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 reads stored metrics over a configurable lookback window controlled by RECOMMENDATION_LOOKBACK_DAYS (default: 7 days) and applies threshold-based detection algorithms. Results are deduplicated at the Deployment level — pods belonging to the same Deployment are grouped so you see one recommendation per workload, not one per replica.

Each recommendation includes:

• Type — one of 11 detection categories
• Priority — high, medium, or low
• Scope — pod, workload, namespace, node, or cluster
• Projected annual savings — potential_savings_co2e_grams and potential_savings_cost

When the API or startup scan runs, metrics from namespaces that no longer exist in Kubernetes are filtered out when the cluster API is reachable. That lets GreenKube reconcile old active recommendations as stale instead of regenerating them forever.

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 Pod Autoscalers. Identifies 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 with a generated CronJob schedule. Identifies nodes with far more capacity than their scheduled pods require. Flags nodes running at very low CPU and memory utilization — consolidation candidates. Flags PersistentVolumes whose claim was deleted, so provisioned storage keeps billing while unused. Flags LoadBalancer Services with no ready backing pods, so the cloud load balancer bills hourly for nothing.

Recommendation Types

🧟 Zombie Pods (ZOMBIE_POD)

What: Pods that are running and consuming resources but show near-zero energy usage.

Detection:

• Energy consumption < ZOMBIE_ENERGY_THRESHOLD
• Cost > ZOMBIE_COST_THRESHOLD (default: $0.01)

Scope: pod

---

📏 Rightsizing CPU (RIGHTSIZING_CPU)

What: Pods with CPU requests significantly higher than actual utilization.

Detection:

• Average CPU utilization < RIGHTSIZING_CPU_THRESHOLD × CPU request (default: 30%)
• Recommendation uses P95 usage, observed max, average usage, and a headroom multiplier for safe reductions
• Only reductions are surfaced — recommendations that would increase a request are discarded

Scope: workload (grouped per Deployment)

---

📏 Rightsizing Memory (RIGHTSIZING_MEMORY)

What: Pods with memory requests significantly higher than actual utilization.

Detection:

• Average memory utilization < RIGHTSIZING_MEMORY_THRESHOLD × memory request (default: 30%)

Scope: workload (grouped per Deployment)

---

📈 Autoscaling Candidates (AUTOSCALING_CANDIDATE)

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

Detection:

• Coefficient of Variation > AUTOSCALING_CV_THRESHOLD (default: 0.7)
• Max/min ratio > AUTOSCALING_SPIKE_RATIO (default: 3.0)
• No existing HPA detected

Scope: workload

---

🌍 Carbon-Aware Scheduling (CARBON_AWARE_SCHEDULING)

What: Workloads running during high-carbon-intensity windows that could be shifted to cleaner periods.

Detection: Grid intensity > zone average × CARBON_AWARE_THRESHOLD (default: 1.5×)

Scope: pod / workload

---

🗂️ Idle Namespace Cleanup (IDLE_NAMESPACE)

What: Namespaces with minimal energy consumption.

Detection: Total namespace energy < IDLE_NAMESPACE_ENERGY_THRESHOLD (default: 1,000 J)

Scope: namespace

---

🌙 Off-Peak Scaling (OFF_PEAK_SCALING)

What: Workloads with sustained idle periods during consistent time windows.

Detection: Idle period >= OFF_PEAK_MIN_IDLE_HOURS (default: 4h) at consistent hours, below OFF_PEAK_IDLE_THRESHOLD of the daily peak.

Output: Suggested CronJob/KEDA scale-down + scale-up schedule.

Scope: workload (grouped per Deployment)

---

🖥️ Overprovisioned Node (OVERPROVISIONED_NODE)

What: Nodes with far more capacity than their scheduled workloads require.

Detection: Node CPU utilization < NODE_UTILIZATION_THRESHOLD (default: 20%)

Scope: node

---

💤 Underutilized Node (UNDERUTILIZED_NODE)

What: Nodes running at very low CPU and memory utilization — consolidation candidates.

Detection: Node has fewer than 3 pods and average CPU utilization below 15%.

Scope: node

---

💾 Orphaned PersistentVolume (ORPHANED_PERSISTENT_VOLUME)

What: PersistentVolumes whose claim is gone — either Released phase (PVC deleted but not reclaimed) or a claimRef pointing at a PVC that no longer exists.

Detection: A new PVCollector lists all PVs and PVCs from the Kubernetes API on every scan.

Savings: Prefers the real per-volume storage cost from OpenCost (OpenCostCollector.collect_pv_costs()); otherwise falls back to provisioned capacity × STORAGE_COST_PER_GIB_MONTH (default $0.10/GiB-month). CO2e savings are not projected — energy estimation currently only covers CPU usage.

Scope: cluster (the PV name is stored in pod_name since PVs have no namespace)

---

⚖️ Orphaned LoadBalancer (ORPHANED_LOAD_BALANCER)

What: Services of type LoadBalancer with no ready backing endpoints — the provisioned cloud load balancer keeps billing hourly while routing traffic to nothing.

Detection: A new LoadBalancerCollector lists Services and Endpoints from the Kubernetes API and flags LoadBalancer Services whose selector matches no pods.

