Phase 7: Advanced Features

[diranged]

Goal

Add sophisticated decision logic to improve cost optimization effectiveness.

Deliverables

• Spot price volatility tracking: Calculate standard deviation, prefer RI/SP when spot fluctuates
• Time-of-day patterns: Learn when RI/SP preference is most valuable
• Workload-specific tuning: Different strategies per NodePool based on labels
  • cost-optimization: prefer-reserved - Always prefer RI/SP when available
  • cost-optimization: prefer-spot - Always prefer spot for transient workloads
  • cost-optimization: balanced - Default algorithm
• Capacity forecasting: Predict future capacity needs to avoid over-committing
• Multi-region optimization: Coordinate decisions across regions
• Historical cost analysis: Track long-term savings trends

Spot Price Volatility Detection

• Calculate rolling standard deviation of spot prices (15-minute window)
• If spot_price_std_dev > 0.20, increase preference for stable RI/SP pricing
• Reduces cost unpredictability

Time-of-Day Patterns

• Track provisioning patterns by hour/day of week
• Learn when capacity demands spike
• Pre-emptively adjust overlays before known busy periods

Workload-Specific Tuning

Support NodePool labels:

apiVersion: karpenter.sh/v1beta1
kind: NodePool
metadata:
  name: batch-jobs
  labels:
    cost-optimization: prefer-spot  # Override for transient workloads

Capacity Forecasting

• Track RI/SP capacity utilization trends
• Predict when capacity will be exhausted
• Proactively adjust overlays before capacity runs out

Success Criteria

• Volatility detection reduces cost unpredictability by 15%
• Time-based patterns improve average cost by 5%
• Per-NodePool strategies documented and configurable
• Capacity forecasting reduces overlay thrashing by 20%

Testing Requirements

• Unit tests for all advanced features
• Integration tests with historical data
• A/B testing in production to validate improvements

References

• RFC-0003: Lines 383-390 (Volatility preference)
• RFC-0003: Lines 388-390 (Critical workloads)
• RFC-0003: Lines 403-407 (Transient workloads)
• RFC-0003: Lines 887-902

[diranged]

Status Update

✅ Completed: Workload-Specific Tuning

The workload-specific tuning deliverable is now complete via the preference annotation system (pkg/preference/).

Original proposal:

labels:
  cost-optimization: prefer-reserved  # Fixed strategies

Implemented (more flexible):

annotations:
  veneer.io/preference.1: "karpenter.sh/capacity-type=on-demand adjust=-100%"
  veneer.io/preference.2: "kubernetes.io/arch=arm64 adjust=-50%"

The preference system provides:

• Per-NodePool customization via annotations
• Arbitrary label matchers (arch, family, capacity-type, size, cpu, memory, etc.)
• Configurable price adjustments (percentage-based)
• Multiple preferences with priority ordering (higher number = higher priority)

Original Strategy	Preference Equivalent
prefer-reserved	karpenter.sh/capacity-type=on-demand adjust=-100%
prefer-spot	karpenter.sh/capacity-type=spot adjust=-50%
balanced	No annotation (default behavior)

Implementation:

• pkg/preference/types.go - Type definitions
• pkg/preference/parser.go - Annotation parsing
• pkg/preference/generator.go - NodeOverlay generation
• pkg/reconciler/nodepool.go - Reconciliation loop
• Full test coverage in *_test.go files

---

❌ Remaining (Future Work)

The following advanced features remain unimplemented:

• Spot price volatility tracking
• Time-of-day patterns
• Capacity forecasting
• Multi-region optimization
• Historical cost analysis

These are analytics/ML features that would require additional data collection and analysis infrastructure.