SIG Autoscaling Bi-Weekly Karpenter Working Group for 2025-09-11

The meeting discusses ongoing discussions around Carpenter's feature development, with a focus on improving node management and disruption handling. Key proposals include an RFC for introducing time-based do-not-disrupt annotations and exploring cluster-wide controls to halt drift and consolidation during incidents. A major use case involves stateful workloads requiring fine-grained disruption control to prevent data loss. The group agrees that a formal RFC is needed to explore these features, with existing limitations in current disruption mechanisms being addressed through proposed enhancements.

Key Takeaways:
• A time-based do-not-disrupt annotation is proposed to allow pods to specify disruption wait times, improving flexibility over binary true/false settings 10:12-15:00.
• An RFC is recommended to explore cluster-wide controls for halting node drift and consolidation during incidents, addressing CI/CD rollout issues and enabling emergency rollbacks 29:05-34:00.
• A dynamic disruption model is being considered to support stateful workloads, where critical pods can be protected from disruption to avoid data loss, with current pod disruption budgets deemed insufficient 20:33-22:47.
• The group acknowledges that existing mechanisms like termination grace periods and pod disruption budgets lack granularity for workload-specific needs, and suggests adding native support in Carpenter or through external operators 25:00-26:00.

The discussion concludes with a consensus to proceed with an RFC to evaluate these features, emphasizing the need for clear use cases and integration with existing workflows before implementation.

Sources:

• 10:12-15:00 Proposal for time-based do-not-disrupt annotat