chart: add optional scheduler/worker split (v0.4.0) and HPA support

[Bierchermuesli]

• Adds an optional scheduler Deployment that runs in scheduler-only mode (NETDISCO_WORKERS_TASKS=0, no pollers)
• When scheduler.enabled=true, the backend Deployment automatically receives NETDISCO_NO_SCHEDULER=1, making it safe to run multiple backend replicas without duplicate job submissions into the PostgreSQL queue
• Adds optional CPU+memory HPA for the backend Deployment (backend.hpa.enabled, disabled by default) with configurable scale-down stabilization to avoid killing pods mid-job
• Omits replicas from the backend Deployment when HPA is enabled so ArgoCD/Flux don't fight the HPA
• Adds Vault agent resource limits/requests annotations to avoid consuming namespace CPU quota unexpectedly
• Adds revisionHistoryLimit: 3 to all Deployments to cap stale ReplicaSets
• Disabled by default — no behaviour change for existing deployments

How?

The netdisco backend runs three MCE roles in one process: Scheduler (submits jobs on cron), Manager (pulls from PG queue), and Poller (executes jobs). Running multiple replicas today causes the Scheduler to submit duplicate jobs every minute.

With this split:

• 1× scheduler pod: Scheduler only, zero pollers
• N× backend pods: Manager + Pollers, no Scheduler

Requires NETDISCO_NO_SCHEDULER and NETDISCO_WORKERS_TASKS env var support in the netdisco backend binary.

HPA example (CPU+memory)

backend:
  hpa:
    enabled: true
    minReplicas: 2
    maxReplicas: 4
    targetCPUUtilization: 80
    targetMemoryUtilization: 80
    scaleDownStabilizationSeconds: 300  # avoid killing pods mid-job

HPA example (KEDA / queue depth) — untested, requires KEDA on the cluster

netdisco exposes netdisco_jobs{status="queued"} on the web /metrics endpoint which makes it a natural KEDA trigger.

Replicas are calculated as ceil(queueDepth / threshold). Tune the threshold to match workers.tasks and your expected queue depth — a full discovery run can queue hundreds of jobs, so ~50 gives a gradual ramp without immediately pegging at maxReplicas:

apiVersion: keda.sh/v1alpha1
kind: ScaledObject
metadata:
  name: netdisco-backend
spec:
  scaleTargetRef:
    name: netdisco-backend
  minReplicaCount: 1
  maxReplicaCount: 4
  triggers:
    - type: prometheus
      metadata:
        serverAddress: http://prometheus:9090
        metricName: netdisco_queued_jobs
        query: netdisco_jobs{status="queued",tenant="netdisco"}
        threshold: "50"   # ceil(queueDepth/threshold) replicas — 200 jobs → 4 replicas

Todo

• helm template with scheduler.enabled: false produces identical output to before
• helm template with scheduler.enabled: true produces both netdisco-backend and netdisco-scheduler Deployments
• Scheduler pod has NETDISCO_WORKERS_TASKS=0, backend pods have NETDISCO_NO_SCHEDULER=1
• HPA applies cleanly with autoscaling/v2

[ollyg]

I've no comments - happy for you to merge @Bierchermuesli when you feel it's ready!

[ollyg]

BTW yes, this is very cool to see, thank you! :-D