volumes, validate application
startup or Native block checksums, measure recovery objectives, and retain cleanup evidence.
The operator tool uses existing Atlas APIs and supports dependency ordering, bounded hooks,
and scheduled drills: Application recovery guide.
AI data services (experimental)
The Python companion ships nine workflows together: shared model chunks, data-ready
job reservations, agent workspace branches, permission-aware discovery, exact-key
KV storage, immutable dataset releases, application recovery checks, per-job cost
estimates and optional Apache Iceberg maintenance. It uses a separate authenticated
API and CLI; external schedulers and serving engines integrate through explicit SDK
adapters. Guide and limitations.
The console
Full tour: Gallery. Screenshots come from a lab deployment.
Quickstart
make run # gateway with a fake driver, no Ceph needed
# Console: http://127.0.0.1:5110
Atlas Native on Kubernetes: helm upgrade --install atlas-native deploy/helm/atlas-native -n atlas-native --create-namespace, or add -f deploy/helm/atlas-native/values-edge.yaml for a one-pod edge site. More: Getting started · Deployment · Native FS · Chart values.
What it is, and is not
Atlas is Apache-2.0 software you can use in production and in commercial products (licensing). What is verified where, and what is lab-only (including the experimental live eBPF attach), is in docs/STATUS.md. History: CHANGELOG.md.
Part of the Zyvor stack
Aether runtime portability plane · Zorvia KubeVirt VM platform · Kryton · Transiva VM migration · zyvor.dev
Support
The current release is Apache-2.0 (older tags used other licenses, see the CHANGELOG). If it helps you, you can donate any amount you like: https://rzp.io/rzp/utwdYh3. A donation is voluntary and supports development and maintenance of the current code. It is not a purchase: it buys no support, SLA, features, priority or license, and it is not tax-deductible (Zyvor AI Labs Private Limited is a company, not a charity). Zyvor Enterprise (supported releases, SLAs, a named contact and enterprise features) is separate: [email protected].
License and support
Free and open source under the Apache License 2.0 (NOTICE). Zyvor Enterprise adds supported releases, deployment and upgrade guidance, priority incident triage, a named technical contact and 24x7 critical intake: plans · pricing · [email protected].
Contributing: CONTRIBUTING.md, under the CLA and DCO (git commit -s), governed by the Code of Conduct. Security issues: SECURITY.md.
Put every storage backend behind one API, and run the filesystem yourself
