AgentSkillsCN

cluster-resource-health

检查 Kubernetes 集群健康状况，包括 Pod 状态、节点条件、资源利用率，以及 EKS 集群中的待处理告警。适用于监控基础设施健康、排查容量问题，或执行集群审计时使用。

SKILL.md
--- frontmatter
name: cluster-resource-health
description: Check Kubernetes cluster health including pod status, node conditions, resource utilization, and pending alerts across EKS clusters. Use when monitoring infrastructure health, investigating capacity issues, or performing cluster audits.

Cluster Resource Health

Query AWS EKS clusters for node health, pod status, resource utilization, and alerts to produce a cluster health dashboard.

Instructions

Phase 1: Cluster Overview (AWS Agent)

  1. •List EKS clusters and their status:
    • •Cluster name, version, and status
    • •Node group configurations (instance types, desired/min/max counts)
    • •Current node count and readiness
  2. •Check Kubernetes version:
    • •Current version vs. latest available
    • •End-of-support date for current version

Phase 2: Node Health

  1. •Inspect node conditions using kubectl via the AWS agent:
    • •Ready, MemoryPressure, DiskPressure, PIDPressure, NetworkUnavailable
    • •Node allocatable vs. requested resources
    • •Unschedulable nodes (cordoned/drained)
  2. •Resource utilization per node:
    • •CPU requested vs. allocatable (%)
    • •Memory requested vs. allocatable (%)
    • •Pod count vs. pod limit

Phase 3: Pod Health

  1. •Identify problematic pods:
    • •CrashLoopBackOff, ImagePullBackOff, OOMKilled
    • •Pending pods (unable to schedule)
    • •Pods with high restart counts (>5)
    • •Evicted pods
  2. •Namespace-level summary:
    • •Pods running, pending, failed per namespace
    • •Resource quotas and limit ranges

Phase 4: Resource Capacity Analysis

  1. •Cluster-wide utilization:
    • •Total CPU requested vs. total allocatable
    • •Total memory requested vs. total allocatable
    • •Headroom for new workloads
  2. •Capacity risks:
    • •Nodes at >80% resource utilization
    • •Namespaces exceeding resource quotas
    • •PersistentVolume claims pending or near capacity

Output Format

markdown
## Cluster Resource Health Report
**Generated**: February 9, 2026

### Cluster Summary
| Cluster | Version | Nodes | Status | Overall Health |
|---------|---------|-------|--------|----------------|
| prod-us-west-2 | 1.29 | 12/12 Ready | Active | HEALTHY |
| staging-us-west-2 | 1.28 | 4/4 Ready | Active | WARNING |

### Resource Utilization (prod-us-west-2)
| Resource | Requested | Allocatable | Utilization |
|----------|-----------|-------------|-------------|
| CPU | 38 cores | 48 cores | 79% |
| Memory | 96 Gi | 128 Gi | 75% |
| Pods | 187 | 440 | 43% |

**Headroom**: Can schedule ~10 more standard pods (1 CPU, 2Gi each)

### Problematic Pods
| Pod | Namespace | Status | Restarts | Node |
|-----|-----------|--------|----------|------|
| payment-api-7d4b8c | production | CrashLoopBackOff | 23 | node-3 |
| data-pipeline-abc | batch | OOMKilled | 5 | node-7 |
| image-proc-xyz | processing | ImagePullBackOff | 0 | node-2 |

### Node Health
| Node | Status | CPU Req% | Mem Req% | Pods | Conditions |
|------|--------|----------|----------|------|------------|
| node-1 | Ready | 82% | 71% | 18 | OK |
| node-7 | Ready | 91% | 88% | 22 | MemoryPressure |

### Capacity Risks
1. **HIGH**: node-7 at 91% CPU / 88% memory - consider scaling node group
2. **MEDIUM**: staging cluster on EKS 1.28 - EOL in 60 days, plan upgrade
3. **LOW**: 3 PVCs at >80% capacity in `data` namespace

### Recommendations
1. **Immediate**: Investigate payment-api CrashLoopBackOff (23 restarts)
2. **Short-term**: Scale prod node group from 12 to 14 nodes (headroom at 79%)
3. **Planned**: Upgrade staging cluster from EKS 1.28 to 1.29
4. **Optimization**: Right-size data-pipeline pods (OOMKilled - increase memory limit)

Examples

  • •"Check the health of our EKS clusters"
  • •"Are there any failing pods in production?"
  • •"Show me cluster resource utilization"
  • •"Which nodes are under memory pressure?"
  • •"Do we have enough capacity for a new deployment?"

Guidelines

  • •Check all clusters unless a specific cluster is requested
  • •Flag any node above 85% resource utilization as a capacity risk
  • •For CrashLoopBackOff pods, suggest checking logs as the immediate action
  • •EKS version end-of-support should be flagged at least 90 days before EOL
  • •Group pods by issue type (crash, OOM, image pull) for easier triage
  • •Include pod restart counts - high restarts indicate chronic issues even if currently running
  • •When capacity is tight, recommend specific scaling actions (node count, instance type)
  • •Use kubectl read-only commands only (never modify cluster state during health checks)