Recommendations

import { Card, CardGrid } from ‘@astrojs/starlight/components’;

GreenKube analyzes your cluster metrics to generate actionable recommendations that reduce both costs and carbon emissions.

Recommendation Engine

The recommendation engine examines metrics collected over a configurable lookback period (default: 7 days) and applies multiple analysis algorithms:

Identifies pods consuming resources but producing little or no value — idle workloads that should be terminated. Finds pods with CPU or memory requests significantly higher than actual usage, suggesting smaller resource allocations. Detects workloads with high usage variability that would benefit from Horizontal or Vertical Pod Autoscalers. Identifies batch workloads that could be time-shifted to periods of lower grid carbon intensity. Spots namespaces with minimal activity that may contain forgotten resources consuming energy and cost. Suggests scaling down workloads during off-peak hours when high resource allocation isn't needed. Identifies nodes with far more capacity than their scheduled pods require. Flags nodes running at very low utilization, wasting energy and money.

Recommendation Types

🧟 Zombie Pods (ZOMBIE_POD)

What: Pods that are running and consuming resources but show minimal CPU/energy usage.

Detection Criteria:

• CPU usage consistently below threshold
• Energy consumption below zombieEnergyThreshold (default: 1000 J)
• Cost above zombieCostThreshold (default: $0.01)
• Running for extended periods

Example Output:

🧟 ZOMBIE POD — High Severity
Pod: legacy-api-deployment-5d8f7c-k2m4p
Namespace: staging
  CPU usage: 2m (avg over 7 days)
  Energy: 450 J
  Cost: $0.85/day
  Estimated savings: $0.85/day, 3.2g CO₂e/day
  Action: Consider terminating this idle workload
    kubectl delete deployment legacy-api-deployment -n staging

📏 Rightsizing (RIGHTSIZING_CPU / RIGHTSIZING_MEMORY)

What: Pods with resource requests (CPU or memory) significantly higher than actual utilization. Two sub-types: RIGHTSIZING_CPU for over-provisioned CPU and RIGHTSIZING_MEMORY for over-provisioned memory.

Detection Criteria:

• Average CPU usage < rightsizingCpuThreshold × CPU request (default: 30%)
• Average memory usage < rightsizingMemoryThreshold × memory request (default: 30%)
• Applies rightsizingHeadroom multiplier (default: 1.2x) for safe recommendations

Example Output:

📏 RIGHTSIZING — Medium Severity
Pod: web-frontend-7b5f8c9d6-x2k4p
Namespace: production
  CPU: Using 45m of 500m requested (9%)
  Recommendation: Reduce CPU request to 55m (45m × 1.2 headroom)
  Memory: Using 128Mi of 512Mi requested (25%)
  Recommendation: Reduce memory request to 154Mi
  Estimated savings: $1.20/day, 5.4g CO₂e/day

📈 Autoscaling Candidates (AUTOSCALING_CANDIDATE)

What: Workloads with high variability in resource usage that would benefit from autoscaling.

Detection Criteria:

• High coefficient of variation (CV > autoscalingCvThreshold, default: 0.7)
• Spike detection (max/avg ratio > autoscalingSpikeRatio, default: 3.0)
• No existing HPA/VPA detected

Example Output:

📈 AUTOSCALING — Medium Severity
Pod: batch-processor-deployment-8c7d6-n3m2p
Namespace: default
  CPU variability: CV = 1.2 (high)
  Peak: 800m, Average: 150m (5.3x spike ratio)
  Recommendation: Configure HPA with:
    minReplicas: 1, maxReplicas: 5
    targetCPUUtilization: 70%

🌍 Carbon-Aware Scheduling (CARBON_AWARE_SCHEDULING)

What: Batch or deferrable workloads that could be scheduled during periods of lower grid carbon intensity.

Detection Criteria:

• Grid intensity during workload execution exceeds carbonAwareThreshold × average (default: 1.5x)
• Workload appears to be a batch job (CronJob, Job owner)
• Lower-intensity periods available in the same zone

Example Output:

🌍 CARBON-AWARE — Low Severity
Pod: nightly-etl-job-28445-k2m4p
Namespace: data-pipeline
  Current intensity: 520 gCO₂e/kWh (peak hours)
  Average zone intensity: 280 gCO₂e/kWh
  Recommendation: Schedule during off-peak hours (02:00-06:00 UTC)
  Estimated savings: 8.5g CO₂e/run

🗂️ Idle Namespace Cleanup (IDLE_NAMESPACE)

What: Namespaces with minimal activity that may contain forgotten resources.

Detection Criteria:

• Total energy below idleNamespaceEnergyThreshold (default: 1000 J)
• Low pod count with minimal resource usage

🌙 Off-Peak Scaling (OFF_PEAK_SCALING)

What: Workloads that maintain high resource allocation during off-peak hours and could benefit from scheduled scaling down.

Detection Criteria:

• Consistently low utilization during off-peak periods
• Workloads without existing time-based scaling policies

🖥️ Overprovisioned Node (OVERPROVISIONED_NODE)

What: Nodes with significantly more capacity than what the scheduled pods require.

Detection Criteria:

• Node resource utilization consistently below threshold
• Opportunity to consolidate workloads onto fewer, right-sized nodes

💤 Underutilized Node (UNDERUTILIZED_NODE)

What: Nodes running with very low utilization that are wasting energy and money.

Detection Criteria:

• Node utilization below nodeUtilizationThreshold (default: 20%)
• Running for extended periods at low usage

Using Recommendations

CLI

# Get all recommendations
greenkube recommend

# Filter by namespace
greenkube recommend -n production

# JSON output for automation
greenkube recommend --format json

API

curl "http://localhost:8000/api/v1/recommendations?namespace=production"

Dashboard

Navigate to the Recommendations tab in the web dashboard for a visual, filterable view of all suggestions.

Tuning Thresholds

All thresholds are configurable via Helm values:

config:
  recommendations:
    lookbackDays: 7
    rightsizingCpuThreshold: 0.3      # 30% usage triggers rightsizing
    rightsizingMemoryThreshold: 0.3
    rightsizingHeadroom: 1.2           # 20% safety margin
    zombieCostThreshold: 0.01         # $0.01/day minimum
    zombieEnergyThreshold: 1000       # 1000 Joules minimum
    autoscalingCvThreshold: 0.7       # High variability
    autoscalingSpikeRatio: 3.0        # 3x spike ratio
    carbonAwareThreshold: 1.5         # 1.5x average intensity

Adjust these based on your cluster size, workload patterns, and organizational priorities.