OpenClaw Auto-Scaling with KEDA

💡 Quick Answer: Use KEDA ScaledObjects with Prometheus metrics from OpenClaw to auto-scale agent pods based on pending message queue depth or session count.

The Problem

Fixed-replica OpenClaw deployments waste resources during quiet periods and drop messages during traffic spikes. Manual scaling doesn’t respond fast enough to sudden bursts from Discord or Telegram channels.

The Solution

Deploy KEDA with custom Prometheus metrics from OpenClaw to scale agents based on actual workload demand.

Expose OpenClaw Metrics

Configure OpenClaw to expose Prometheus-compatible metrics:

apiVersion: v1
kind: ConfigMap
metadata:
  name: openclaw-config
  namespace: openclaw
data:
  openclaw.json: |
    {
      "metrics": {
        "enabled": true,
        "port": 9090,
        "path": "/metrics"
      }
    }

ServiceMonitor for Prometheus

apiVersion: monitoring.coreos.com/v1
kind: ServiceMonitor
metadata:
  name: openclaw-metrics
  namespace: openclaw
spec:
  selector:
    matchLabels:
      app: openclaw
  endpoints:
    - port: metrics
      interval: 15s
      path: /metrics

KEDA ScaledObject

apiVersion: keda.sh/v1alpha1
kind: ScaledObject
metadata:
  name: openclaw-scaler
  namespace: openclaw
spec:
  scaleTargetRef:
    name: openclaw-agent
  pollingInterval: 15
  cooldownPeriod: 300
  minReplicaCount: 1
  maxReplicaCount: 10
  triggers:
    - type: prometheus
      metadata:
        serverAddress: http://prometheus.monitoring:9090
        metricName: openclaw_pending_messages
        query: |
          sum(openclaw_pending_messages{namespace="openclaw"})
        threshold: "5"
        activationThreshold: "2"
    - type: prometheus
      metadata:
        serverAddress: http://prometheus.monitoring:9090
        metricName: openclaw_active_sessions
        query: |
          sum(openclaw_active_sessions{namespace="openclaw"})
        threshold: "10"

Advanced: Scale by Channel

apiVersion: keda.sh/v1alpha1
kind: ScaledObject
metadata:
  name: openclaw-discord-scaler
  namespace: openclaw
spec:
  scaleTargetRef:
    name: openclaw-discord
  pollingInterval: 10
  cooldownPeriod: 120
  minReplicaCount: 0
  maxReplicaCount: 5
  triggers:
    - type: prometheus
      metadata:
        serverAddress: http://prometheus.monitoring:9090
        metricName: openclaw_channel_queue
        query: |
          sum(openclaw_pending_messages{channel="discord"})
        threshold: "3"
  advanced:
    restoreToOriginalReplicaCount: true
    horizontalPodAutoscalerConfig:
      behavior:
        scaleUp:
          stabilizationWindowSeconds: 30
          policies:
            - type: Pods
              value: 2
              periodSeconds: 60
        scaleDown:
          stabilizationWindowSeconds: 300
          policies:
            - type: Pods
              value: 1
              periodSeconds: 120

Cost-Optimized: Scale to Zero

apiVersion: keda.sh/v1alpha1
kind: ScaledObject
metadata:
  name: openclaw-batch-scaler
  namespace: openclaw
spec:
  scaleTargetRef:
    name: openclaw-batch-worker
  pollingInterval: 30
  cooldownPeriod: 600
  minReplicaCount: 0  # Scale to zero when idle
  maxReplicaCount: 3
  triggers:
    - type: prometheus
      metadata:
        serverAddress: http://prometheus.monitoring:9090
        metricName: openclaw_queued_tasks
        query: |
          sum(openclaw_queued_tasks{type="batch"})
        threshold: "1"
        activationThreshold: "1"

graph TD
    A[Discord or Telegram Messages] --> B[OpenClaw Gateway]
    B --> C[Message Queue]
    C --> D[Prometheus Metrics]
    D --> E[KEDA Scaler]
    E --> F{Queue Depth > 5?}
    F -->|Yes| G[Scale Up Pods]
    F -->|No| H{Queue Empty for 5m?}
    H -->|Yes| I[Scale Down]
    H -->|No| J[Keep Current]

Common Issues

• Flapping — set cooldownPeriod ≥ 300s and use stabilizationWindowSeconds to prevent rapid scale up/down
• Cold start latency — keep minReplicaCount: 1 for latency-sensitive channels; use 0 only for batch workers
• Metric scrape delay — reduce pollingInterval to 10s for responsive scaling; align with Prometheus scrape interval
• API rate limits — Discord/Telegram rate limits apply per bot token regardless of replica count

Best Practices

• Separate ScaledObjects per channel type (Discord, Telegram, Slack)
• Use activationThreshold to prevent scaling on noise
• Set aggressive scaleUp (30s window) but conservative scaleDown (300s)
• Monitor with keda_scaler_metrics_value to verify KEDA sees your metrics
• Combine with PodDisruptionBudget for graceful scaling

Key Takeaways

• KEDA enables event-driven scaling based on actual message workload
• Scale-to-zero saves costs for batch and low-traffic agents
• Per-channel scalers allow independent scaling policies
• Prometheus integration provides flexible, query-based triggers
• Stabilization windows prevent flapping during bursty traffic