Kubernetes resource limits: predictability vs. efficiency

Milan Plzík's blog post explores the trade-offs between predictability and efficiency when setting Kubernetes resource limits, arguing that while excessive limits can lead to unused compute power and latency increases, they also offer valuable predictability and insights into workload behavior under stress. He discusses the inherent challenges in achieving optimal resource utilization within Kubernetes clusters and suggests that resource limits, though tricky to configure, are crucial for understanding workloads during extreme conditions and preventing potential catastrophes. He highlights two strategies: setting a fixed-fraction headroom above requests to allow for some shared resource use while maintaining predictability, and configuring workloads with requests equal to limits for a more predictable performance. The post underscores the economic implications of spiky resource usage, emphasizing that having easily accessible "freebie" resources may disincentivize performance improvements and could compromise service level agreements (SLAs) during resource spikes.