Cluster autoscaling is not straightforward. Scaling solely based on pod resource requests (CPU, memory, etc.) often leads to inefficiencies and waste. To solve this, we introduce a fine-grained, cost-aware, and workload-specific autoscaling plugin for Karpenter:
- Cost-Aware Scaling – Dynamically provisions non-preemptible (expensive) and preemptible (cheap) VMs based on workload stability, maximizing cost efficiency. A simple YAML file can define scaling policies for different scenarios.
- Performance-Aware Scaling – Uses a scoring system to assess application performance needs and VM capabilities (CPU, I/O, etc.), scaling VMs to meet the needs while ensuring an optimal balance between performance and cost.
- ARM-Optimized Scaling – Detects ARM-compatible workloads and prioritizes cost-efficient ARM instances.
With this fine-grained, cost-aware, and workload-specific Karpenter plugin, we enable precise scaling and sharp cost reductions without compromising availability compliance.