BareStax keeps its cluster boundaries explicit

BareStax manages an installed Ubuntu server through job scheduling. Hardware ownership, storage servers, and external load balancing remain with the institution.

Infrastructure defines the cluster boundary

Does BareStax use Kubernetes?

No. BareStax is not a Kubernetes distribution and does not place Kubernetes between jobs and GPUs. It applies declarative reconciliation to system settings, kernel, drivers, fabric, BMC, and job scheduling.

Does BareStax virtualize GPUs?

No. BareStax assigns the full GPU directly to a job and does not use MIG or vGPU. There is no virtualization layer in the compute path.

Does BareStax install the operating system?

No. Your institution installs Ubuntu 24.04 LTS on amd64 before a server joins. BareStax preserves that installation, does not erase the disk, and configures the compatible stack above it.

Which GPUs does BareStax support?

BareStax supports NVIDIA data center and workstation-class GPUs. This includes HGX systems from Hopper and Blackwell generations and servers with PCIe GPUs. Each release note identifies the hardware classes validated for that release.

Which cluster networks are supported?

BareStax supports InfiniBand and RoCE, and records NVLink or NVSwitch where present. Neither fabric is required, so Ethernet-only hardware can join. Discovery also records PCIe and NUMA topology before profile selection.

What cluster sizes are covered?

The small profile covers 2–4 GPU nodes. Standard scale runs from 5–128 GPU nodes, with a ceiling of 1,024 GPUs and a separate head node. Different GPU classes keep their own queues.

What storage systems does BareStax manage?

NFS and S3 are client targets. When shared storage is missing, BareStax blocks fallback writes to local disk and leaves the storage server under institutional control. Capabilities carries the health and impact details.

Operations cover the node lifecycle

How are updates moved into an air-gapped network?

barestax-bundler collects and signs a content-addressed release bundle on a connected network. Approved portable media or a data diode transfers it into the isolated network, where import verifies signature, integrity, compatibility, and the scan report. Four-eyes approval precedes canary and budgeted rollout from the cluster-local repository.

Does any runtime function require internet access?

No. Runtime components stay on the institution's network without a remote-server connection. The Security page lists what remains available offline.

Can BareStax take over an existing cluster in place?

No. BareStax does not import an existing scheduler configuration in place; it generates its own BareStax Core configuration, accounting included. Servers join one by one with their Ubuntu installation intact. Queues, tenants, quotas, and fairshare are defined in BareStax, while researchers submit work from the command line, templates, or the portal. Your team and BareStax plan the sequence around the existing hardware and queue structure.

What happens if the control plane stops responding?

Running jobs continue, and compute nodes stay on their last known safe state. Local health gates keep a faulty node from accepting new work while the control plane is unavailable. Backups cover identity, PKI, the database, and BareStax Core state for restoration to clean hardware.

Can a researcher see why a job is waiting?

Yes. The portal gives each researcher the waiting reason, quota, logs, and controls for personal jobs. Operators see queue, fairshare, accounting, health, and capacity context in the console.

How does BareStax work with Prometheus and Grafana?

Exporters use standard Prometheus format, so an institution's existing Prometheus and Grafana installation can read the same data. BareStax also integrates with Alertmanager. The heat map and deep monitoring dashboards inside the console are BareStax panels, not Grafana. A test alert verifies that the alert chain reaches the on-call recipient.

How are faulty nodes handled?

BareStax withholds new placement from an unhealthy node and lets its current job finish. Health and recovery covers escalation, flap protection, and the storm brake.

Can BareStax expose long-running model services?

BareStax runs long-lived services in a dedicated BareStax Core queue with service registration and protection. External access comes through the institution's own load balancer. BareStax does not provide load balancing or automatic scaling.

Which languages are available in the product interface?

English and Turkish are the product interface languages. Arabic is available on the website, not in the product interface. Console details cover themes and keyboard use.

Security controls identity and software trust

Which identity providers can connect to BareStax?

BareStax supports LDAP, Active Directory, OIDC, SAML, and local accounts. Built-in TOTP provides a second factor in connected and air-gapped environments. Critical operations ask for that factor again.

How does BareStax decide which container images may run?

Jobs run through Apptainer. An image enters the packaged OCI registry, passes scanning, receives approval from two distinct people, and is signed by the cluster key. Compute nodes reject an image without an accepted signature, while institutional registries and counter-signatures can join the same trust path.

Commercial terms remain explicit

How long is a release line supported?

A minor release line stays supported for 24 months and continues to receive security fixes. The contract defines response and resolution times by severity.

How does licensing work in an isolated environment?

Licensing remains offline. Price depends on cluster scope and support level; Deployment explains the commercial boundary.

A missing question can go straight to the product team

Write to hello@barestax.com if your question is not here; someone on the product team replies.