01
Raw ClusterQueue quota does not express team or namespace GPU caps, allowed GPU flavors, required FinOps labels, GPU-hour or budget ceilings, or image registry allowlists. GPU jobs need an auditable gate before Kueue admits them.
Case 28
GPU Admission on Kueue
Fail-closed GPU admission on upstream Kueue: an AdmissionCheck controller that returns ALLOW, HOLD or DENY per Workload and writes an evidence envelope for every decision.
01
Evidence
From problem to proof.
How the system is framed, operated, and proven — not a slide deck summary.
02
A controller owns an AdmissionCheck (controllerName gpu-admission.justrunme.com/admission-check). After Kueue reserves quota, it evaluates a cluster-scoped GPUAdmissionPolicy and sets the Workload check to Ready (ALLOW), Retry (HOLD on soft quota pressure) or Rejected (DENY). Each decision emits an ASC-compatible evidence envelope stored as a ConfigMap. Upstream Kueue, not a fork.
03
Fail-closed by design: if the controller is down, checks stay Pending and Workloads are not admitted; a malformed or missing policy and undecidable input are Rejected. GPUs are emulated: the kind demo advertises nvidia.com/gpu as an extended resource, so nothing here is measured on physical GPUs.
04
v1 merged to main. 8 e2e scenarios on kind with upstream Kueue on this corpus: allow_golden, missing_labels, disallowed_flavor, bad_registry, ghcr_evil_prefix, over_quota_hold, malformed_policy, controller_down_fail_closed. CI run https://github.com/justrunme/gpu-admission-kueue/actions/runs/39766009986. Claims apply on this corpus only.
Constraints
Upstream Kueue only (no fork), AdmissionCheck API, fail-closed defaults: controller down keeps checks Pending, malformed or missing policy and undecidable input are Rejected.
02
Guides
Read the guide.
03
Related
Nearby systems.
Agent Security Contract
Agent and MCP stacks expand the attack surface beyond prompt text: tool misuse, poisoned tool descriptions, argument injection, and unapproved high-impact actions run without a pre-execution gate
0 false-ALLOW · 0 false-BLOCK on this corpus
AI Runtime Platform
AI workloads often start as API calls but become production systems that need identity, policy, scaling, telemetry, and clear ownership
Every model call: policy + trace + budget
LLM Infrastructure Runtime
LLM workloads move faster than traditional platform controls and can quickly become expensive, opaque, and hard to operate
LLM serving with operability signals
Available for meaningful infrastructure conversations