gcloud-lab/infrastructure/gpus/base/keda-gpu-scaling/keda-vllm.yaml

apiVersion: http.keda.sh/v1alpha1
kind: HTTPScaledObject
metadata:
  name: rtx6000-scaling
  namespace: customer1
spec:
  hosts:
  - rtx6000-brain-service.customer1.svc.cluster.local
  scaleTargetRef:
    name: rtx6000-brain-vllm
    kind: Deployment
    apiVersion: apps/v1
    service: rtx6000-brain-service
    port: 8000

  replicas:
    min: 0
    max: 1
  pollingInterval: 10
  cooldownPeriod: 30
  scaledownPeriod: 900
fix: restrict KEDA scaling to /v1/(chat/)?completions paths only - Switch to HTTPScaledObject (matches a100 pattern) - Probes (/health, /v1/models) now ignored for scaling - Keeps fast polling (10s), quick cooldown (30s), 15min idle scale-to-0 Health/readiness/startup probes hit pod IP directly (bypass service), so never triggered scaling anyway — but paths ensure only inference traffic scales. 2026-04-29 04:14:46 +00:00			`apiVersion: http.keda.sh/v1alpha1`
			`kind: HTTPScaledObject`
feat: add KEDA scaling for A100 vLLM - Scale to 0 when idle (saves money) - 15 min cooldown before scale-down - Triggers on 3+ pending HTTP requests - Targets openclaw-brain-vllm deployment 2026-04-22 17:07:05 +00:00			`metadata:`
Update keda-vllm.yaml 2026-04-28 23:58:26 -04:00			`name: rtx6000-scaling`
feat: add KEDA scaling for A100 vLLM - Scale to 0 when idle (saves money) - 15 min cooldown before scale-down - Triggers on 3+ pending HTTP requests - Targets openclaw-brain-vllm deployment 2026-04-22 17:07:05 +00:00			`namespace: customer1`
			`spec:`
fix: restrict KEDA scaling to /v1/(chat/)?completions paths only - Switch to HTTPScaledObject (matches a100 pattern) - Probes (/health, /v1/models) now ignored for scaling - Keeps fast polling (10s), quick cooldown (30s), 15min idle scale-to-0 Health/readiness/startup probes hit pod IP directly (bypass service), so never triggered scaling anyway — but paths ensure only inference traffic scales. 2026-04-29 04:14:46 +00:00			`hosts:`
			`- rtx6000-brain-service.customer1.svc.cluster.local`
feat: add KEDA scaling for A100 vLLM - Scale to 0 when idle (saves money) - 15 min cooldown before scale-down - Triggers on 3+ pending HTTP requests - Targets openclaw-brain-vllm deployment 2026-04-22 17:07:05 +00:00			`scaleTargetRef:`
Update keda-vllm.yaml 2026-04-28 23:58:26 -04:00			`name: rtx6000-brain-vllm`
fix(keda): correct HTTPScaledObject for A100 vLLM wake-on-request - Use proper scaleTargetRef structure for http.keda.sh/v1alpha1 - Uncomment keda-vllm.yaml in kustomization - This should allow Flux to reconcile the scaler 2026-04-23 02:55:07 +00:00			`kind: Deployment`
			`apiVersion: apps/v1`
fix: restrict KEDA scaling to /v1/(chat/)?completions paths only - Switch to HTTPScaledObject (matches a100 pattern) - Probes (/health, /v1/models) now ignored for scaling - Keeps fast polling (10s), quick cooldown (30s), 15min idle scale-to-0 Health/readiness/startup probes hit pod IP directly (bypass service), so never triggered scaling anyway — but paths ensure only inference traffic scales. 2026-04-29 04:14:46 +00:00			`service: rtx6000-brain-service`
			`port: 8000`
fix(keda): remove invalid spec.paths from HTTPScaledObject HTTPScaledObject CRD (http.keda.sh/v1alpha1) does not support .spec.paths field. Scaling now triggers on all HTTP requests to the host (health/liveness ignored via targetPendingRequests:1 + pollingInterval:10s). Refs https://keda.sh/docs/scalers/http-addon/ 2026-04-29 04:50:10 +00:00
update keda 2026-04-23 01:37:09 +00:00			`replicas:`
			`min: 0`
			`max: 1`
fix: restrict KEDA scaling to /v1/(chat/)?completions paths only - Switch to HTTPScaledObject (matches a100 pattern) - Probes (/health, /v1/models) now ignored for scaling - Keeps fast polling (10s), quick cooldown (30s), 15min idle scale-to-0 Health/readiness/startup probes hit pod IP directly (bypass service), so never triggered scaling anyway — but paths ensure only inference traffic scales. 2026-04-29 04:14:46 +00:00			`pollingInterval: 10`
			`cooldownPeriod: 30`
			`scaledownPeriod: 900`