KUBELET WATCHES INODES. JUST NOT UNTIL IT'S AN EMERGENCY

Summary from the TLDR DevOps newsletter

Kubelet's background image garbage collection only monitors disk bytes, so workloads heavy on small files receive no early intervention before kubelet evicts pods. On one node, unpacked layers from a single-stage build containing the @mui/icons-material package consumed millions of inodes. A multi-stage Dockerfile can reduce the deployed image from more than 40,000 files to a few dozen. Adding a Prometheus alert for inode usage above 80% can catch the problem well before the 95% eviction threshold.