OpenClaw Resource Limits and Tuning on Kubernetes

💡 Quick Answer: Start with 512Mi/250m requests, 2Gi/1 CPU limits for the gateway. Add 1-2Gi for Chromium sidecar. Use VPA to right-size after a week of real usage data. The init container needs minimal resources (64Mi/100m limits).

The Problem

OpenClaw resource needs vary dramatically: a simple chatbot uses 200MB, while an agent running browser automation, multiple concurrent sessions, and heavy tool use can consume 4GB+. Under-provisioning causes OOMKills and slow responses; over-provisioning wastes cluster resources.

The Solution

Resource Profiles

Light — Single user, text-only

resources:
  requests:
    memory: 256Mi
    cpu: 100m
  limits:
    memory: 512Mi
    cpu: 500m

Standard — The official default

resources:
  requests:
    memory: 512Mi
    cpu: 250m
  limits:
    memory: 2Gi
    cpu: "1"

Heavy — Browser automation + multi-agent

# Gateway container
resources:
  requests:
    memory: 1Gi
    cpu: 500m
  limits:
    memory: 4Gi
    cpu: "2"
# Chromium sidecar
resources:
  requests:
    memory: 1Gi
    cpu: 500m
  limits:
    memory: 2Gi
    cpu: "1"

What Consumes Resources

graph TD
    A[OpenClaw Memory Usage] --> B[Node.js Runtime ~150MB base]
    A --> C[Active Sessions ~50MB each]
    A --> D[Tool Execution ~100-500MB]
    A --> E[File Operations ~variable]
    A --> F[Chromium Sidecar ~500MB-2GB]

Init Container Resources

The init container runs briefly and needs minimal resources:

initContainers:
  - name: init-config
    resources:
      requests:
        memory: 32Mi
        cpu: 50m
      limits:
        memory: 64Mi
        cpu: 100m

Right-Sizing with VPA

Deploy Vertical Pod Autoscaler to get recommendations:

apiVersion: autoscaling.k8s.io/v1
kind: VerticalPodAutoscaler
metadata:
  name: openclaw-vpa
  namespace: openclaw
spec:
  targetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: openclaw
  updatePolicy:
    updateMode: "Off"  # Recommendation only
  resourcePolicy:
    containerPolicies:
      - containerName: gateway
        minAllowed:
          cpu: 100m
          memory: 256Mi
        maxAllowed:
          cpu: "4"
          memory: 8Gi

# Check recommendations after ~1 week
kubectl get vpa openclaw-vpa -n openclaw -o jsonpath='{.status.recommendation}'

Storage Sizing

Common Issues

OOMKilled — Memory Limit Too Low

kubectl describe pod -n openclaw -l app=openclaw | grep -A3 "Last State"
# OOMKilled → increase memory limit

# Check actual usage
kubectl top pod -n openclaw

CPU Throttling — Slow Responses

# Check throttling
kubectl exec -n openclaw deploy/openclaw -- cat /sys/fs/cgroup/cpu.stat
# Look for nr_throttled > 0

Increase CPU limit or switch to burstable QoS (no CPU limit):

resources:
  requests:
    cpu: 250m
    memory: 512Mi
  limits:
    memory: 2Gi
    # No CPU limit → burstable, can burst to node capacity

Best Practices

• Start with standard profile — 512Mi/250m requests, 2Gi/1 CPU limits
• Use VPA in recommendation mode — right-size after real usage data
• Consider burstable CPU — omit CPU limits for bursty workloads
• Always set memory limits — prevent OOM from affecting other pods
• Size init container separately — 64Mi is plenty for config copy
• Monitor with kubectl top — check actual vs requested regularly

Key Takeaways

• OpenClaw base usage is ~150MB; each session adds ~50MB
• Browser automation (Chromium sidecar) needs 1-2GB additional memory
• Use VPA in recommendation mode to right-size after real usage
• Consider burstable QoS (no CPU limit) for responsive agent sessions
• The official default (512Mi/250m → 2Gi/1CPU) works for most single-user deployments