Title

AWS re:Invent 2023 - Harness the power of Karpenter to scale, optimize & upgrade Kubernetes (CON331)

Summary

• Karpenter is a new Kubernetes scaler that directly interacts with EC2 APIs, bypassing the cluster autoscaler and autoscaling groups, resulting in faster and more flexible scaling.
• It is Kubernetes native, open source, and now a CNCF project.
• Karpenter can optimize costs, support diverse workloads, and assist with upgrades and patches.
• It combines the features of cluster autoscaler, node groups, node termination handlers, and descheduler into a single cohesive stack.
• The tool uses YAML files (NodePool and EC2 NodeClass) to control instance provisioning and offers features like automatic instance selection, spot interruption handling, and weighted provisioning strategies.
• Karpenter's node disruption workflow relies on pod disruption budgets and annotations to minimize disruption during scaling down.
• It supports automatic upgrades and patching through drift detection, which reconciles differences between desired and actual node states.
• The tool can be installed via AWS Public ECR and supports custom AMI pipelines.
• Karpenter emits Prometheus metrics for monitoring and supports day two operations for seamless Kubernetes version upgrades and AMI updates.
• The project is actively developed with community engagement encouraged.

Insights

• Karpenter's direct interaction with EC2 APIs and Kubernetes native design provide a more efficient scaling solution compared to traditional cluster autoscalers.
• The tool's ability to optimize costs through intelligent bin-packing and instance selection can lead to significant savings, as demonstrated by Sentinel-1's reported 50% cost reduction.
• Karpenter's support for custom AMIs and integration with existing AMI pipelines ensures that organizations can maintain their custom configurations while benefiting from Karpenter's scaling capabilities.
• The project's active community and CNCF backing suggest a strong future for Karpenter in the Kubernetes ecosystem, with ongoing improvements and support.
• The detailed explanation of Karpenter's internal algorithms for scheduling, batching, bin-packing, and launch decisions highlights the complexity of the problem space and the sophistication of the solution.
• The distinction between voluntary and involuntary disruptions in Karpenter's node disruption workflow is crucial for maintaining cluster stability and minimizing the impact on running applications.
• The tool's ability to handle drift and automatically reconcile node states can greatly simplify Kubernetes version upgrades and security patching, reducing the operational burden on teams.
• Karpenter's design considerations, such as minimizing pod disruption and preferring older nodes for termination, reflect a thoughtful approach to maintaining high availability and minimizing negative impacts on workloads.