Spot Capacity Orchestrator  ·  View 30 of 32  ·  7 · Assurance

Security and Trust Zones

Where the power to acquire, terminate and drain lives, and what a compromised workload can and cannot reach.

Editable source SVG draw.io All views
Workloads Workload no capacity rights Compromised workload the assumed attacker Node Node agent own node only SPIRE agent node-bound SVID Control plane sco-controller only acquirer OpenBao credential · 1 h Provider Nova · sco projects nobody else has a role Reclaimer notices, never commands unix socket bundle only scoped own SVID only Security — Who Can Acquire, Terminate and Drain External / third party Risk / gap Application we own Security / platform synchronous failure / alternate Reclaimable instances live in projects where only the orchestrator holds a role. Every capacity action is written to the audit log service through the controller's outbox (view 31). v 1.0 · owner Compute Platform Architecture · date 2026-09

Decisions

  • Termination authority is scoped by OpenStack project, not by tag. Reclaimable and fallback instances live in projects where only the orchestrator's identity holds a role; tags are mutable and a policy that trusts them can be talked out of its scope.
  • The orchestrator's provider credential is an application credential issued through OpenBao for an hour. The node agent holds no provider credential at all.
  • When an instance is terminated, the controller bans its SPIRE agent and revokes OpenBao leases issued to that node, rather than relying on the instance disappearing.

The assumed attacker

  • A compromised workload can reach the local agent socket, which answers notice and deadline queries only, and can obtain its own SVID. It cannot drain another node, launch or terminate capacity, change a contract or read another workload's checkpoints.

Omitted

  • Checkpoint isolation per workload is on view 13; the authority to spend beyond a cap is on view 31.