PLATFORM · INTEGRATION
How the Governance Layer Connects
"Above the scheduler" is an architectural claim. This page makes it concrete. Four integration patterns — eBPF probe, Kubernetes admission webhook, kernel module, and sidecar — each with the exact connection point, the state vector it reads, and the control signal it emits.
The Architectural Claim, Made Precise
The Omni-Compass governing convergence control plane law operates above the scheduler. This means it does not replace the scheduler, does not patch the scheduler, and does not require the scheduler to be modified. It observes the system state, evaluates the admissibility criterion, and emits a control signal that the scheduler receives as a constraint — not as a replacement for its own logic.
The governance layer reads the six-dimensional state vector S = (utilization, memory_pressure, thermal_load, latency_delta, power_draw, queue_depth) from the governed system. It evaluates V(S) = ‖S − S*‖² and checks the admissibility criterion. If the criterion is satisfied, it passes. If not, it emits a corrective control signal. The scheduler executes the corrected signal.
Four integration patterns implement this architecture at different layers of the stack. The pattern you use depends on where in the stack you are governing. All four read the same state vector and emit the same control signal format. The governing law is identical across all four — only the integration surface differs.
eBPF Probe
Linux kernel 5.8+. Zero kernel modification. Reads state from kernel data structures via BPF maps. Emits control via cgroup v2 resource limits.
Admission Webhook
Kubernetes 1.16+. Intercepts Pod/Deployment admission. Reads state from the Kubernetes API server. Emits control as mutated resource requests/limits.
Kernel Module
Linux kernel 4.15+. Loadable kernel module. Direct access to scheduler data structures. Emits control via scheduler hooks (sched_setattr).
Sidecar
Any container runtime. Runs as a container in the same Pod. Reads state from the container runtime API. Emits control via the runtime resource update API.
State vector — read by all four patterns
S = [
utilization, // CPU utilization [0.0, 1.0]
memory_pressure, // Memory pressure [0.0, 1.0]
thermal_load, // Thermal load [0.0, 1.0]
latency_delta, // Latency delta from baseline [ms]
power_draw, // Power draw [W]
queue_depth, // Scheduler queue depth [count]
]Control signal — emitted by all four patterns
control_signal = {
cpu_quota: float, // CPU quota adjustment [0.0, 1.0]
memory_limit: int, // Memory limit adjustment [bytes]
io_weight: int, // I/O weight [1, 10000]
net_priority: int, // Network priority [0, 7]
admit: bool, // Admit or reject the state transition
}DOWNLOAD VERIFIED DATA
EXTERNAL REFERENCES