Status
Short maturity matrix for Atlas 0.7.0. History and narrative live in ROADMAP.md. When the two disagree, this file wins.
A cell is yes only for that column. Lab verification is not production support. Nothing below is bank or enterprise GA.
| Capability | Implemented and unit-tested | Verified on real infrastructure | Production-supported | Experimental | Planned |
|---|---|---|---|---|---|
| Ceph / Rook control plane (RBD, CephFS, RGW, jobs) | yes | yes, single-node Rook lab | no | ||
| NFS and ZFS drivers | yes, fake and real modes | no remote production target | no | real mode | SAN, cloud block, external Ceph import |
| Object storage (Ceph RGW default; bring-your-own S3) | yes | yes, single-node Rook lab (RGW) | no | RustFS was previously a first-party backend (driver, admin proxy, Helm chart, console pages) — removed; see docs/RUSTFS.md and CHANGELOG.md | |
macOS disks (atlas-driver-macos, atlas-mac-agent): discovery, guarded external-disk lifecycle, Mac fleet | yes, parsers on recorded plists, refusal matrix, fake mode, fleet tests | read-only discovery and the real agent against a real 1 TB external drive on a Mac (2026-10-17); erase/addVolume/deleteVolume/eject on real hardware, the launchd installer, HTTPS agent-to-gateway and NFS/SMB mounts from a Mac are not verified | no | real mode, write operations | Disks page UI, container resize, local snapshots, remote writes via the agent, SMART/capacity alerts, macFUSE mount (no-go for now) — docs/MACOS.md |
| Longhorn backend | yes, read-only (nodes/volumes via CRDs) | no | no | native write path (PVC provisioning goes through the Kubernetes path today, not this driver) | |
Raw disk provisioning (ZFS pool via zpool create; Ceph OSD via Rook CephCluster patch) | yes, fake mode + safety-check unit tests + GET /zfs/devices/GET /ceph/nodes/{node}/devices picker endpoints | ZFS: verified live end-to-end on the lab host through the console — wiped a disk carrying a real, stale Ceph OSD partition table, then zpool create succeeded (zpool status: tank0/ONLINE, backed by the real device, zero errors). Root-caused and fixed two real bugs along the way: zpool create's partition-reopen polls udev's device database (needs /run/udev hostPath-mounted, not just /dev), and the root/boot-disk check was blind to the host's real mount table in a container (needs a hostPath-mounted /proc/1/mountinfo, ATLAS_HOST_MOUNTINFO_PATH) — see docs/DISKS.md. Ceph: never run against a Rook cluster with the device-discovery daemon enabled (no Rook cluster on this lab host to test against) | no | yes — depends on the target Rook install running ROOK_ENABLE_DISCOVERY_DAEMON | stand up a Rook cluster to verify the Ceph OSD path; (pool destroy → wipe → re-provision was run live on the lab: tank0 destroyed, sdb re-provisioned as tank1) |
SQLite default, Postgres query layer, Helm database.kind, cross-replica rate limiting | yes | yes, CI postgres-test and deploy/postgres-lab/ | no — needs a real HA Postgres | enterprise IdP for OIDC (Dex lab only) | |
| Auth, tenants, quotas, audit export, Vault resolution | yes | yes, lab (Dex, Vault, SIEM receiver) | no | ||
| DataBridge Postgres | yes | yes, through cutover | no | TLS still sslmode=disable | verified TLS migration |
| DataBridge MariaDB | yes | yes, through cutover | no | repeatable CI/lab automation | |
| DataBridge MongoDB | yes | yes, through cutover | no | change-stream edge cases, repeatable verification | |
| DataBridge MySQL | yes | through cutover, in-cluster on Rook Ceph, incl. TIMESTAMP columns and composite keys (2026-10-04) | no | ||
| DataBridge SQL Server and Oracle | yes | through cutover into CloudNativePG, in-cluster on Rook Ceph (2026-10-04); the Debezium initial snapshot is the full-load, validate is advisory | no | rollback drill; Oracle TCPS | |
| Cross-cluster DR (RBD mirror) | yes, incl. the clean-failback promote guard | yes, two lab Rook clusters: one-way (2026-10-04) and two-way with clean failback (2026-10-06), journal and pool modes and local consistency-group snapshots/rollback (2026-10-06), docs/DR.md; dataplane_verified stays a per-deployment flag | no | cross-site group mirroring (not in released Ceph) | |
| Atlas Native filesystem (Raft metadata, EC, tiering, NFS/SMB/S3/CSI, replication, edge profile) | yes | yes, lab k3s and Rook Ceph (docs/NATIVE_STORAGE.md); no NVMe or RDMA benchmarks | no | yes | RDMA, GPUDirect Storage, IO500 submission |
| Application recovery operator tool | yes, dependency/HTTP/hook failure scenarios | actual gateway plus one metadata/one data Native node on temporary files; no live Ceph/PVC or database hook qualification | no | operator hooks and capture coordination | multi-site application recovery and HA workflow resumption |
| Ops Advisor, incidents, what-if, anomalies | yes, read-only | console exercised on the lab gateway | no | persisted findings; no execution | |
MCP server (gateway: 27 read-only tools; atlas-io agent: 6; opt-in guarded create_snapshot and ack_alert) | yes, incl. two-phase token, tenant and role tests | lab gateway and BPF agent: remote Host header, io_risk, agent bearer auth (docs/MCP.md); write tool not enabled on the lab | no | ATLAS_MCP_WRITE_ENABLE, ATLAS_IO_TOKEN | MCP prompts/resources, more read tools, guarded alert ack |
eBPF storage I/O sensor (atlas-io-agent) | yes — fake source, histograms, attribution, RCA, fail-open leases, HTTP/Prom, atlasctl io | yes — aya/CO-RE block-layer attach, queue time and the Atlas Native maps against a real native cluster, on a kernel 7.0 lab host (docs/IO_EBPF.md) | no | live mode (bpf feature) | NFS/ZFS/uring probes, cgroup→volume map, Observatory heatmaps |
Product integrations beyond gRPC Owner | gRPC owner surface yes | no | Transiva import, Veyron, GuestKit, PacketWolf, then a small SDK |
Transiva's owner id on the wire remains hyper2kvm. v0.5.0 does not rename it.
Unresolved: Zeus OS already has an atlas module ("Machine Finder") and a Storage Center
UI. Atlas does not yet absorb, replace, or sit beside that UI. Do not treat the names as settled.