Request a demo

Everything you need to evaluate BareStax

The technical overview and two working checklists, open to download without a form. Use them to prepare a pilot, brief your security team, or compare approaches.

Downloads

  • Technical overview

    Architecture, lifecycle, operations, security, deployment, and licensing in one document. Download the PDF

  • Pilot acceptance checklist

    What to prepare before a pilot, and the drills that prove acceptance. Download the PDF

  • Air-gap update checklist

    Every step a release takes, from the connected side to the last node. Download the PDF

  • Security contact

    Report a vulnerability through the published contact in our security.txt file. Open security.txt

Pilot acceptance checklist

A paid pilot runs on your own hardware for 90 days, starting when the hardware is ready. Acceptance is proven by technical drills your team witnesses.

Before the pilot

  • A separate head node and up to 4 GPU nodes, racked and cabled.
  • Ubuntu 24.04 LTS, amd64, installed on every server. BareStax keeps your installation.
  • The cluster network. InfiniBand, RoCE, NVLink, and NVSwitch are supported; plain Ethernet also works.
  • Your identity source: LDAP, Active Directory, OIDC, SAML, or local accounts.
  • Your shared storage targets: NFS or S3.
  • For an air-gapped site, your transfer method: approved portable media or a data diode.

Drills that prove acceptance

  • Node join: a server joins with one command and one approval.
  • Two-node validation: the GPU, NVLink, InfiniBand, and two-node NCCL checks pass before the node takes jobs.
  • Canary update: a signed release reaches one node first, then the rest within your disruption budget, while running jobs continue.
  • Drain and recovery: a faulty node stops taking new work, running jobs finish, and recovery follows policy.
  • Air-gapped import: a release bundle crosses the boundary, its signature, integrity, and compatibility are verified, and two people approve it.
  • Audit export: the hash-chained audit record is exported for your review or your SIEM.
  • Control plane restore: a backup is restored to clean hardware, and the old head node is prevented from answering.

After the pilot

The pilot license is time-limited. If you place a production order within 60 days, the full pilot fee counts toward the first year.

Download the checklist as a PDF

Air-gap update checklist

On the connected side

  • The bundler collects everything the release needs.
  • Upstream repositories are pinned to timestamped snapshots, so the same release produces the same bytes on every install.
  • The output is a signed, content-addressed release bundle.

Across the boundary

  • The bundle moves on approved portable media or through a data diode.

Inside the air-gapped network

  • The import verifies signature, integrity, and compatibility, and attaches a scan report.
  • Two people approve the promotion; the requester cannot approve their own request.
  • Nodes install packages from the cluster's local repository.

Rollout

  • The release reaches one canary node first, then the rest of the cluster within your disruption budget.
  • Running jobs continue. Changes that need downtime wait for your maintenance window.
  • The control plane is upgraded the same way.
barestax-bundler export
barestax-bundler verify
barestaxctl release import <bundle>
barestax-server upgrade --offline

Download the checklist as a PDF

Glossary

TermMeaning in BareStax
Air-gappedA network with no connection to the internet. BareStax installs, runs, updates, and licenses the cluster fully inside it.
Control planeThe software on the head node that holds the declared state of the cluster and runs the console, API, and audit record.
Head nodeThe server, or virtual machine, that runs the control plane. It is not counted as a license unit.
NodeA GPU server in the cluster. Its agent connects to the control plane; nothing pushes commands into it.
TenantA group of users with its own queues and quotas. Tenants cannot see each other's work.
QueueWhere jobs wait for GPUs. Each GPU class gets its own queue.
DriftA difference between a node's declared and observed state. Drift without downtime is corrected in service; drift that needs downtime waits for the maintenance window.
Four-eyes approvalA critical action needs a second person, who reads the impact and types the target name. The requester cannot approve their own request.
Release bundleA signed, content-addressed package with everything a release needs, built on the connected side.
Data diodeA one-way link that lets release bundles into an air-gapped network.
Impact viewThe jobs, tenants, and services a node, storage target, or alarm affects, with verified and possible impact kept apart.
RunbookThe offline guide attached to every error code.
Canary updateA release that reaches one node first before the rest of the cluster.
Disruption budgetThe largest number of nodes that may be out of service at the same time.
DrainStopping new work on a node while its running jobs finish.
CordonStopping new work on a node by hand, for maintenance or deeper diagnosis.

Bring these to your first call

GPU node count and network type are enough to start. We go through the rest with you.