Savings: Prefers the real per-Service LoadBalancer cost from OpenCost (OpenCostCollector.collect_lb_costs()); otherwise falls back to LOAD_BALANCER_COST_PER_MONTH (default $18.00/month). CO2e savings are not projected — energy estimation currently only covers CPU usage.

Scope: cluster

---

Recommendation Lifecycle

Each recommendation is persisted in the database and reconciled across scans.

active -> applied
  |
  +-> ignored -> active
  |
  +-> stale

Status	Meaning
active	Current valid recommendation, shown in active lists, top recommendations, Prometheus active gauges, Grafana cards, and the frontend Active tab
applied	Recommendation explicitly implemented through the API; excluded from active lists and included in realized savings
ignored	Recommendation intentionally hidden with an audit reason; preserved for review and can be restored
stale	Previously active recommendation that no longer appears in the latest generated set

Managing Lifecycle in the Dashboard

On the /recommendations page:

• Active tab — current actionable recommendations with type filter and annual savings summary
• Ignored tab — previously ignored recommendations with restore action
• Realized Savings tab — applied recommendations and cumulative realized savings
• Ignore flow — reason is optional; restore is available from the Ignored tab

The current frontend does not expose an Apply button yet, even though the API supports applying recommendations.

Managing Lifecycle via API

# Mark a recommendation as applied
PATCH /api/v1/recommendations/{id}/apply

# Permanently ignore with a reason
PATCH /api/v1/recommendations/{id}/ignore

# Restore an ignored recommendation
DELETE /api/v1/recommendations/{id}/ignore

# Get active recommendations (optionally trigger a live refresh)
GET /api/v1/recommendations/active?refresh=true

# Get applied recommendations
GET /api/v1/recommendations/applied

# Get ranked top recommendations
GET /api/v1/recommendations/top?limit=5&metric=co2

# Get savings summary
GET /api/v1/recommendations/savings

Savings Ledger

When a recommendation is marked applied, GreenKube estimates realized annual savings on the recommendation row, then the SavingsAttributor converts those annual values into per-period ledger rows. Over time this accumulates into:

• greenkube_co2e_savings_attributed_grams_total — cumulative CO₂e savings (Prometheus gauge)
• greenkube_cost_savings_attributed_dollars_total — cumulative cost savings (Prometheus gauge)

The savings ledger is visible in:

• The Grafana dashboard’s Impact Command Center section (attributed savings timeline)
• The /api/v1/recommendations/savings endpoint

For fixed Grafana windows, prefer the pre-computed dashboard gauges:

• greenkube_dashboard_savings_co2e_grams_total
• greenkube_dashboard_savings_cost_dollars_total

Ranked Top Recommendations

GreenKube also exposes a ranked view of active recommendations with positive projected savings:

GET /api/v1/recommendations/top?limit=5&metric=co2
GET /api/v1/recommendations/top?limit=10&metric=cost&refresh=true

This API powers:

• The Grafana Actionable Recommendations row
• The greenkube_top_recommendations Prometheus gauge
• CO₂e-first or cost-first ranking via metric=co2|cost

Using Recommendations

CLI

# Get all recommendations (reads from database)
greenkube recommend

# Filter by namespace
greenkube recommend -n production

# Longer lookback window
greenkube recommend --last 7d

# Live collection (re-runs the full collection pipeline)
greenkube recommend --live

# CI/CD gate: exit 1 if any recommendations exist
greenkube recommend --fail-on-recommendations

API

# Live recommendations (runs the recommender)
GET /api/v1/recommendations

# Persisted active recommendations
GET /api/v1/recommendations/active

# Persisted applied recommendations
GET /api/v1/recommendations/applied

# Ranked active recommendations by projected annual savings
GET /api/v1/recommendations/top

# History
GET /api/v1/recommendations/history

Tuning Thresholds

All thresholds are configurable via Helm values.yaml or environment variables:

config:
  recommendations:
    lookbackDays: 7                   # Analysis window used by API/startup scan
    rightsizingCpuThreshold: 0.3      # 30% usage triggers CPU rightsizing
    rightsizingMemoryThreshold: 0.3   # 30% usage triggers memory rightsizing
    rightsizingHeadroom: 1.2          # 20% safety margin on new request
    zombieCostThreshold: 0.01         # $0.01/day minimum to flag
    zombieEnergyThreshold: 1000       # 1,000 Joules minimum to flag
    autoscalingCvThreshold: 0.7       # High variability coefficient
    autoscalingSpikeRatio: 3.0        # 3× max/min ratio
    carbonAwareThreshold: 1.5         # 1.5× average intensity
    nodeUtilizationThreshold: 0.2     # 20% CPU for overprovisioned node
    offPeakIdleThreshold: 0.05        # 5% of daily peak
    offPeakMinIdleHours: 4            # 4h idle to suggest off-peak scaling
    minCpuMillicores: 10              # Floor for generated CPU requests
    minMemoryBytes: 16777216          # Floor for generated memory requests
    applyTolerance: 0.25              # Reserved for future auto-apply detection
    storageCostPerGibMonth: 0.10      # Fallback cost for orphaned PV savings ($/GiB-month)
    loadBalancerCostPerMonth: 18.0    # Fallback cost for orphaned LoadBalancer savings ($/month)

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