Phase 2: Cluster Autoscaler Priority-Based Scale-Up for 10k+ Nodes

Problem

At 2 workers per node (400m CPU request), 10k workers requires 5,000+ nodes. Node provisioning becomes a bottleneck — OpenStack API rate limits, node bootstrap time (2-5 min per node), and cluster autoscaler max-node-provision-time (20 min default) all constrain scale-up speed (B8).

What Phase 1 Did

• Reduced worker CPU request from 600m→400m for better pod density (~2 per 1-core slice)
• Increased worker memory request from 700Mi→1Gi
• Added worker resource limits (CPU: 2, memory: 2Gi)

What Remains (This Issue)

1. Priority-based scale-up: Add workerPriority and webPriority values that map to priority class names, enabling cluster-autoscaler to prioritize worker nodes over other pods during scale-up events
2. ProvisioningConfig: Add cluster-autoscaler provisioningConfig support with multiple node group definitions (worker, web, system) and per-node-group minSize / maxSize / instanceType
3. Over-provisioning: Add a buffer node group that creates extra nodes ahead of scale-up demand using PDB-violating low-priority pause pods
4. Bottleneck documentation: Update values_large.templateyaml with node provisioning guidance per provider:
   • OpenStack: Recommended CAPI node group sizes, Nova quotas
   • AWS: Recommended ASG warm pools, instance families
   • GKE/AKS: Recommended node pool configs

Design Notes

• Priority-based scale-up: priorityClassName on pending pods determines which node pool expands first
• Over-provisioning uses pause containers with very low priority that PDBs don't protect; when real pods arrive, pause pods are preempted
• This is provider-specific — AWS has ASG warm pools, OpenStack CAPO has ServerGroup/ServerTemplate
• The chart should document but not implement provider-specific autoscaler config (defer to cloud-provider subdirectories)

References

• .agents/results/result-tf-infra.md — B8: Node Provisioning (HIGH)
• openstudio-server/values.yaml — worker.container.resources.requests.cpu: 400m
• openstudio-server/templates/_scheduling.tpl — priority classes

Acceptance Criteria

• Priority class hierarchy documented in values.yaml
• Buffer/over-provisioning pause pod template (opt-in, disabled by default)
• values_large.templateyaml updated with node provisioning guidance per provider
• OpenStack-specific CAPO node group sizing documented
• AWS-specific ASG warm pool config documented

Start with openstudio-server/values.yaml and openstudio-server/templates/_scheduling.tpl to understand the existing worker resources and priority classes. Then inspect values_large.templateyaml and the chart structure for where provisioningConfig, node groups, buffer pods, and provider guidance would fit. Done means the opt-in buffer configuration, documented priority hierarchy, and OpenStack and AWS provisioning guidance satisfy the listed acceptance criteria.