maxprocs

Overview ¶

Package maxprocs caps GOMAXPROCS to the Kubernetes CPU request so the Go scheduler doesn't create more OS threads than the pod is guaranteed.

Why not CPU limits / automaxprocs ¶

Real-time media workloads (egress, ingress) intentionally omit CPU limits to avoid CFS throttling — even brief throttle events cause audio glitches and video frame drops. Without a limit the cgroup quota is "max", so libraries like uber/automaxprocs fall back to runtime.NumCPU (the full node). This package uses the CPU request instead, which reflects what the Kubernetes scheduler actually guarantees to the pod.

When a CPU limit IS set, the cgroup quota gives a tighter bound and Go (or automaxprocs) already derives GOMAXPROCS from it. This package still works in that case — it caps GOMAXPROCS down to the request but never increases it, so the stricter of (limit, request) wins.

Usage ¶

The CPU request is read from the LIVEKIT_CPU_REQUEST environment variable, which should be populated e.g via the Kubernetes Downward API:

env:
  - name: LIVEKIT_CPU_REQUEST
    valueFrom:
      resourceFieldRef:
        containerName: <name>
        resource: requests.cpu

Without this variable the package is a no-op. A blank import is sufficient — init runs at process startup:

import _ "github.com/livekit/protocol/utils/hwstats/maxprocs"